Welcome. This is a collection of articles, images, and whatever else ends up here.
Folders so far:
The Reading List on the desktop is the curated list of books, podcasts, and newsletters that everything here was built from.
Double-click icons to open them. You can have multiple windows open, drag them around, and minimize them to the taskbar. On mobile, tap twice to open.
The Recycle Bin is empty.
Wooldridge and Jennings defined the properties of an intelligent agent in their 1995 survey: autonomy, reactivity, pro-activeness, and social ability. Thirty years later, those properties still hold. But the implementation looks nothing like what they imagined.
A chatbot generates text. You give it a prompt, it gives you a response, and the interaction ends. An agent does something different. It has tools, which are functions it can call to interact with the outside world. Read a file. Run a shell command. Search the web. Create a pull request. Query a database. The agent calls a tool, reads the result, decides what to do next, and repeats until the task is done.
That loop is the defining feature. The model reasons, acts, observes the outcome, and reasons again. Without it, you have a chatbot. With it, you have an agent.
Yao et al. formalized this pattern in their 2022 paper "ReAct: Synergizing Reasoning and Acting in Language Models." The name combines Reasoning and Acting. The structure is simple:
Each cycle through the loop is called a "turn." The number of turns varies with the task. A simple question might take 3 turns: search, read, summarize. A complex feature implementation might take 500: read code, write code, run tests, see failures, fix the failures, commit, push. A chatbot always takes exactly one turn. The agent's ability to take many is what gives it the capacity to handle work that requires sustained effort.
A tool is a function the agent can call. The term is specific. It refers to a discrete capability with a defined input and output that the agent invokes during its reasoning loop. Some common examples:
The agent sees a list of available tools and their descriptions at the start of a session. When it decides to use one, it formats the input as structured JSON. A separate runtime executes the tool and returns the output into the conversation context. The model never runs code directly. This separation between reasoning and execution is a deliberate design choice that makes the system easier to control and audit.
Because agents act on the real world, they need constraints. Reading a file is low-risk. Deleting one is not. Running arbitrary shell commands sits somewhere in between, depending on the command.
Most agent frameworks implement a permission system that determines which tools the agent can use, under what conditions, and whether human approval is required. Some tools get auto-approved. Others require explicit confirmation before each use. The specifics vary by framework, but the principle is consistent: the agent's autonomy should be proportional to the risk of the action it wants to take.
Every agent operates within a context window, the same token limit that constrains any LLM conversation. Every tool call, every result, every piece of reasoning takes up space. Long-running agents can exhaust this limit, which is why most frameworks support compaction, a process where older parts of the conversation are summarized to free up room for new information.
Some agents also maintain persistent memory that survives across sessions. This might be a JSON file on disk, a database, or a structured memory system with different scopes (per-user, per-project, global). The distinction matters: context is what the agent knows during a single conversation, memory is what it retains between conversations.
Complex tasks sometimes benefit from multiple agents working together. A coordinator agent decides what needs to happen and spawns specialized subagents to handle parts of the work. Each subagent gets its own context, its own tool set, and its own focus. Results flow back to the coordinator for synthesis.
This pattern has roots in the multi-agent systems research of the 1990s and 2000s, particularly the FIPA standards for inter-agent communication. The modern version is simpler. Subagents are typically defined by a system prompt, a list of allowed tools, and optionally a different model. They run in isolation and return a single result.
The practical benefit is separation of concerns. A code review task might spawn a security reviewer and a performance reviewer in parallel. Each one stays focused on its own domain without the conversation getting muddled.
There are several ways to build agents today. The choice depends on how much control you need over the agent loop and how tightly the agent integrates with your existing systems.
None of these is universally best. The right choice depends on your constraints.
Anthropic ships three things that matter for building agents: the Claude model (the LLM itself), Claude Code (an interactive agent product), and the Claude Agent SDK (a library for building your own agents in Python or TypeScript). A fourth piece, MCP (Model Context Protocol), is an open standard that lets agents connect to external tools. These pieces layer on top of each other.
Claude is the large language model. The current generation has three sizes: Opus (the most capable and most expensive), Sonnet (balanced between capability and speed), and Haiku (fast, lightweight, cheap). All three support tool use, which means they can generate structured requests to call external functions as part of their response.
You access Claude through the Anthropic API. You send messages, the model generates responses. If you include tool definitions in the request, Claude can produce "tool use" blocks alongside its text. Your code executes the tool and sends the result back. That exchange, repeated in a loop, is how agents work. But managing that loop yourself is tedious. That's where the SDK comes in.
Claude Code is the product most people interact with directly. It runs as a CLI (claude command), a VS Code extension, a JetBrains plugin, a desktop app, or through the web. It comes with a built-in set of tools: Read, Edit, Write, Bash, Glob, Grep, WebSearch, WebFetch, and several others. It manages its own agent loop, permission system, and session persistence.
For automation, Claude Code has a headless mode. You pass a prompt via -p and get structured output back:
claude -p "Fix the login bug in auth.ts" \
--output-format stream-json \
--max-turns 50This is how many custom agents call Claude. You spawn the CLI as a subprocess, parse the JSON stream, and extract the results. It works well for systems where you want the full agent capability without managing the loop yourself.
The Agent SDK gives you the same agent loop and tools as Claude Code, exposed as a Python or TypeScript library. You import it, configure your tools and permissions, pass a prompt, and receive a stream of messages as Claude works through the task.
from claude_agent_sdk import query, ClaudeAgentOptions
async for message in query(
prompt="Find all TODO comments and create issues",
options=ClaudeAgentOptions(
allowed_tools=["Read", "Glob", "Grep", "Bash"],
max_turns=50
)
):
print(message)The SDK handles tool execution for you. You declare which tools are available, Claude decides when to call them, and the SDK executes the call and feeds the result back into the conversation. You never write the tool loop manually.
Every conversation is persisted as a session on disk (JSONL files at ~/.claude/projects/). You can resume a session later with full context, which means the agent picks up where it left off with complete memory of what it did and found. This is useful for long-running tasks and for maintaining conversation continuity across multiple interactions.
The SDK supports spawning subagents for subtasks. You define named agents with their own system prompts, tool access, and model preferences. Each subagent runs in isolation with its own context. They can run in parallel, and their results return to the parent as a single message. This keeps the parent context clean and prevents different subtasks from interfering with each other.
Hooks are callbacks that fire at specific points in the agent loop: before a tool executes, after it returns, when a subagent starts or stops, before context compaction. They give you control over the agent's behavior without modifying the loop itself. Typical uses include audit logging, blocking writes to production paths, injecting extra context, and forwarding status updates to external systems.
MCP is an open standard for connecting agents to external tools and data sources. An MCP server exposes a set of tools (like "list GitHub issues" or "query a database"), and any MCP-compatible agent can discover and call them. The naming convention follows the pattern mcp__<server>__<tool>.
Three transport methods are supported:
MCP tools require explicit permission, the same as built-in tools. You can allow all tools from a server with a wildcard (mcp__github__*) or grant access individually.
The SDK evaluates permissions in a specific order, and each layer can override the ones below it:
This layered design lets you set restrictive defaults and selectively open things up, or set permissive defaults and block the dangerous operations with deny rules and hooks. Most production agents use a combination.
The SDK and CLI access the same underlying agent. The choice depends on what you need. The CLI is well-suited for interactive use, one-off scripts, and automations where you want to call Claude as a subprocess and parse the output. The SDK is better when you need programmatic control over sessions, custom tool definitions, hook-based interception, or tight integration with an existing application.
Many teams start with the CLI and move to the SDK as their requirements grow. Both approaches are valid, and the concepts transfer directly between them.
-p flag) is how many custom agents call ClaudeThe concept of an autonomous software agent predates most of the technology we associate with AI today. What we now call "AI agents" draws on work from automated planning, behavior-based robotics, multi-agent coordination, reinforcement learning, and natural language processing. Each era contributed ideas that the next one absorbed, often without acknowledging the debt.
John McCarthy proposed the Advice Taker in 1958, describing a program that could receive knowledge, reason about it, and decide what to do. The system was never built, but the proposal established what an agent should look like: goal-directed, knowledge-driven, and capable of acting on its own conclusions.
The first working planner was STRIPS, developed at the Stanford Research Institute in 1971 by Fikes and Nilsson. STRIPS took a goal and a set of possible actions, each defined by preconditions and effects, and produced a plan to achieve the goal. Its action representation became the standard for planning research and influenced agent design for the next three decades.
Terry Winograd's SHRDLU (1972, MIT) demonstrated natural language understanding within a simulated blocks world. You could tell it to move objects and it would comply. The system worked well within its narrow domain. Outside that domain, it failed completely. This brittleness, the inability to generalize beyond hand-coded rules, became the defining limitation of the symbolic era.
Expert systems followed in the late 1970s and 1980s. MYCIN (Stanford, 1976) diagnosed bacterial infections. R1/XCON (DEC, 1980) configured computer hardware orders. Both were commercially successful. Both required extensive manual rule authoring to build and maintain. Scaling them to new domains was expensive and error-prone, which is a large part of why the field eventually moved on.
Rodney Brooks at MIT challenged the symbolic approach directly. His subsumption architecture (1986) argued against internal world models entirely. Instead, agents should be built as layered sets of reactive behaviors, where each layer handles a specific concern (avoid obstacles, wander, explore) and complex behavior emerges from the interaction of simple rules. He demonstrated this on physical robots.
Running in parallel was the BDI model (Belief-Desire-Intention), which took the opposite approach. Originally a philosophical framework from Michael Bratman (1987), it was formalized computationally by Rao and Georgeff in 1991. BDI agents maintain beliefs about the state of the world, desires they want to achieve, and intentions they have committed to pursuing. The model provided a structured way to think about agent decision-making and is still used in multi-agent system design.
Wooldridge and Jennings published their survey "Intelligent Agents: Theory and Practice" in 1995. It defined the properties that distinguish an agent from ordinary software: autonomy, reactivity, pro-activeness, and social ability. These definitions became the standard vocabulary for the field. Most current descriptions of what makes an "agent" an agent trace back to this paper.
With individual agent architecture better understood, the research community turned to how agents work together. FIPA (Foundation for Intelligent Physical Agents) standardized inter-agent communication, defining an Agent Communication Language with message types drawn from speech act theory: request, inform, propose, accept, reject.
JADE (2000), a Java framework implementing FIPA standards, became the dominant platform for multi-agent research in academic settings. The era also produced work on auction mechanisms for resource allocation, contract nets for task distribution (originally proposed by Smith in 1980), and cooperative problem-solving protocols.
Much of this coordination research sat in academic papers for two decades. LLM-based multi-agent frameworks have been rediscovering these patterns since 2023, sometimes explicitly and sometimes by reinventing them from scratch.
Deep reinforcement learning shifted the approach from hand-coded rules to learned behavior. Instead of programming what an agent should do, you define a reward signal and let the agent discover the strategy through trial and error.
DeepMind's DQN (Mnih et al., 2013 preprint, 2015 Nature publication) learned to play Atari games at superhuman level by training a neural network directly on pixel inputs. No domain knowledge, no handcrafted features. The network learned which actions maximized the score through repeated play.
AlphaGo (Silver et al., 2016) defeated the world champion Lee Sedol 4-1 in Go by combining Monte Carlo Tree Search with deep neural networks trained through self-play. AlphaZero (2017) generalized the approach to chess, shogi, and Go, learning entirely from self-play with no human game knowledge at all.
By 2019, OpenAI Five had beaten world champions at Dota 2 and DeepMind's AlphaStar reached Grandmaster level in StarCraft II. These results demonstrated that learned agents could handle real-time, incomplete-information, multi-player environments.
The limitation was generality. Each system was trained for a single game or environment. A general-purpose agent that could handle arbitrary tasks remained out of reach.
Large language models made general-purpose agents possible by providing a reasoning engine that already understood a wide range of tasks, domains, and concepts. The remaining problem was connecting that reasoning to action.
Yao et al. solved this in their October 2022 paper "ReAct: Synergizing Reasoning and Acting in Language Models." The key insight was that LLMs could interleave chain-of-thought reasoning with tool-use actions in a loop. The model reasons about what to do, calls a tool, reads the result, and reasons again. This pattern became the foundation of every major agent framework that followed.
Schick et al. at Meta published Toolformer in February 2023, showing that LLMs could learn when and how to call external APIs autonomously, without step-by-step instruction.
Auto-GPT launched in March 2023 and reached over 100,000 GitHub stars within weeks. It attempted fully autonomous task execution with GPT-4. Practical reliability was poor, but the project demonstrated that the concept of an autonomous AI agent had broad public appeal.
OpenAI introduced function calling in their API in June 2023, providing a structured mechanism for LLMs to invoke tools. This replaced the fragile prompt-based tool invocation that earlier agents relied on and made tool use far more reliable.
Through 2024 and 2025, the ecosystem matured. Anthropic released Claude Code as a production coding agent. Frameworks like LangChain, CrewAI, and AutoGen made multi-agent orchestration accessible to a wider range of developers. Microsoft's Magnetic-One and Anthropic's computer use capabilities pushed agents toward general-purpose computer interaction.
The current generation of agent architectures combines ReAct-style reasoning loops, tool use, persistent memory, and multi-agent coordination. Open problems include reliability on multi-step tasks, evaluation methodology for open-ended work, and the question of how much real-world authority it is appropriate to delegate to a model.
Modern LLM agents carry forward ideas from every prior era. Planning from STRIPS. Goal-directed behavior from BDI. Coordination protocols from the multi-agent systems community. Learned strategies from reinforcement learning. The language model unifies these as a general-purpose reasoning engine that can be connected to tools and data sources.
Previous agent generations were each built for a single task or domain. LLM agents get broad reasoning capability from the pre-trained model and acquire domain-specific capability through tools. That combination is new. The individual components are not.
Ask ten researchers what "trust in AI" means and you'll get twelve definitions. Some treat it as a measurable cognitive state. Others call it a social relationship. A few insist it's just risk tolerance dressed up in fancier language. This confusion isn't academic navel-gazing. It matters because how you define trust determines what you try to fix.
The practical version: when should a person follow an AI's recommendation, and when should they override it? Get this wrong in one direction and the doctor rubber-stamps a bad diagnosis. Get it wrong in the other and the pilot ignores a system that's trying to save the plane.
Both of those things happen. They happen a lot.
The research community has two terms for the failure modes. Automation bias is when people defer to an AI even when it's wrong. Algorithm aversion is when they refuse to follow an AI even when it's right. Both are well-documented, and the frustrating finding is that they're both common, sometimes in the same person, depending on the task.
Parasuraman and Manzey laid out the automation bias framework back in 2010, and it's held up. People skip their own analysis when a machine gives them an answer. It's worse when the system is usually right, because you learn to stop checking. Then the one time the system fails, you miss it.
Algorithm aversion goes the other way. Dietvorst, Simmons, and Massey at Wharton showed that people who watch an algorithm make a mistake will sometimes abandon it entirely, even if it still outperforms their own judgment. One wrong answer and they're done. This happens more in subjective domains. People tolerate bad AI predictions about weather but not about hiring or criminal sentencing.
De Vries and colleagues found that the split is domain-dependent. Quantitative tasks push people toward automation bias. Moral and subjective tasks push them toward aversion. The same person might blindly follow a navigation app and refuse to trust a medical AI on the same afternoon.
The goal the field keeps chasing is calibrated trust, where a person's confidence in an AI matches the AI's actual reliability. Almost nobody achieves this naturally.
For years, the pitch went like this: if we just explain what the AI is doing, people will trust it the right amount. Make the black box transparent and the calibration problem goes away.
The research has not been kind to this idea.
Kaur et al. (2020, CHI) studied data scientists using interpretability tools. These were technical people who should have known better. They over-trusted the explanations. They saw feature importance charts and treated them as ground truth without questioning whether the explanations were faithful to what the model actually did.
Chromik and Butz described what they called "dark patterns of XAI," where explanations create a false sense of understanding. A plausible-sounding explanation increases trust even if it's wrong. Jacovi and Goldberg formalized this distinction between "faithfulness" (the explanation actually matches the model's reasoning) and "plausibility" (the explanation sounds reasonable to a human). These two properties are independent. An explanation can sound great and be completely misleading.
Schemmer and colleagues found that for simple tasks, adding explanations sometimes made things worse. People spent cognitive effort parsing the explanation instead of evaluating the prediction. Performance went down.
The current view is more nuanced. Explanations can help, but only when the user has enough expertise to evaluate them, the explanation type matches the task, and the explanation is actually faithful. That's a narrow set of conditions. For most real-world deployments, slapping an explanation on an AI output and hoping for the best is not a strategy.
Most trust research before 2023 studied traditional decision-support systems. An AI makes a recommendation, a human accepts or rejects it. The interaction is structured and bounded.
LLMs broke that model. When the AI talks to you in natural language, uses your name, responds to your emotions, and sounds confident about everything, the trust dynamics shift from analytical to social. People start applying interpersonal trust heuristics. They evaluate the AI the way they'd evaluate a colleague.
This has specific consequences. Anthropomorphism effects are significant. The more human-like the interaction feels, the more people rely on social cues (politeness, confidence, fluency) instead of accuracy cues (track record, uncertainty signals, domain expertise). A smooth, articulate wrong answer is more persuasive than a halting, hedging right one.
Jakesch et al. at Cornell studied how AI-generated opinions shift human attitudes. The effect is real and measurable. When an LLM presents a position confidently, people adjust their views toward it, even on subjective questions where the model has no special authority.
Sycophancy makes this worse. Models trained with RLHF have a tendency to agree with the user, tell them what they want to hear, and validate their existing beliefs. Anthropic's own research has flagged this as a trust calibration risk. If the AI always agrees with you, you never get the corrective signal that builds appropriate calibration.
AI systems fail. The question is what happens to trust afterward.
De Visser and colleagues established the basics of trust repair in automation. The most effective strategy is acknowledgment combined with explanation combined with corrective action. Just saying sorry doesn't work as well with machines as it does with people. You need to show you understand what went wrong and that you've changed.
Esterwood and Robert at the University of Michigan added an important distinction. Competence violations (the AI couldn't do the task) are easier to repair than integrity violations (the AI seemed to be deceptive or biased). If your AI system gives a wrong medical diagnosis, you can rebuild trust by improving accuracy. If your AI system appears to be biased against certain patients, the damage is much harder to undo.
LLMs introduce a new dimension here. Because they can talk about their failures in natural language, they have a richer repair toolkit than traditional automation. An LLM that explains its mistake and adjusts its approach in real-time can recover trust faster than a system that just outputs a corrected number. But this also means an LLM that confabulates an explanation for why it was wrong, without actually understanding the failure, can fake repair in ways that are hard for users to detect.
Individual differences matter. Not everyone responds to AI the same way.
Personality is a factor but a weaker one than you might expect. Openness to Experience correlates with higher AI trust. Neuroticism correlates with lower trust. Agreeableness is associated with more deference to AI recommendations. These effects are consistent but modest in size.
Technical literacy is a stronger predictor than any personality trait. People who understand how these systems work tend to have better-calibrated trust, neither too high nor too low. This isn't because they trust more or less. It's because they have a framework for evaluating when trust is warranted.
Age and generational effects are messy. Some studies find older adults show higher automation bias. Others find them more aversion-prone. The inconsistency probably reflects different tasks and contexts more than a real generational pattern.
The hardest trust problems show up when AI is a teammate, not just a tool.
McNeese, Demir, and Cooke at Arizona State have studied human-AI teaming extensively, much of it funded through DARPA and Army Research Lab programs. Their central finding: effective human-AI teams need shared mental models. Both the human and the AI need some understanding of what the other is doing and why. But opaque AI teammates prevent this. You can't build a shared model with something you can't read.
Healthcare provides the clearest examples. Cai et al. at Google studied AI-assisted pathology and found that AI recommendations anchored physician diagnoses. When the AI highlighted a region of a scan, doctors focused on that region, even when the critical finding was somewhere else. The AI was right often enough that this usually helped. But when it was wrong, the anchoring effect meant doctors missed things they would have caught on their own.
Driving automation shows the trust handoff problem. The SAE levels of automation assume clean transitions between human and machine control. In practice, the handoff moment is the most dangerous part. A driver who has been monitoring their phone while the car drives itself needs several seconds to regain situational awareness. Trust in the automation makes this worse, because higher trust means less monitoring and slower handoff.
The field is moving away from treating trust as a property of individuals. The older framing was cognitive: trust is a mental state, measured through surveys, predicted by personality traits. The newer framing is relational and systemic. Trust is something that emerges from the interaction between a person, a system, a task, an organization, and a context.
This matters for design. If trust is cognitive, you fix it by giving people better information (explanations, confidence scores, accuracy metrics). If trust is relational, you fix it by changing the interaction (how failures are communicated, how control is shared, how the AI presents uncertainty, what social cues it sends).
The biggest open question is whether the social trust dynamics of LLMs are a feature or a bug. Conversational AI makes interactions more natural and accessible. It also exploits social heuristics that evolved for human relationships and don't necessarily transfer well to software. We trust confident humans because confidence usually correlates with knowledge. We trust agreeable humans because agreeableness usually signals honesty. Neither of these correlations holds for LLMs.
The research community hasn't resolved this. I don't think they will anytime soon. But the work is getting sharper, the measurements are getting better, and the stakes keep going up as these systems take on more responsibility.
On April 20, 2026, Apple announced that John Ternus would become its next CEO, effective September 1. Tim Cook, who has run the company since August 2011, moves to executive chairman. Arthur Levinson becomes lead independent director. Johny Srouji, who has run silicon since 2008, takes the new title of Chief Hardware Officer and reports to Ternus.
The transition was visible in the reporting before the announcement. Mark Gurman at Bloomberg named Ternus as the leading internal candidate as early as 2024, repeated the call through 2025, and ran a long Businessweek profile in March 2026 that read in retrospect as a soft launch. Outside the Apple analyst circuit, almost nobody knew who he was. Inside the company, he had been running the organization that ships the products that pay for everything else.
Ternus graduated from the University of Pennsylvania in 1997 with a degree in mechanical engineering and applied mechanics, and a minor in psychology. He swam competitively for Penn's men's team. His senior project was a mechanical feeding arm for individuals with quadriplegia, operable by head movement.
His first job was at Virtual Research Systems, a small VR-headset company in Northern California, where he spent roughly four years as a mechanical engineer. He joined Apple in 2001 on the Product Design team. His first project was the Apple Cinema Display.
By 2013 he was vice president of hardware engineering, reporting to Dan Riccio, with AirPods, Mac, and iPad in his portfolio. Around 2020 he picked up iPhone hardware. On January 25, 2021, Apple promoted him to senior vice president of hardware engineering when Riccio moved to a new project that became the Vision Pro. Apple Watch hardware moved into his org in late 2022. In late 2025, Cook handed him oversight of industrial design, which Bloomberg read as the signal that succession was being executed rather than merely planned.
Ternus's track record is unusually heavy on hard, expensive transitions, and most of them landed.
The Apple Silicon transition for the Mac began in November 2020 with the M1. Hardware engineering owned the integration of the new chip across the laptop and desktop lineup, which meant rebuilding thermal systems, board layouts, and product geometry across every Mac in parallel. Apple completed the transition in June 2023 with the M2 Ultra Mac Pro, ahead of the original two-year commitment. Mac revenue grew through the period instead of dipping the way platform shifts often cause.
The iPad Pro lines through 2018, 2020, 2021, 2022, and 2024 went through Ternus's organization. The May 2024 M4 iPad Pro is one of the thinnest mainstream computing devices ever shipped at 5.1 millimeters. Fitting a tandem OLED display, an active thermal structure, and an M-series chip into that volume is the kind of problem that surfaces every weakness in a hardware org. It worked.
The Vision Pro, announced at WWDC 2023 and shipped in February 2024, was developed under Riccio's separate group but transitioned into Ternus's organization for production scaling. The product launched on schedule. Sales sit below the most optimistic projections, but the hardware itself shipped as promised, which is the engineering bar.
AirPods, including the AirPods Pro 2 and the hearing-aid functionality the FDA cleared in September 2024, came out of his organization as well.
The clearest expression of Ternus's philosophy is a story he told at the University of Pennsylvania School of Engineering commencement on May 18, 2024. He had been at Apple a few months. He flew to a supplier to inspect production of the Apple Cinema Display. The spec called for screw heads with a specific groove pattern, a cosmetic detail nobody outside Apple would ever notice. The supplier had cut a different number than the spec called for.
The supplier's argument was reasonable. Nobody can see the difference. Nobody is going to count grooves on a screw with a magnifying glass. The product will ship on time if you accept this.
Ternus, late at night in the factory, took out a magnifying glass and counted. The grooves got fixed.
"The care that you put into your work really matters," he said in the speech. The line is the kind of thing CEOs say at commencements. The story underneath it is the thing worth absorbing. Care is not a virtue that lives in mission statements. It lives in the moment somebody chooses to push back on a supplier at 11 PM on a detail that nobody will notice. Apple has a culture of that choice. Ternus has been making it for twenty-five years.
The 2024 commencement speech is the closest thing to a public statement of Ternus's operating philosophy. Four lines from it are worth pulling out.
"Always assume you're as smart as anyone else in the room, but never assume that you know as much as they do." Confidence in your reasoning, humility about your information. The two have to coexist or the work breaks. Engineers who lack the first defer to authority and ship bad designs. Engineers who lack the second talk over experts and miss the actual problem.
"The care that you put into your work really matters." Stated above. The proof is the supplier visit, not the line.
"Build what interests you, build what excites you, but above all else, build it in a way that aligns with your values." A version of integrity-as-constraint. Not a values statement; a build constraint. If the project requires you to compromise something you care about, the work will not survive contact with the next decade.
"Go out there and make a dent in the universe." The Jobsian closing line. Worth noting that Ternus chose it. He works in a register that is generally calm and technical, and he ended his most public speech with a Jobs quote. The Apple cultural memory is intentional, not residual.
Profiles published around the CEO announcement converge on a few traits.
He sits in open-plan engineering space rather than an executive office. Bloomberg's Gurman has noted this several times. The choice is not stylistic; it shapes the information he gets. Engineers who walk past him deliver problems faster than engineers who have to schedule a meeting.
He is described, almost universally, as a "super nice guy" who is also decisive. The two qualities are usually treated as opposites in tech leadership. They are not. The combination is what allows technical organizations to run at high velocity without the political overhead that slows most large companies. People bring him bad news because they trust the response. He acts on the bad news because he understands what it means.
His outside life is engineering-shaped. He cycles. He races a Porsche at Laguna Seca and has been reported to take off-road rally trips in Washington state with co-workers. His LinkedIn account is publicly empty. None of that is decoration. It signals a person whose identity is in the work, not in the visibility around the work.
Ternus takes over an Apple that has not had a major new product category land at scale since the Apple Watch in 2015. The Vision Pro shipped on time but has not produced a mass market. Apple Intelligence, the 2024 AI strategy, has been criticized as late and partial. The next two years of his tenure are likely to be defined by three programs: a foldable iPhone (Gurman has reported September 2026), AI-driven smart glasses, and the rebuilt Siri/Apple Intelligence stack.
The pattern that should worry watchers is that Ternus is, by background, a hardware engineer running a company whose competitive position depends on AI software. That is the same shape as the Sundar Pichai gap in 2015 (operations background, AI emerging) and the Satya Nadella gap in 2014 (cloud background, AI emerging). Both succeeded by hiring the missing capability and giving it real authority. Ternus has the new Chief Hardware Officer slot for Srouji and existing services and software leaders. The test is whether the AI work gets the same kind of seat.
Three patterns are visible in the way Ternus reached this job, and they are useful for anyone building hardware.
The first is tolerance for long timelines. Apple Silicon had been in development for more than a decade before the M1 shipped. The Vision Pro took roughly seven years from project start. Ternus did not start either program, but he ran both through their hardest phases. Keeping a multi-year program funded and on track without visible course correction is rare, and it is the discipline that hardware careers compound around.
The second is product judgment under cost pressure. The M-series chips replaced Intel CPUs across the Mac line at margins Apple would not have accepted with an outside silicon provider. The decision to use tandem OLED on the iPad Pro, with the cost premium that involves, was made by hardware engineering working with operations and design. Ternus has owned several of these calls and the calls have been right.
The third is presentation discipline. Ternus has presented at WWDC and at September iPhone events repeatedly since 2017. He is calm, technical, and unembellished. He does not perform. That register matches what Apple sells, and it matches the Cook era's tone. Steve Jobs's successor needed to be operational rather than charismatic. Cook's successor needed to be technical rather than operational. Apple has been preparing for this for fifteen years.
It is easy to read a career as inevitable once it is done. John Ternus joined Apple in 2001 and became its CEO in 2026. The line connecting those two points looks straight. It was not. Five decisions, made over twenty-five years, are visible in retrospect as the load-bearing ones. None of them were obvious at the time.
The decisions are not unique to Ternus. They show up in the careers of most people who run hardware organizations at scale. The reason to study them through one person is that the alternative is to study them as abstractions, which is how they become useless.
Ternus spent his first decade at Apple inside iPad hardware engineering. iPad was, for most of that time, the second-class Apple product. iPhone got the attention, the keynote slots, and the executive promotions. iPad got the sustaining engineering work and the question of whether it was a real computer.
The decision to stay on iPad through that period is easy to miss because it does not look like a decision. It is not glamorous to spend a decade making a product slightly thinner and slightly faster than the version before it. The work is also exactly what builds a hardware engineering career. You learn how the supply chain actually behaves under volume, where the thermal and structural margins are, what suppliers will and will not do for you. Those lessons do not transfer from a deck. They transfer from shipping the same product five times.
The pattern is general. The product line that gets less attention is usually the one where you have room to make decisions, run experiments, and develop judgment. The flagship product is run by people who have already been promoted. If you want to build the kind of resume that gets you to the flagship, you build it on the secondary line first.
The standard arc for an engineering manager is that the technical depth atrophies as the team grows. The job stops being engineering and starts being calendars, headcount, and politics. By the time you are running a few hundred people, you have not opened a CAD file in years.
Ternus's career runs against this. By every account from people who have worked with him, he stays in the technical detail of the products his organization ships. He is reported to walk a factory line, count features on a part, and know what the tolerances are. He sits in open engineering space, not an executive office. The choice shapes the kind of information he receives. Engineers who walk past him deliver problems in real time. Engineers who have to schedule a meeting deliver problems on a delay.
The reason this matters for hardware specifically is that the failures in hardware are physical. They are about a thermal margin that is two degrees too tight, a connector that fails at the eight-thousandth insertion, a yield curve that flattens at 92 percent when the plan needed 96. You cannot run a hardware organization on a dashboard. You can run a hardware organization on a dashboard only if you already know what the numbers on the dashboard mean. That knowledge atrophies if you stop using it.
Apple Silicon for Mac was a public commitment in June 2020. The work that made the M1 possible was at least a decade old by then. The Vision Pro was a project from roughly 2016. Both programs had a structural feature that makes them rare in industry: they ran for years without visible course correction.
Most large engineering programs do not run that way. They get started, hit a hard problem in year two, get scope-cut in year three, and either ship as a diminished version of themselves or get cancelled. The reason is that organizations under quarterly pressure cannot fund work that will not produce revenue for five or seven years. The discipline to keep a multi-year program intact is usually a CEO-level discipline. Tim Cook had it. Ternus inherited it as a hardware leader and demonstrated it on programs that he did not start.
The general lesson is that hardware careers compound on long programs. A one-year project teaches you how to ship. A three-year project teaches you how to manage scope. A seven-year project teaches you how to keep a team intact through several Hype Cycles, several budget reviews, and the one moment in year four when the program is six months from cancellation and you have to decide whether the whole bet is still right. The judgment that comes out of that experience is not available any other way.
Apple keynotes are extremely high-cost productions. The presenters rehearse for weeks. The product reveals are choreographed, and the language is written and re-written and tested. Ternus has been on those keynotes since 2017. He started with side segments, moved to product reveals, and by 2024 was anchoring the May "Let Loose" iPad event.
The decision to learn how to present is not the same as the decision to seek the spotlight. Ternus is reported as deeply private. The Bloomberg profile noted his LinkedIn account has no posts. He does not perform on stage. What he does is communicate technical decisions clearly to a non-technical audience, in a register that does not flatter the audience and does not condescend.
That skill matters because the alternative is that someone else explains your work for you. If your products are technical and the people who explain them publicly are not, the explanations drift. The way Apple talks about its products is a function of the fact that the people who built them are also the people who present them. That is a defensible position only if the engineers can hold an audience.
The Bloomberg profile, the Fortune profile, and the PBS NewsHour piece all converge on the same observation: Ternus has been one of the most senior people in tech for half a decade and almost no one outside the industry could pick him out of a lineup. His public profile is engineered to be small. His LinkedIn is empty. He gives few interviews. His one widely circulated public talk was a commencement speech to engineering students at his own alma mater.
The pattern is deliberate, and it is a contrast to the reigning model of executive visibility, which assumes that being known is a job requirement. It is not. Apple's leadership style under Cook treats personal visibility as a cost rather than a benefit, on the theory that the products are what the company is supposed to be known for. Ternus inherits that style and extends it.
The lesson for engineers building careers is that visibility and credibility are not the same thing. The market for visibility is crowded and noisy and often rewards shallow work. The market for credibility is smaller, slower, and rewarded almost entirely on whether the things you ship work. Pick which market you want to compete in. The two are not always compatible.
On June 22, 2020, at WWDC, Tim Cook announced that Apple would replace Intel processors with Apple-designed silicon across the entire Mac lineup, and that the transition would be complete in two years. Five months later the M1 MacBook Air, MacBook Pro, and Mac mini shipped. By June 2023, with the M2 Ultra Mac Pro, the transition was done. Apple beat its own deadline by roughly six months.
The bet looks obvious in retrospect. It was not obvious at the time. Intel had been the heart of the Mac since 2006. Several previous attempts to replace it, including the 2008 PA Semi acquisition that started Apple's silicon program, had taken more than a decade to mature. The decision to bet the entire Mac line on internal silicon was the largest platform transition Apple had executed since the move from PowerPC to Intel itself.
The technical surface of the problem is the chip. The actual surface is everything else.
Every Mac SKU has its own thermal envelope, board layout, port complement, and product geometry. Replacing the CPU means rebuilding the thermal solution and the board for every product, in parallel. The MacBook Air went fanless. The 14-inch MacBook Pro got a new chassis. The 24-inch iMac became thin enough that the logic board had to move into the chin rather than behind the display. None of these mechanical changes were optional. They were the consequence of a chip that ran much cooler than the Intel parts it replaced.
The software surface was harder. macOS had to run on the new architecture without breaking the existing application library. Apple's answer was Rosetta 2, a binary translation layer that ran existing x86 software on the new chips at performance levels that, for many applications, exceeded native execution on the Intel Macs they replaced. Rosetta 2 was the bridge that made the transition tolerable for users. Without it, the platform shift would have looked, from the outside, like a reset.
The supply surface was hardest. Apple's silicon is fabricated by TSMC in Taiwan. The transition committed Apple to TSMC at lead-customer status on every advanced node from N5 forward. That is a single point of failure for the company's most strategic product line, in a region with significant geopolitical risk. The decision to accept that risk was not made lightly. The alternative was to remain dependent on Intel, which by 2018 was several process nodes behind TSMC and had no credible plan to catch up.
The transition is studied as a case for two reasons. The first is that it landed. The second is that it landed without the financial damage that platform shifts usually cause.
Mac revenue grew during the transition. This is the unusual outcome. Most platform shifts produce a revenue trough as customers wait out the early generation of the new platform. The Mac transition did not produce a trough because Apple managed three things in parallel.
It managed the message. The June 2020 announcement included a precise commitment ("two-year transition") and a clear product cadence. Customers knew what was coming and when. The uncertainty that usually depresses sales during a platform shift was minimized because Apple gave the market a schedule and met it.
It managed the developer story. Rosetta 2 ran existing software acceptably from day one. Universal binaries, a development model Apple had used in the PowerPC-to-Intel transition fifteen years earlier, allowed developers to ship a single application that ran natively on both architectures. The transition cost to the developer ecosystem was lower than any prior platform shift in Apple's history.
It managed the product cadence. The first generation, the M1, was released in low-risk products: the MacBook Air, the entry MacBook Pro, the Mac mini. The professional products, where customers were most sensitive to a regression, came later. By the time the M2 Ultra Mac Pro shipped in 2023, the architecture had two and a half years of in-market validation.
The lesson most often drawn from Apple Silicon is about vertical integration. That lesson is real and worth its own treatment. The lesson less often drawn is about the discipline of running a long program.
The PA Semi acquisition was 2008. The first A-series chip, the A4, shipped in the iPad and iPhone 4 in 2010. The A-series architecture matured through ten generations before it crossed into the Mac. The decision to fund that work for twelve years, through three economic cycles and several leadership transitions, is what made the 2020 announcement possible. Apple did not announce the Mac transition and then build the chips. It built the chips for a decade and then announced the transition.
The discipline that this required is not technical. It is organizational. A program that takes twelve years to produce a strategic outcome is a program that has to survive twelve annual budget cycles, twelve sets of board questions about return on investment, and twelve opportunities for a competing executive to argue that the resources should go somewhere else. Most companies do not have the institutional patience for that. Apple does. The patience is partly Cook's, partly the cumulative result of cash reserves that absorb the question, and partly the cultural memory of having done it before with the original A-series program.
The version of this story that gets told is the success version. There is a parallel version where the transition went differently.
If Rosetta 2 had not been good enough, customer complaints in 2021 would have stalled adoption and given Microsoft Surface and the PC ecosystem an opening. If TSMC's N5 yields had been worse, the M1 ramp would have been gated on chip supply rather than on demand. If Intel had executed on Alder Lake and the rumored 18A node, the comparative advantage of Apple Silicon in 2024 would have been smaller, and the transition would have been a draw rather than a victory.
None of those things happened. They could have. The reason to keep them in mind is that the lesson "Apple Silicon was a great bet" is not useful in your own work. The useful lesson is that the bet was a high-variance one, and the variance was managed through a combination of long lead-time investment, technical conservatism in the user-facing layer (Rosetta 2), and a phased product introduction. Each of those is a decision that can be made, or not made, by other organizations facing platform transitions.
The biographies of Apple's leadership tend to start at Apple. The story of Tim Cook usually begins with the March 1998 hiring announcement and skips the eighteen years that came before. That is the wrong place to start if you want to understand why Apple operates the way it does.
Cook graduated from Auburn University in 1982 with a degree in industrial engineering. He earned an MBA from Duke in 1988. He spent twelve years at IBM in personal computer operations, finishing as director of fulfillment for the company's North American PC business. He moved to Intelligent Electronics for a brief tenure as chief operating officer of its reseller division, then to Compaq as vice president of corporate materials. He had been at Compaq for six months when Steve Jobs called him in early 1998.
The interview is part of Apple lore. Cook described it later as a meeting with someone whose mind was on the right side of the next decade. That description was specific. The "right side of the next decade" was not about products. It was about how Apple was going to operate.
Apple in early 1998 was approximately 90 days from bankruptcy by some estimates. The product line had collapsed under management. Inventory was a disaster. The company had warehouses of unsold Performa and PowerBook variants. Cash burn was catastrophic. The product strategy that Jobs would announce four months after Cook's hire (the iMac, the consumer-versus-pro grid, the dramatic SKU reduction) is the strategy that gets the historical credit. The operations work that made it possible is what Cook did first.
Within two years he had collapsed Apple's inventory from months on hand to days. He shut domestic warehouses, consolidated suppliers, and shifted final assembly to contract manufacturers concentrated in Asia, primarily Foxconn. The financial impact was direct: working capital that had been trapped in unsold inventory was released. Apple, which had been close to insolvent in 1997, was cash-flow positive by 1999.
The line that Cook is usually quoted as saying about that period is "inventory is fundamentally evil. You kind of want to manage it like you're in the dairy business. If it gets past its freshness date, you have a problem." The dairy framing is precise. Components age. A chip that is current today is a discount chip in six months and a write-off in twelve. The cost of holding inventory is not just the carrying cost. It is the depreciation of the components themselves.
The pattern that Horace Dediu at Asymco named the "Bank of Apple" is the second-order consequence of the operational discipline. Once Apple was generating consistent cash, Cook used the cash differently than most companies do.
Most companies use cash to fund operations and return the rest to shareholders. Apple used cash to pre-purchase manufacturing capacity. The pattern is well-documented. Apple identifies a critical component (advanced display glass, advanced camera sensors, custom audio chips, machine tools for milling unibody enclosures), pays a supplier for the tooling and capacity to produce that component at Apple-volume scale, and locks in exclusivity for a year or two before the technology becomes available to competitors.
The strategic effect of this pattern is that Apple's competitors face a one-year delay on every component that Apple chose to invest in. The financial effect is that Apple's cost structure on those components is lower than its competitors', because the tooling has been amortized against guaranteed Apple volume rather than speculative open-market volume.
The pattern requires capital reserves at a scale that almost no other hardware company has. The cash position that Apple uses to execute this strategy is itself a moat. A competitor who wants to copy the pattern needs not just the money but the willingness to deploy it at the same scale, and the supplier relationships to make the bets stick.
The classical example is the unibody MacBook in 2008. Apple bought the CNC milling machines, in volume, before the consumer electronics industry had decided that aluminum unibody was the design vocabulary it wanted. The decision committed Apple to a manufacturing approach that took years for competitors to copy and gave Apple a structural manufacturing advantage during that period.
The 2017 iPhone X OLED transition is another. Apple committed multi-year volume to Samsung Display for OLED panels at a scale that locked up the supply. The Android competition that wanted OLED panels for flagship phones during 2018 and 2019 had to take what Samsung had left. Apple's display strategy, which is generally credited as a hardware decision, was at least as much an operations decision.
The Apple Silicon transition is the most recent. Apple's status as TSMC's lead N-1 process customer is paid for in long-term volume commitments and capacity reservations that no other fabless company can match. The strategic effect is that Apple is consistently a process node ahead of its silicon competitors.
Most hardware companies cannot afford the Bank of Apple strategy. The capital requirement is too high. The lesson is not that you should pre-purchase tooling. The lesson is that operations is strategy, not back-office work.
The decisions that Cook made between 1998 and 2005 were operational on the surface. Inventory targets, supplier consolidation, contract manufacturing. They were strategic in their effect. They turned Apple from a company that was constrained by working capital into a company that could move faster than its competitors on every product transition. The operations work created the conditions under which the product work could be ambitious.
The pattern generalizes. If your operations are at parity with your competitors, your product strategy is competing on product alone. If your operations are better, you have more room to be ambitious on product. The relative size of that room is roughly proportional to the operational gap. This is a thing that engineering-driven companies under-invest in, because the returns are not visible in any single product cycle. They are visible in the third product cycle, when you can do something the competition cannot afford.
Ternus inherits the Cook operations machine. He did not build it, and the Bloomberg reporting suggests that the operations function will continue to report to Jeff Williams (and Williams's successor) rather than to Ternus directly. The structural risk is that an engineering CEO will, over time, treat operations as a service function and let the operational discipline atrophy.
The structural advantage is that Ternus has spent twenty-five years inside the engineering organization that depends on the operations work. He has been the customer of the Bank of Apple strategy on multiple programs. He understands what it produces. The optimistic case is that he extends the strategy by directing it at the next set of strategic components: AI accelerators, mixed-reality displays, miniaturized batteries. The pessimistic case is that he treats it as someone else's problem. The next two years will tell which.
Hardware ships through a sequence of phase gates that the industry has standardized over the last forty years. The vocabulary is not glamorous and it is mostly acronyms. The reason to learn it is that most of the failure modes in hardware development are failures to respect what each phase is for. People who skip a phase, or who treat a phase as a checkbox, are people whose products ship late or do not ship at all.
The general structure is concept, prototype, validation, ramp. The industry-standard names for the validation gates are EVT, DVT, PVT, and MP. Each one has a specific purpose, and the work that should happen at each gate is different.
EVT is the first build that uses production-intent parts in something close to the production process. The question EVT exists to answer is "does the design work as engineered?" Components arrive on a real PCB. The mechanical pieces are made on real tooling, even if the tooling is soft (aluminum) rather than hardened steel. Build volumes are typically thirty to a hundred units.
The work at EVT is functional verification. Power-on yield. Thermal performance under load. RF performance against the antenna design. Drop testing against the structural model. The output is a long list of bugs, both electrical and mechanical, that the engineering team has to triage and fix before the next gate.
The failure mode at EVT is treating it as a demo build. EVT is for finding problems, not for showing the product. If your EVT looks too clean, the team is probably hiding issues that will show up at DVT, where they are more expensive to fix.
DVT is the build where the design is locked. Components are at production specification. Tooling is hardened. The build is large enough to run the regulatory testing the product needs (FCC, CE, safety certifications) and to do reliability testing at statistical scale. Build volumes are typically a few hundred units.
The question DVT answers is "is this actually the product we are going to ship?" Cosmetics matter at DVT in a way they do not at EVT. The answer to a DVT issue can be a tooling change, a component substitution, or a process change at the contract manufacturer. The answer cannot be "we will fix it in the next build" because the next build is PVT, which is supposed to be the production process running at low volume.
The failure mode at DVT is scope creep. Engineering teams discover late issues and want to redesign components. Operations teams discover yield issues and want to add screening tests. Product teams discover use cases the spec missed. The DVT review is the hardest meeting in the program, because it is the meeting where everyone has to agree that the product is done. Most slipped programs slip here.
PVT is a low-volume run of the actual production process at the actual factory. Build volumes are typically a few thousand to ten thousand units. The question PVT answers is "can the line build this at the rate and quality we need?"
The work at PVT is line-level. How long does the assembly take? Where are the bottlenecks? What is the first-pass yield? What does the test process catch and what does it miss? Are the operators trained? Is the line balanced? PVT units are sometimes sold (often as "early access" or limited geographic launches) and the customer experience of those units is a real signal about whether the product is ready for general availability.
The failure mode at PVT is yield. The line builds the product, but the first-pass yield is too low. Maybe the design is marginal. Maybe the process is not capable. Maybe the test coverage is wrong. PVT is the gate where you find out if the production economics actually work, and it is the gate where most novice hardware programs discover that they have built a product that is not manufacturable at scale.
MP is the production ramp. The line is running at rate. The product is shipping. The question now is "is yield converging to the plan, and is the field reliability matching the lab data?"
The work at MP is sustaining engineering. Failure analysis on returns. Cost-down on the bill of materials. Yield improvement projects. Component qualification for second sources, because every component you single-source is a supply risk that will eventually bite you. The product engineering team starts to thin out as people move to the next program. The sustaining team takes over.
The failure mode at MP is field reliability. Issues that did not surface in lab testing show up in the field at low rates that are statistically significant when you ship a million units. The classic Apple example is the 2015-2019 butterfly keyboard, which passed its lab reliability testing and failed at unacceptable rates in the field. The lesson is not that lab reliability testing is bad. The lesson is that the field is a longer test than the lab and a noisier one, and that a hardware program is not actually done at MP. It is done six months after MP, when the early field data has stabilized.
Every hardware program eventually faces pressure to compress the phase gates. The pressure is structural. The product is late, the launch date is fixed, and the schedule has to come from somewhere. The temptation is to skip a phase or to combine two of them.
This rarely works, and the reason is that each phase exists to find a specific class of problem. EVT finds engineering problems. DVT finds design problems. PVT finds manufacturing problems. MP finds field problems. If you skip EVT, the engineering problems show up at DVT, where the tooling is harder and the changes are more expensive. If you skip DVT, the design problems show up at PVT, where the line is committed and the changes ripple through the manufacturing process. If you skip PVT, the manufacturing problems show up in MP, where you are now finding out about them at scale, with units in customer hands.
The phase gates are not a process tax. They are a sequence in which problems are cheapest to fix. Skipping one moves the problems to the next gate at higher cost. The good hardware programs are the ones where each gate is treated as an opportunity to find problems early, not as a milestone to hit on the schedule.
The clearest statement of Apple's vertical integration doctrine is one Tim Cook has repeated, in some form, on multiple earnings calls and at multiple shareholder meetings. The version Asymco's Horace Dediu has documented is: "We believe that we need to own and control the primary technologies behind the products we make, and participate only in markets where we can make a significant contribution."
Two phrases in that sentence are doing the work. The first is "primary technologies." The second is "significant contribution." Neither is intuitive. Both are testable.
A primary technology is one whose performance shows up in the user's experience of the product. A secondary technology is one that does not. The distinction matters because vertical integration is expensive. You do not integrate everything; you integrate the parts that differentiate the product.
Apple integrates silicon (A-series, M-series), operating systems (iOS, macOS), display technology (custom panels with custom drivers), camera systems (custom image signal processors and computational photography), and audio processing (the H-series chips in AirPods). All of these show up in the user's experience as performance, battery life, image quality, or audio quality.
Apple does not integrate the things that do not. The aluminum that goes into the unibody chassis comes from outside suppliers. The OLED panels (until recently, all of them) came from Samsung Display and LG Display. Most of the discrete components on the logic board, the connectors, the screws, the packaging, the shipping logistics, the retail point-of-sale software, all of it is bought from someone else. The principle is consistent. Integrate where the integration changes what the user sees. Buy where it does not.
Vertical integration imposes costs that are easy to underestimate. Ben Thompson at Stratechery has framed these costs as the "integration tax," and the framing is useful.
The first cost is capital. Designing your own silicon requires a chip team. A chip team is hundreds of people across architecture, RTL, verification, physical design, and software. Apple's silicon team has been described as one of the largest in the industry. The salary cost alone runs into the billions per year. The compounding cost is that the team has to keep producing competitive silicon every year. There is no point at which you have "finished" the chip.
The second cost is opportunity. Every component Apple integrates is a component Apple has to keep up with. The pace of innovation in silicon, displays, batteries, and image sensors is set by the merchant suppliers (TSMC, Samsung, Sony) competing with each other. An integrated company that falls behind the merchant pace finds itself in the worst position: paying integration costs while shipping inferior components. This is the position Intel found itself in by 2018, and it is the structural risk for any vertical integrator.
The third cost is flexibility. A company that buys components can switch suppliers when a better one emerges. A company that integrates is committed. If your custom modem is two generations behind, you cannot switch to Qualcomm without abandoning years of internal investment. Apple has paid this tax visibly in modems, where the C1 and C2 internal modem programs have lagged the merchant Qualcomm parts for several generations.
The economics of vertical integration are roughly as follows. Integration pays when the integrated component is materially better than the merchant alternative, and when "materially better" shows up in the user experience in a way the user values, and when the volume justifies the fixed cost of the integration program.
The Apple Silicon Mac transition is the canonical case where all three conditions held. The merchant alternative was Intel, which had stalled at 14 nm for several years. The user-visible difference was battery life that doubled and performance per watt that exceeded anything in the laptop market. The volume was the entire Mac line, which is around twenty million units per year. The integration paid.
The custom display drivers in the iPhone are another case where the conditions held. The user-visible difference is color accuracy, refresh rate management, and display power consumption. The volume is enormous (200+ million iPhone units per year). The merchant alternative was generic display controllers that did not exploit the panel capabilities. Apple integrated, and the iPhone display experience has been a competitive advantage for ten years.
The custom modem is the case where the conditions did not hold. The user-visible difference is small. Modem performance is increasingly invisible to the user as long as it is good enough. The merchant alternative (Qualcomm) is excellent. Apple has been pushing on internal modems for almost a decade and has not yet shipped one that is materially better than what it could buy. The integration is being pursued anyway, on the theory that it will eventually pay. That theory has been expensive to maintain.
The second phrase from Cook's doctrine is "significant contribution." The test it implies is: do not enter a market where you cannot offer something materially different. This sounds like a marketing test. It is actually an integration test.
The reason is that "materially different" in a hardware market almost always requires integration somewhere in the stack. Apple did not enter the music player market until it could integrate iTunes with the iPod. It did not enter smartphones until it could integrate iOS with the silicon and the touch controller. It did not enter wireless audio until it could integrate the H1 chip with iOS pairing. The pattern is consistent. The market entry is gated on having something integrated to bring.
The corollary is that markets where Apple cannot offer integration get skipped. Apple has, at various points, looked at and declined to enter televisions, cars (Project Titan, ultimately cancelled in 2024), and home robotics (so far). The pattern in each case is that the integration story did not close. Without integration, Apple would have been a premium-priced commodity entrant. The discipline to skip those markets, even when there was internal momentum to enter them, is part of what the doctrine produces.
Most hardware companies cannot vertically integrate at Apple's scale. The capital is not there. The volume is not there. The talent is not there. The lesson is not that you should integrate. The lesson is that you should be specific about which layer of your stack actually differentiates your product, and integrate that layer if you can.
For most small hardware companies, the differentiating layer is software, firmware, and industrial design. The components are bought. The PCB is designed by an outside firm. The contract manufacturer handles assembly. The integration story is the user experience that the company controls directly. That is a defensible position if the user experience is genuinely better than what a competitor with the same components could ship. It is not defensible if the user experience is generic. The discipline is the same as Apple's, scaled down: identify the primary technology, own it, and buy everything else.
Supply chain is usually treated as plumbing. The product team designs the product. The supply chain team makes sure the parts arrive. The financial reporting separates them. The career paths separate them. The conversation about strategy is mostly about products and markets, with supply chain treated as a constraint that has to be respected rather than a position that can be built.
The companies that have learned to treat supply chain as a competitive position have produced durable advantages that are hard to copy. The classical examples are Walmart in retail (cross-docking, EDI mandates on suppliers) and Toyota in automotive (just-in-time, kanban, supplier development). The hardware example, since the late 1990s, is Apple. The pattern is the same. Operations work that looked like back-office plumbing turned out to be the load-bearing structure under the product strategy.
The supply chain plays that produce a competitive moat in hardware are not exotic. They are three moves that compound over time.
The first move is capacity reservation. You pay a supplier in advance, often by funding the tooling and equipment, to reserve capacity for your product. The supplier agrees not to sell the same capacity to your competitors for some period. The financial commitment is significant. The strategic effect is that your competitors are delayed on the same component by the length of the reservation.
The second move is multi-year volume commitments. You sign for two or three years of volume in exchange for pricing and capacity that a one-year buyer cannot get. The supplier gets predictability, which lets them invest in capacity. You get cost structure your competitors cannot match. The classical Apple version of this is the relationship with TSMC. Apple's commitment to TSMC volume is what funds TSMC's leading-edge capacity. TSMC's leading-edge capacity is what makes Apple Silicon possible.
The third move is supplier development. You invest in the supplier's process. You send your engineers to live at their factory. You help them improve their yield, their tooling, their quality systems. The supplier's capability becomes part of your capability. This is the Toyota pattern, and Apple does a version of it at scale. The depth of Apple's relationships with Foxconn, Pegatron, and the major component suppliers is the reason Apple can ship hundreds of millions of devices per year on aggressive product cadences.
The pattern Asymco named the "Bank of Apple" is the integrated version of all three moves, at a scale that almost no other company can match.
The mechanic is roughly: Apple identifies a strategic component (advanced display glass, custom CNC tooling for unibody machining, OLED panels, advanced camera sensors). Apple commits cash to fund the supplier's tooling and capacity, in exchange for exclusivity and pricing. The capital outlay is held on the balance sheet as a long-term asset, not as a one-time expense. The supplier gets capital that they could not raise on their own market terms. Apple gets first access to the technology and a pricing structure that reflects the depth of the commitment.
The strategic effect is that Apple's competitors operate on a delayed timeline for every component Apple chose to invest in. The financial effect is that Apple's component costs are lower than the competition's because the tooling has been amortized against guaranteed Apple volume. The capital position is itself a moat. A competitor that wanted to copy this strategy would need not only the cash but the willingness to deploy it at the same scale and the supplier relationships to make the bets stick.
The supply chain that Cook built is concentrated in China. Final assembly, for most of Apple's product line, happens at Foxconn, Pegatron, and Luxshare facilities in Zhengzhou, Chengdu, and Shenzhen. The concentration is the source of the speed and the source of the risk. Apple can ship products at volumes and cadences no other company can match. Apple is also exposed to a single political and geographic environment in a way that has become uncomfortable since 2018.
The diversification work that has been visible since 2019 is non-trivial. Final assembly for some iPhone models has shifted to India, primarily through Foxconn's facilities in Tamil Nadu. NPI for entry-level iPad has reportedly moved to Vietnam. The MacBook supply chain has diversified into Thailand and Malaysia. The pace of the diversification is slower than political pressure would suggest, and the reason is that the supplier ecosystem in Shenzhen took twenty-five years to build. Recreating it elsewhere takes time. The companies that are betting on a sudden China decoupling are mostly going to be wrong on the timeline.
The Bank of Apple strategy is not available to companies that do not have Apple's cash. The lesson is not the strategy. The lesson is the disposition.
The disposition is that supply chain decisions are strategic, not tactical. Where you build, who you build with, and how deep your relationship goes are decisions that shape what you can do for the next five years. They cannot be unwound quickly. A small hardware company that picks a contract manufacturer based on the quote on the first build is making a five-year decision based on a one-quarter signal. The right way to pick a manufacturer is to look at where you want the relationship to be in three years, what the manufacturer's capacity and capability roadmap is, and whether the relationship can grow at the rate your product can.
The other lesson is that the supplier relationships you invest in early compound. Engineers who have worked with your team for two years on the previous product are an asset. The yield curve they helped you tune on the previous product is an asset. The trust that lets you ship a new product on a tight schedule is an asset. None of these show up on the balance sheet. All of them are real, and they are the reason that hardware programs that change manufacturers between generations almost always have a worse second generation than they had a first.
TSMC is the most strategic supplier relationship Apple has, and it is the most consequential supply chain relationship in the world. The dependency is structural. Apple's lead products are built on TSMC's leading-edge nodes (N5, N4, N3, N2). The relationship is mutually exclusive in practice (no other foundry has the capability) and mutually committed (Apple's volume funds TSMC's leading-edge capacity).
The Acquired podcast's January 2025 interview with TSMC founder Morris Chang is worth two and a half hours of attention if you want to understand how this came to be. The short version is that Chang's commercial design (a pure-play foundry that does not compete with its customers on chip design) is the structural reason fabless silicon design exists at all. Apple's silicon strategy is downstream of Chang's. Without TSMC, Apple Silicon does not exist. Without Apple, TSMC's leading-edge volume is much smaller. Each is the other's most important customer or supplier in the sense that mattered most.
The book to read on the broader geopolitics is Chris Miller's Chip War (Scribner, 2022). The argument is that semiconductor supply chains are the single most consequential industrial topology of the next twenty years, and the West's exposure to Taiwan-concentrated leading-edge capacity is the largest strategic risk in the industrial economy. The book won the Financial Times Business Book of the Year for 2022 and is the standard reference on this topic.
Geoffrey Boothroyd and Peter Dewhurst, working at the University of Massachusetts and then at the University of Rhode Island in the early 1980s, formalized what is now called Design for Manufacturing and Assembly (DFMA). Their core observation was that the cost of a product is determined far more by the design than by the manufacturing process. Most cost-reduction work in factories is fighting the wrong battle. The right battle is fought before the design is released.
The Boothroyd-Dewhurst method is a structured way of asking, for every part in an assembly, whether the part needs to exist. Three questions: does this part move relative to the parts it connects to? Does this part need to be a different material? Does this part need to be separable for assembly or service? If the answer to all three is no, the part should be combined with an adjacent one. The method, applied rigorously, typically reduces part count by 30 to 50 percent.
The reason this matters is that part count is the leading indicator of assembly cost, supplier complexity, and quality risk. A product with 100 parts is roughly twice as expensive to assemble as a product with 50 parts, even if the bill of materials cost is similar, and the quality risk is more than twice as high because every additional interface is an additional place where something can be wrong.
The vocabulary distinguishes Design for Manufacturing (DFM) from Design for Assembly (DFA), with the integrated practice called DFMA. The distinction is useful.
DFM is about whether each part can be made cheaply, accurately, and reliably. The questions are about manufacturing process. Can this part be injection-molded with the geometry it has, or does it need machining? Are the tolerances appropriate to the process? Does the design avoid features (undercuts, thin walls, sharp corners) that drive yield down? Each manufacturing process has design rules, and DFM is the discipline of designing inside those rules.
DFA is about whether the parts can be put together cheaply, quickly, and reliably. The questions are about the assembly process. Can this part be picked and placed by an automated tool? Does it self-locate? Does it need to be flipped or repositioned during assembly? Is the assembly direction consistent? Each interface is an opportunity for error, and DFA is the discipline of minimizing the number and complexity of interfaces.
The combined practice, DFMA, is what Apple's hardware engineering organization runs at high discipline. The reason Apple products feel solid and assemble cleanly is not magic. It is that the parts have been designed for the manufacturing process and the assembly sequence, often through tens of design iterations before the product is ready for tooling.
Hardware engineering teams new to manufacturing make a predictable set of mistakes. The mistakes are easy to describe and hard to avoid without experience.
The first is treating tolerances as wishes. A drawing that calls for 0.05 mm tolerance on every dimension is a drawing that has not engaged with the manufacturing process. Each manufacturing process has a tolerance capability. Injection molding holds 0.1 to 0.2 mm reliably. CNC machining holds 0.025 mm reliably. Sheet metal fabrication holds 0.5 mm reliably without secondary operations. A tolerance tighter than the process can deliver requires either a more expensive process or secondary operations to bring the part into spec. Both add cost and yield risk.
The second is over-using fasteners. Every screw is a part, an operation, an opportunity for cross-threading, and a customer service call when it strips. The Apple unibody MacBook reduced screw count dramatically by going to a milled aluminum chassis with snap-fit and adhesive joints where possible. The lesson is not to eliminate screws (some interfaces require them) but to be honest about which interfaces actually need them.
The third is single-sourcing critical components without a second-source plan. The component you single-source is the component that will gate your production when the supplier has a fire, a labor strike, a yield crash, or an export-control issue. The right pattern is to design with the constraints of two suppliers' parts, even if you only buy from one of them, so that the second source is a real option when you need it.
The fourth is ignoring assembly access. The product looks good in CAD because all the parts are visible and the operator's hand is not modeled. The product is impossible to assemble in production because the operator cannot reach the screw without disassembling the part that goes in next. DFA reviews exist to find this. They have to be done with the people who will actually assemble the product, not in a room with the design team.
The relationship between design decisions and unit cost is non-linear. A useful approximation is that 70 to 80 percent of a product's unit cost is locked in by the time the design is released for tooling. Manufacturing engineers can pick up another 10 to 15 percent through process optimization in the first year of production. The remainder is roughly fixed.
This is the structural reason that DFM has to happen during design. By the time the product is in pilot production, the levers that matter have been used. Any cost reduction that has not been designed in by DVT is a cost reduction that requires a redesign, which means new tooling, which means time and capital that the program does not have.
The corollary is that the time spent on DFM during design is the highest-return hardware engineering work in the entire program. An engineer who spends an extra week negotiating part count, tolerances, and assembly sequence with the manufacturing team is doing work that produces returns over the entire production life of the product. The same engineer's work on a yield improvement project after launch is fighting against the design that already shipped.
Yield is the percentage of units coming off the production line that meet specification. First-pass yield is the percentage that pass on the first try. Final yield is the percentage that ship after rework. Both numbers matter, and the gap between them is a cost.
A program with 95 percent first-pass yield and 99 percent final yield is healthy. A program with 80 percent first-pass yield and 95 percent final yield is paying for the rework, which is real cost in labor, time, and risk that the rework introduces a different defect. The DFM work is what produces the gap between these two scenarios. A design that absorbs manufacturing variation produces high first-pass yield. A design that is marginal produces low first-pass yield, which is the leading indicator of a program that will struggle to ramp.
The yield curve over the life of a product is also predictable. First-pass yield typically starts low at the beginning of mass production and improves over the first six months as the line tunes. The shape of that curve is a leading indicator of program health. A curve that flattens early at a low number is a design problem that no amount of manufacturing optimization will fix. A curve that climbs steeply is a process problem that the manufacturing team can solve. Reading the curve correctly, and acting on it, is one of the things that distinguishes mature hardware programs from immature ones.
Somewhere right now, a person is opening a small white box. They lift the cardboard sleeve, tug a paper tab about half a millimetre thick, and a charging cable slides free of its coil. The whole motion takes around three seconds. Almost no one notices it.
Jony Ive spent Sunday afternoons on that tab. "I had such a clear awareness that in designing a certain solution, for example, how we managed a cable that's in a box, designing that, I knew that millions of people would engage with this little tab," he told Patrick Collison at Stripe Sessions in San Francisco on May 8, 2025. The conversation ran an hour. It is the closest he has come, in public, to laying out the philosophy under thirty years of work.
Jonathan Paul Ive was born on the eastern edge of London in February 1967. His father, Michael John Ive, was a silversmith who lectured at Middlesex Polytechnic and went on to help shape the UK's national curriculum for Design Technology. People who worked alongside Mike say, in retrospect, that he laid the groundwork for a generation of British designers, his son among them.
The transmission was domestic and slow. Mike and Jony walked the streets of north London together and talked about the things they passed. Why the streetlamp on this side of the road had a different mounting than the one across. Why the bus shelter faced north when the wind came from the west. Each question assumed an object had been made by a person who had decided things, and that the decisions could be examined.
For Christmas, Mike's gift to Jony was a single day in the empty Polytechnic workshop. They had the run of the lathes and benches. The condition was that Jony had to draw what he wanted to make first, in pencil, before anything was cut. Decades later he would describe it as the closest thing he ever had to a private apprenticeship.
The drawing rule was not arbitrary. The hand has to commit to an idea before the materials do, and a poorly conceived object cannot be drawn convincingly: the drawing exposes what hasn't been thought through.
By 1985, when Ive enrolled at Newcastle Polytechnic, his working method was already settled. He sketched first, in volume, and built foam models in volume too. Where most classmates produced six or seven physical iterations of a project, he routinely made fifty or a hundred, refining the shoulder of a handle by a millimetre at a time. His tutors describe a student who treated every exercise as if the object would actually be manufactured.
Some of them were. As an intern at the London consultancy Roberts Weaver Group, Ive worked on a ballpoint pen for the Japanese stationery brand Zebra. The pen, the TX2, sold in millions and stayed in production for years. There is a small mechanical detail at the cap, a clip with a ball that doubles as a fidget, giving the user something to do with the hand while thinking.
He won the RSA Student Design Award in 1988 and again in 1989. The cash prize, modest, came with a travel grant. He used it for a trip to California, and while there he met Robert Brunner, the head of industrial design at Apple. The meeting was informal. Ive showed Brunner his student work, including a tubular phone he had built for his final year. The model was not a styling exercise; it included internal components and a cost-and-tooling argument for how the device would actually be made. Brunner remembered it for years.
Ive graduated with a first in 1989, went back to Roberts Weaver for a year, then moved with several colleagues to a London consultancy called Tangerine in 1990. The studio, in Hoxton Square, did consumer-product work for clients including LG, Ideal Standard, and Apple. The objects he designed during this period included microwaves, drills, toothbrushes, and a sanitary-ware range. None are remembered now, except by the people who used them, which is consistent with how he thinks about design.
The Mac arrived in Ive's life as an undergraduate. He has told the story several times and the version he tells is consistent. The personal computers available in the UK in the late 1980s, in his account, felt indifferent to the people who used them, sometimes hostile. The Macintosh was the first computer he met that didn't feel that way. Someone on the other end, he said, had cared.
"What we make stands testament to who we are," he said at Stripe. He has used the line for years. The argument under it is that an object is a kind of letter from its makers to the people who use it, and the user reads the letter accurately whether or not they have language for what they are reading. "People know carelessness," Ive added.
The Mac was the letter that told him there was a way to build computers without contempt for the user. Brunner's offer in 1992 followed from that. He has described his arrival in Cupertino, talking to Collison, as walking into "an innocent euphoria, I think, of like-minded people driven by values clearly in service of humanity gathering together." The phrasing is unguarded.
Steve Jobs returned to Apple in late 1996. Within months the company was preparing what would become the iMac. Ive, then 30, ran the project. He had been at Apple five years, and the studio was eight people.
The decision to use translucent plastic for the iMac G3 enclosure began with a piece of beach glass that somebody brought into the studio. The team had a habit of bringing objects in to look at: water bottles, sweet wrappers, a Game Boy, a piece of greenish-blue glass worn smooth by the Pacific. The argument for translucency was partly practical, since solid plastic in cheap colours would have looked cheap, and partly about the relationship between the user and the machine. A translucent enclosure stops asking the user to ignore the inside of the object.
It is also harder to make. Solid pigments hide variation in the moulding process; clear or tinted plastic does not. Every internal component had to be redesigned to look acceptable through the case, because the user could now see it. Ive has called the decision "cheeky" in interviews. It forced the rest of the company's hardware engineers to a standard that had previously been invisible.
The Bondi Blue iMac shipped in August 1998 and sold roughly a million units in its first six months. It was the first product to put Ive's signature on a mass-market computer, and the Industrial Design Studio at One Infinite Loop became the centre of the next twenty years of Apple's output.
None of Ive's philosophy is invented. The people he has read, looked at, worked with, and learned from are part of how he thinks, and the list is worth knowing if you want to read his work the way he reads it.
His father, again. Mike Ive, the streetlamps, the workshop, the rule about drawing first. Most of what Ive met later, in conversation with other designers, was already implicit in those walks down the road.
Then Dieter Rams. Rams ran design at Braun in Frankfurt from 1961 to 1995 and produced, with a small team, one of the most coherent product catalogues of the twentieth century. His ten principles of good design, written in 1976, are the closest available articulation of what Ive practices: good design is innovative, it is unobtrusive, it is honest, it is as little design as possible. The connection between the two designers is documented. Around 2008 Ive sent Rams an iPhone in the post, with a note thanking him for the inspiration; Rams kept the letter and described it in a 2010 interview with Die Zeit. In 2011 Ive wrote the foreword to Sophie Lovell's monograph on Rams. The Apple iOS Calculator, with its convex circular buttons and high-contrast palette, is a near-direct quotation of the Braun ET 66 calculator that Rams and Dietrich Lubs designed in 1987.
Naoto Fukasawa runs his own studio in Tokyo and is responsible for much of MUJI's visual identity. His wall-mounted CD player for MUJI, the one that hangs from a string and starts when you pull, was an object the Apple team passed around in the late 1990s; Ive has called it "an important and resonant signal" for the studio. He and Fukasawa collaborated on the Twentieth Anniversary Macintosh in 1997, when Fukasawa was at the IDEO predecessor ID Two. Fukasawa's interest in "without thought" design, objects that fit so well into a routine that the user does not register them, surfaces in the iPhone home button and, later, in the Apple Watch's haptic tap.
Marc Newson, the Australian industrial designer, has been Ive's closest creative friend for thirty years and is co-founder of LoveFrom. Newson made his name with the Lockheed Lounge in 1986 and has produced furniture, watches, cameras, and luggage that share a vocabulary of soft, sensuous geometry. He and Ive co-curated a 2013 (RED) auction at Sotheby's that raised over $26 million for the Global Fund. Their collaboration has continued through Leica cameras, the LoveFrom catalogue, and the Ferrari programme.
Robert Brunner ran Apple's industrial design from 1989 to 1996, hired Ive twice (once at Tangerine, once as an Apple employee in September 1992), and now runs Ammunition, the San Francisco studio behind Beats by Dre. The line Ive used at Stripe, "they persuaded me to move to San Francisco," is mostly about Brunner.
And Steve Jobs. Ive and Jobs ate lunch together most days for fourteen years. The partnership produced the iMac, iPod, MacBook, iPhone, iPad, the original Apple Watch (still in development at Jobs's death), and the early sketches of Apple Park. The Jobs line Ive returns to most often is "you can express gratitude to the species through what you make." If you want one sentence to explain why Apple's products feel different from what most other companies ship, that is it.
The list is not exhaustive, but the shape of it matters. Mostly European industrial designers who privileged restraint, plus one Japanese designer who made restraint feel emotional, plus the partner who taught Ive how to argue for what restraint required inside a public company. Silicon Valley barely figures.
The Rams principles deserve a closer look, because they are the curriculum behind the curriculum. The 1976 essay Weniger, aber besser ("Less, but better") is the source. Rams wrote it as a self-assessment after a decade running design at Braun, in response to what he saw as a flood of products that mistook decoration for design. The principles aren't commandments; they're tests, and the test is whether the maker can answer the question honestly.
The ten:
The tenth runs through the others. Subtraction is harder than addition because subtraction requires deciding what is essential, and a designer who keeps adding has not committed. The Braun catalogue is the proof: the SK 4 record player from 1956, the T 3 pocket radio from 1958, the L 2 speaker from the same year, the ET 66 calculator from 1987. One argument made repeatedly under different constraints.
Ive applied a version of the same test, but at a different scale. Where Rams worked inside a German appliance company in the post-war Wirtschaftswunder, Ive was eventually shipping products in hundreds of millions of units. The eighth principle, thoroughness in detail, has to carry through tooling, supply chain, and final assembly across that volume. His contribution to the Rams tradition was showing that it could.
Every Friday at the Industrial Design Studio, one designer cooked breakfast for the rest of the team. The rotation was strict, the menus loose. Korean banchan, English fry-ups, sourdough, scrambled eggs with chives from somebody's garden in Los Altos. The studio had a kitchen and a long communal table, and people sat together and ate.
Ive's reasoning at Stripe was specific. Teams that care for each other listen to each other; teams that don't, wait their turn to speak. He described himself as "an extremely shy and introverted person," and his deepest fear in any meeting was that the best idea would come from the quietest voice in the room and never reach the table. The breakfast was an attempt to build the conditions in which it would.
"Make things for each other," he said at Stripe. "It makes you more vulnerable. It makes others grateful." The line is more radical than it sounds. The standard model of a corporate design studio treats the customer as the audience and the team as the production crew. Ive's model treats the team as the first customer. If the people next to you cannot feel the care in the work, the millions of strangers downstream will not either.
The cable tab is what this looks like in practice. The functional spec is "holds the cable." The actual spec, the one Ive built around, is something closer to "is the first thing a person feels when they meet our work." Sunday afternoons on the geometry of a paper pull are the cost of meeting the second specification, and the team that paid that cost was the team that cooked for each other on Fridays.
Throughout the Stripe interview, Ive pushed back at what he called the "functional imperative" school of product development: the idea that a thing solves the problem and the work is done. He thinks that framing is the source of most of what's wrong with what the technology industry ships.
His argument is not that beauty matters more than function. He thinks the problem is misstated when function is treated as the whole of it. A product is also a relationship, and beauty, in the way he uses the word, is a property of the relationship the object offers to the person who lives with it. A tool that works but feels indifferent has done half its job.
At Stripe he named a specific failure mode. "Joy in humans has been missing," he said, "and sometimes joy gets confused with being trivial." Designers and engineers, he argued, steer away from joy because they want to be taken seriously. The result is products that solve problems while transmitting nothing of the makers' care for the people using them. Building for joy, in his account, is the harder work; what he keeps naming as the failure is joyless work that mistakes its joylessness for seriousness.
"What kills an idea is people's urge to express their opinion." Ive said this at Stripe, and the studio he ran was built deliberately against the urge.
He talked about the practices he used to protect ideas. Working from each other's homes, where the architecture of the room is not a conference table. Long pauses in meetings, where the silence is functional rather than awkward. Letting the quietest person speak last instead of first. Treating a half-formed idea as a fragile thing, rather than as raw material for a critique session. Most design organisations get this exactly the wrong way around. They optimise for the loudest voice, on the assumption that confidence is a proxy for quality. Ive's bet is that it is not, and that the most considered idea is often the one that takes longest to form, made by the person who needs the most cover to bring it out.
The connection back to joy isn't accidental. Joy lives in the specific, and specificity comes from noticing, and noticing happens in silence. Loud meetings tend to produce generic ideas because they reward whoever can produce a confident statement first. Quiet meetings tend to produce specific ideas because they reward whoever has been paying attention.
The most pointed line of the interview, for an audience of fintech founders, came near the end. "The lie is, because we spend all our time talking about what we can measure, that's all that matters."
The contrast he drew was with the things he believes actually drive the experience of a product: trust, delight, the sense that the maker cared. None of these are measurable in the way conversion or retention or churn are. The temptation, when something cannot be measured, is to ignore it, because the unmeasured thing does not produce a dashboard. The result is products that succeed on the dashboard and fail in the room.
The implication is that the measurements you choose to attend to are themselves a values statement. A team that measures only what is measurable will design only for what is measurable, and the unmeasurable in such a team turns up where it has always turned up: in the seams, the edges, the cable tabs, the haptic feedback that nobody asked for and everybody felt.
Ive and Jobs spent fourteen years in something close to constant conversation. Lunches almost daily. The studios linked by a covered corridor between buildings at One Infinite Loop. By the time Jobs died in October 2011, the partnership had produced the iMac, iBook, Power Mac G4 Cube, iPod, MacBook, MacBook Pro, MacBook Air, Mac mini, iPhone, iPad, the original Apple Watch (still in development at the time), and the early sketches of Apple Park.
Apple Park is the artefact of the partnership most often overlooked. Jobs began the project with Norman Foster in 2010, his last year. The brief was specific: a building that combined the collegiate atmosphere of Stanford with the apricot orchards Jobs remembered from his Cupertino childhood. The eventual ring, with 80 percent of its 175-acre site planted with drought-resistant indigenous trees, was nailed down about six months in. "Every pane of glass in the main building will be curved," Jobs told Foster. "We have a shot at building the best office building in the world."
When Jobs died, Foster and Ive carried it. Tim Cook, who became CEO in August 2011, gave them the runway to finish without changes. Foster has called the transition "seamless" (his word). Ive has called the building "our house, where we go to work together," and added, "We made it for us, to help us be better, to make better products." It opened in April 2017. Ive, by then, had taken a step back from day-to-day product work, and left Apple two years later.
Ive announced LoveFrom in June 2019, with Marc Newson as co-founder, and left Apple at the end of that year. The studio's offices are in Jackson Square, San Francisco, in four buildings on Montgomery Street that Ive bought between 2020 and 2022 for an estimated $90 million. The team is around 60 people, and the disciplines on staff range across industrial design, graphic design, user-interface design, architecture, CAD sculpting, writing, filmmaking, and music. The breadth is what makes the studio different from other industrial design firms.
The published work has been small in volume. Christie's commissioned LoveFrom to design a new auctioneer's rostrum to mark the auction house's 260th anniversary in March 2026; the rostrum, in oak with stainless-steel detailing, was the first the firm has commissioned in a generation. Ferrari, a multi-year partnership announced in 2021, produced the interior and interface of the all-electric Ferrari Luce in February 2026. Airbnb engaged the studio to redesign its app and visual identity. There has been work for Moncler, Coca-Cola, the King's Foundation, and a Coronation logo for King Charles III. The studio's stated mission is to "sincerely elevate the species." The mission is grandiose; the work, so far, has been more modest than the mission.
The largest current project is the AI hardware device LoveFrom is building with OpenAI. The collaboration was reported by The New York Times in September 2024 and remains, as of writing, undisclosed in form. The worry Ive has about AI's "rate of change" is, by his own statement, a worry he is trying to design against from the inside. "The ownership of that has been driving a lot of what I've been working on that I can't talk about," he told Collison.
The Stripe interview is also where Ive said the thing he doesn't usually say in public: he is worried about what the technology industry is shipping.
The phrasing was direct. "Even if you're innocent in your intention, if you're involved in something that has poor consequences, you need to own it." He drew a contrast with social media, which arrived without a culture of safety conversations attached to it, and said he is cautiously hopeful that AI is being built differently. His specific concern is the rate of change: the Industrial Revolution, in his framing, gave societies decades to adapt; the current curve does not.
The ownership clause is the design lesson under the worry. The maker is responsible for the second-order effects of the made thing, not only the first-order ones. A designer who ships a product that delights one user while hollowing out a generation has not done good work, even if the first user is happy. Most makers, when asked about consequences, claim innocence on intent. Ive will not let himself off the hook that easily.
None of what Ive teaches is best learned from a portfolio review. It is closer to apprenticeship than research, and the practices are simple enough to begin with even if they take a career to do well.
The Mike Ive rule is the place to start: draw what you intend to make before you make it. The hand has to commit to an idea before the materials do, and a sloppy idea is exposed in pencil before it is exposed in plastic.
Read Dieter Rams. The 1976 essay Weniger, aber besser, the ten principles, then the Braun catalogue from 1955 to 1995. Sophie Lovell's monograph, with Ive's foreword, is worth owning.
Look at ordinary objects on purpose. Streetlamps, door handles, cable ties, the seam of a paper bag. Mike Ive trained his son by walking down the street together and asking why; the question still works.
Build the same thing many times. Ive made a hundred foam models for a project his classmates made six. The discipline is iteration: each model rules out a small set of possibilities and refines the surviving set.
Spend disproportionate time on the small touches: the cable tab, the cable wrap, the first second of a power-on. The user does not have a vocabulary for what they are noticing, which is precisely the point.
Build conditions for the quiet voice. A team that talks over its quietest member ships its second-best ideas. The ritual does not have to be Friday breakfast, but it has to do the same work.
Treat function as half the spec. The other half is the relationship the product offers the person who lives with it.
Refuse the lie of measurement. Trust, delight, and care do not show up on the dashboard, and they are still the work. A team that ignores them ships products that test well and live badly.
Take responsibility for the second order. If what you ship has costs you did not intend, those costs are still your work.
The Stripe interview is on the public record. Watch it twice. The second time, watch his hands and the pauses, not the words. He is showing you the practice as much as describing it.
On January 9, 2007, Steve Jobs stood on stage at Macworld in San Francisco and showed a phone that did not really work yet. He scrolled through music, made a call, loaded a web page, and the room believed it was watching a finished product. It was not. The iPhone would not go on sale until June 29, 2007, and in January much of its software still crashed if you touched the wrong thing in the wrong order.
That gap between the announcement and the ship date is the whole story. Apple committed publicly to a product months before that product could survive a day of normal use. The question worth asking is why a company would do that on purpose, and what the choice did to the people building the thing.
Apple did not usually preview hardware. The standard pattern was to announce and ship close together, partly to deny competitors a head start and partly to keep the surprise. The iPhone broke that pattern, and the main reason was regulatory.
A phone with a cellular radio has to be certified by the FCC before it can be sold in the United States. FCC certification filings become public, and they include photographs, internal diagrams, and technical detail. Apple submitted the iPhone to the FCC in early 2007, and those documents would eventually open to anyone who asked. Jobs did not want the world's first look at the iPhone to be a regulatory PDF. Announcing in January let Apple reveal the phone on its own terms, with its own framing, before the filing could leak the secret out from under them.
The early date did other work too. It gave the press almost six months to build anticipation, and it pressured carriers and developers into the orbit of a product that did not yet exist in stores. Once Apple had said the words out loud, the deadline was real and external.
Inside Apple the effort was Project Purple. The team was sealed off in a locked building in Cupertino that people called the Purple Dorm, named in part for a smell that reminded them of pizza. Scott Forstall recruited engineers without being able to tell them what they would build. He told them the project was so secret he could not describe it, and that they should expect to give up their nights and weekends for a couple of years.
The door of the Purple Dorm had a Fight Club sign on it, and the joke carried the rule: the first rule was that you did not talk about it outside those doors. Jobs scattered pieces of the work across the campus so that no single group could see the full shape of what was coming.
Apple did not start with one design. Two competing concepts ran in parallel. P1 was built on the iPod, steered by the click wheel that people already knew. P2 was the riskier idea: a slab of glass driven by multitouch, descended from work that would also feed into the OS X side of the company.
Running both at once was a hedge. Rather than bet early and risk betting wrong, the team developed the two paths together and let them prove themselves. P2 won. The click wheel could not handle the demands of a phone with a browser and a keyboard, and multitouch could. The phone Jobs showed in January was the descendant of P2, and the operating system underneath it was a stripped-down relative of the Mac's, not the iPod's.
The most revealing detail of the January demo came out years later, mostly through Fred Vogelstein. Reporting for the New York Times Magazine in a 2013 piece titled "And Then Steve Said, 'Let There Be an iPhone,'" and again in his book "Dogfight," Vogelstein described how fragile the demo unit really was.
The prototype had only 128MB of memory and ran software that was unfinished, slow, and prone to freezing. The apps were incomplete. Tap the wrong icon and the phone could lock up or shut off. So the engineers worked out what they called the golden path: a precise sequence of actions, performed in a specific order, that the phone could get through without crashing. Jobs could send an email and then browse the web, but if he reversed the order the phone tended to fall over.
Jobs rehearsed against that script for days. The team kept several phones on hand so they could swap to a fresh unit if one died mid-demo. The signal was so unreliable that engineers hard-coded the status bar to show five bars and put a portable cell tower backstage. By the accounts of people who were there, they were drinking Scotch in the audience by the time it was over, because they could not quite believe nothing had broken.
A launch date you have announced to the world is a forcing function. It does not care whether the software is ready. It converts every open question into a countdown, and it forces decisions that would otherwise drag on. June 29 was not negotiable once Apple had said it, and that fact organized everything the team did between January and summer.
That pressure has a real cost. People burned out, features got cut, and the demo had to be staged because the actual product could not yet do what the demo claimed. The deadline bought focus, and it charged for it. The lesson is not that public deadlines are good or bad, but that they are powerful, and that a team should understand what kind of pressure it is signing up for before it picks a date and says it out loud.
There is a separate question the January announcement could not answer: could Apple actually build these things at scale, by the millions, and get them into people's hands. Sourcing, manufacturing, and the deals that put the phone in stores are their own subjects, and later articles in this folder take them up. Here the point is narrower. Apple chose a date, told the world, and then spent six months making the claim true.
Steve Jobs had been carrying a prototype iPhone in his pocket for weeks, and he came back angry. The screen was plastic, and his keys had scratched it. Walter Isaacson recounts the moment in his biography: Jobs decided the phone would ship with glass, and he wanted it changed weeks before the June 2007 launch. People put glass and keys and other junk in their pockets, he reasoned, and a plastic screen would look beaten up before the phone was a month old. The decision was simple to state and very hard to execute. The product was nearly done, the launch date was fixed, and glass that could survive a pocket did not exist on Apple's shelf or anyone else's that Apple had lined up.
This is a story about the part of a launch that customers never see. By the time you are holding a finished phone, someone has already solved the problem of where the glass comes from, who builds the thing, and how you get from zero to enough units to sell on day one. The original iPhone is a good case to study because almost everything about it was being attempted for the first time, on a deadline, in secret.
Jobs went to Corning, the glassmaker founded in 1851, and to its CEO Wendell Weeks. As Isaacson tells it, Jobs started explaining to Weeks how glass was made, and Weeks cut him off: "Can you shut up and let me teach you some science?" Corning had a chemically strengthened glass it had developed in the 1960s under the name "Project Muscle," a glass strengthened by an ion-exchange process that puts the surface under compression so cracks do not spread easily. It had never found a real market and had been shelved for decades. Corning later branded it Gorilla Glass.
Jobs wanted as much of it as Corning could make, in a matter of months. Weeks told him plainly that Corning did not have the capacity, that none of its plants were making the glass at all. Jobs told him not to be afraid, to get his mind around it, that he could do it. Weeks, in Isaacson's account, shook his head when he retold the story. Corning repurposed a plant in Harrodsburg, Kentucky to make the glass at volume on a brutal timeline. Reporting on the episode, including a Fast Company account years later, describes the Harrodsburg facility being switched over to LCD glass and then to the strengthened glass for Apple. Corning delivered. The phone shipped with glass.
One thing deserves precision, because the legend tends to compress it. The glass that shipped on the first iPhone was the product of this crash effort. The "Gorilla Glass" brand name and the broad licensing business came later. What matters for the launch is that a supplier with a dormant technology was pushed to industrialize it in months, and that Apple was willing to put money and its own deadline behind that push.
The glass is the famous part, but a phone is a few hundred parts, and most of them came from somewhere specific. The picture below comes from teardowns done right after launch, mainly by iSuppli (later part of IHS) and reported through outlets like AppleInsider and EDN. Treat the exact part numbers and dollar figures as teardown estimates rather than Apple's own disclosures, because Apple disclosed almost nothing.
Samsung supplied the application processor, an ARM-based system-on-chip that teardowns identified as the S5L8900, reported as a stacked package that also carried memory dies. Samsung also supplied the NAND flash that held the storage, an 8-gigabyte part on the high-end model. The cellular baseband came from Infineon, the PMB8876 "S-Gold 2," along with Infineon's radio transceiver and power management parts; iSuppli put Infineon's combined silicon at around 15 dollars of the bill of materials. Wi-Fi came from a Marvell chip, the 88W8686, reported at roughly 6 dollars. The touch panel module was attributed to Balda of Germany working with TPK in China, estimated near 27 dollars, which made it one of the more expensive single items. The touchscreen controller was a Broadcom part, the BCM5973. iSuppli's much-quoted total put the hardware and manufacturing cost of the 8-gigabyte model around 266 dollars against a 599-dollar price.
I list these not to be exhaustive but to show the shape of the problem. A new product type pulls parts from a dozen vendors across several countries, each with its own lead time, yield curve, and minimum order. Several of these were single-source: one processor vendor, one baseband vendor. Single-sourcing buys you the best part and tight integration, and it means a problem at one supplier can stop the whole line.
The pieces came together in China. Final assembly of the original iPhone was done by Hon Hai, the Taiwanese contract manufacturer better known by its Foxconn trade name, at its operations in Shenzhen. Contract manufacturing at this scale is not a matter of handing over a design and waiting. Apple had to deliver tooling, the molds and fixtures and test rigs specific to this product, and tooling has long lead times of its own. It had to define the assembly steps, the test points, and the criteria for a unit to pass. And it had to do the work of new product introduction, the NPI ramp, where a line that has never built a given product is brought up from hand-built samples to thousands of identical units a day.
The hard part of an NPI ramp is that yield starts low. The first units off a new line are slow and many of them fail test, and you learn why by building them, finding the failures, and fixing the process. For a known product you have done this before. For the iPhone, nobody had built a capacitive multi-touch phone with a bonded glass front at volume, so every station on the line was being figured out at the same time the clock was running toward June 29.
A launch with a date announced to the public removes the easiest pressure valve. If your ramp goes badly, you cannot quietly slip the launch by a quarter, because the date is already a promise. You can ship fewer units, you can ration supply, but you have to have working phones in stores on the day. That turns every supply question into a scheduling question. The long-lead components, the ones with the slowest path from order to delivery, set the real start date for the whole program. If the application processor or the glass or the touch module is the bottleneck, no amount of effort downstream buys back the time.
Secrecy made it harder. Apple kept the product tightly held, which limited how much it could tell suppliers and how openly it could second-source or shop a part around. A vendor building to a spec it only partly understands, under a code name, cannot help you the way a fully briefed partner can. Apple traded some of that openness for control of the surprise, and the surprise was part of the product.
The keys-in-the-pocket decision is a good lesson precisely because of when it happened. Changing a material weeks before launch is one of the most expensive moves you can make in a hardware program. The glass front is not a sticker you swap. It changes the supplier, the tooling, the bonding process, the strength testing, the way the panel mates to the touch sensor, and the assembly line that puts it all together. A change that would be cheap in early design becomes enormous in late ramp, because by then a long chain of decisions has been built on top of the old choice.
Jobs made the call anyway, and it worked, which is why it gets told as a triumph. The honest reading is narrower. It worked because Apple could put 200 million dollars and its own engineers behind Corning, because Corning had a real, if dormant, technology to revive rather than inventing one from scratch, and because there was just barely enough calendar left. The lesson for most teams is not "demand glass at the last minute." It is the opposite: decide the things that touch your suppliers and your tooling early, because late changes are paid for in money, in risk, and in nights nobody gets back.
You will probably not be sourcing a phone. The structure of the problem carries over anyway. Whatever you are launching, some part of it comes from outside your walls, some part of it has the longest lead time and quietly sets your schedule, and some part of it can only be made by one supplier who can therefore stop you cold. Manufacturing readiness for a launch is mostly the work of finding those parts before they find you, and giving yourself enough slack that a bad yield week or a late mold does not turn into a missed date you already announced.
Pick up an original iPhone and look at the front. There is no carrier logo. There was no carrier software loaded on it, no apps you could not delete, no branded start screen. In 2007 that was strange enough to be the story. American phones came to you through the carrier, who subsidized the hardware, stamped its name on the case, loaded its own software, and decided which models lived or died. Apple sold the first iPhone into that world while refusing almost every part of how it worked. This is a story about the deal that let them do it, the price they picked, and the day the phone went on sale.
A phone needs a network, and in the United States that meant making a deal with one of the carriers. What Apple wanted was close to a reversal of the normal arrangement. Apple would control the device, the software, and the brand. The carrier would not put its name on the phone, would not load software onto it, and would not decide how it was sold. Apple wanted to own the relationship with the customer. And reportedly it wanted a share of the monthly service revenue, around ten dollars per subscriber per month, on top of selling the hardware.
By most accounts Apple went to Verizon first, and Verizon said no. Reporting at the time, traced back to USA Today, says Verizon balked at exactly these terms: a cut of monthly fees, Apple's control over how and where the phone could be sold, Apple's say over customer service and repairs. From Verizon's seat that was the carrier handing its business to a hardware company. They passed.
Cingular took the deal. The agreement was an exclusive one, reported as a five-year US exclusive, and Cingular got the iPhone on Apple's terms. Cingular would share a portion of monthly subscriber revenue with Apple. There is a wrinkle of timing worth stating plainly: the company that signed was Cingular, but it was in the middle of rebranding to AT&T, the name it had acquired, and around the launch the iPhone's carrier was being called AT&T. Same company, two names depending on when you read about it.
The unusual part was not that a phone had one carrier. Exclusives were common. The unusual part was who set the terms. A carrier was agreeing to be a pipe: provide the network, share the revenue, and stay off the product. Apple traded away the carrier's subsidy and its distribution muscle, and in return kept the thing it cared about most, which was control of the product and the customer.
The iPhone launched at 499 dollars for the 4-gigabyte model and 599 dollars for the 8-gigabyte model, and you still had to sign a two-year contract with the carrier on top of that. By the standards of 2007 this was a lot of money for a phone. The going expectation was that a contract bought down the hardware to a low or sometimes zero up-front price, the cost buried in your monthly bill. The iPhone asked for the contract and a high sticker price.
That price did work for Apple. The phone read as a premium object, the contract and the price both signaling that this was not a giveaway handset. Whether buyers experienced 599 dollars as positioning or just as expensive is a fair question. What is clear is that Apple chose a number that sat well above what a subsidized phone felt like, and held it as part of the product's identity.
The price held for about ten weeks. On September 5, 2007, Apple cut the 8-gigabyte iPhone from 599 dollars to 399 dollars, a 200-dollar drop, a little over two months after launch. People who had paid full price in June and July were angry, and they said so loudly. They had bought the thing at its most expensive, partly to have it first, and now it cost two hundred dollars less while they were still inside their two-year contracts.
Steve Jobs answered with an open letter on Apple's site. He defended the cut, writing that lowering the price was the right decision and the right time to make it, that the iPhone was far enough ahead of the competition that a lower price would reach more customers. He also gave ground. Apple offered early buyers a 100-dollar credit at the Apple store. The letter acknowledged, in plain terms, that the people who had paid early were upset and that Apple owed them something.
There is a lesson sitting in here about launch pricing, and it cuts two ways. A high launch price captures the buyers who want the thing most and will pay most, and you can lower it later to reach everyone else. That is ordinary. The trouble is the speed. Ten weeks is fast enough that your earliest, most enthusiastic customers feel punished for their enthusiasm. The 100-dollar credit was the cost of repairing that, and the fact that Apple paid it tells you the goodwill of early adopters was worth protecting. The honest version is not that the cut was a mistake or a masterstroke. It was a real cost, mostly paid in trust, and Apple chose to buy some of it back.
The marketing leaned on wanting the phone rather than on what it did. The first iPhone television ad, called "Hello," aired during the 79th Academy Awards on February 25, 2007, months before anyone could buy one. It was a montage of film and television characters answering phones and saying "hello," dozens of them, from old movies and shows, cut together so the word passed from one to the next. The iPhone appeared at the end with the line that it was coming in June. The ad did not even say the product's name out loud.
What the ad did not do is as telling as what it did. It did not list the things the phone could do. It did not explain the touchscreen or the browser or the voicemail. It put the phone at the end of a long line of phones people already knew and loved from the movies, and let you fill in the rest. That is a confident way to advertise a product, and it only works if the product can carry the desire the ad builds. Apple bet it could.
The iPhone went on sale June 29, 2007. People lined up outside Apple Stores and AT&T stores, some for days, and the lines were part of the coverage as much as the phone was. The scarcity and the waiting were doing work that no ad could buy.
The quiet but important decision was how you turned the phone on. Instead of activating it in the store, the way carrier phones worked, you took the iPhone home, plugged it into your computer, and activated it through iTunes. This broke the carrier-store model on purpose. The store was no longer the place where the carrier signed you up and learned who you were. Apple put activation through its own software, on your own machine, which kept the experience inside Apple's world and kept the carrier at arm's length from the moment of setup. It also strained the activation servers. With everyone activating at once, some buyers waited hours to get a working phone, which was the predictable cost of moving a step that used to happen in thousands of stores into one funnel.
The early numbers came with caveats. Apple said it sold 270,000 iPhones in the first 30 hours, a figure it reported a few weeks later. That number drew some skepticism at the time, partly because it landed below what some analysts had floated and partly because counting "first 30 hours" is its own choice. Apple then announced it had sold its one millionth iPhone on September 9, which it put at 74 days after launch. Both figures come from Apple, so read them as Apple's framing rather than audited counts, but the shape is not in dispute: fast out of the gate, a million inside two and a half months.
Most of what made this launch unusual was not the phone. It was the choices around the phone. Apple picked its channel and bent the carrier to it instead of accepting the carrier's terms, which is the choice of who owns the customer. It put activation through iTunes, which kept that ownership even at the moment of setup. It priced high and let the price say something, then paid in credits and an apology when the cut came too fast. And it spent its first ad building desire rather than explaining anything.
None of these requires a phone to copy. The questions underneath them are the general ones for any launch. Who controls the relationship with the buyer, you or your distributor. What your price is saying before anyone reads the spec sheet. Whether your earliest customers will feel rewarded or burned by what you do in the months after they buy. The iPhone got most of these right and paid visibly for the one it rushed, which is about as instructive as a launch gets.
When people say a product launched on a certain date, they mean the day the press release went out, the keynote happened, or the store page flipped to "buy now." That day is real, and it matters. But it is the visible tip of something much larger, and treating it as the work is how launches go wrong. The date is when the program gets tested in public. The program is everything that had to be true for that test to pass.
I find it useful to define launch by what a customer can actually do, not by what a company announces. A product has launched when a customer can find it, buy it, receive it, and use it successfully, all four, in sequence, without a person from the company stepping in to rescue the transaction. If any one of those breaks, you have made an announcement, not a launch. The order matters too. A customer who can find and buy a product but cannot get it working has a worse experience than one who never found it, because now they have paid and they are disappointed.
So the question that organizes a launch is not "when do we tell people," but "when is every step of that path solid enough to send a stranger down it." Most of this article is about how those two questions diverge, and what to do about the gap.
Before any of the readiness work, you have to be able to say what the product is and who it is for in a way that survives contact with a real buyer. This sounds obvious and is the step teams most often skip, because the people building the product already know what it does and assume everyone else will too.
Clayton Christensen's jobs-to-be-done framing is the cleanest way I know to force the question. Christensen argued that customers do not buy products so much as hire them to do a job. His standard example came from a fast-food chain trying to sell more milkshakes. Demographic research and product tweaks did nothing. The insight came from watching when shakes sold: early morning, to solo commuters, who were not really hungry but faced a long, dull drive and wanted something that lasted and could be held in one hand. The job was "keep me occupied and not-hungry on my commute." Once you see that, you stop competing with other milkshakes and start competing with bananas, bagels, and boredom. The product you would build is completely different.
The launch consequence is direct. If you do not know the job, you will describe your product by what it does feature by feature, and a list like that is not a reason to buy. The buyer has to translate that list into "will this do my job," and most buyers will not do that translation for you. They will leave.
April Dunford makes the case in Obviously Awesome that positioning is context-setting: you are telling the buyer what kind of thing your product is, so they bring the right expectations and the right comparisons. Get the context wrong and even a good product reads as confusing or overpriced. Dunford breaks positioning into a few connected pieces: the competitive alternatives a customer would use if you did not exist, the unique capabilities you have that those alternatives lack, the value those capabilities enable, the customers who care most about that value, and the market category that frames all of it.
What makes this a launch problem and not just a marketing problem is that positioning decides the comparison set. If you launch a project-management tool and let people file it next to the market leader, you are judged against that leader's full capability list on day one. If you position it as the tool for a specific kind of team doing a specific kind of work, you are judged against whatever that team uses now, often a spreadsheet and a group chat, and you look great. Same product, different verdict, because the context was set differently before anyone clicked.
Dunford's practical point is that positioning is deliberate. The default, which is to position against whatever the most obvious competitor is, is rarely the position that makes you look best. You choose it, and you choose it before you write a word of launch copy.
A launch is the one moment when product, engineering or manufacturing, marketing, sales, support, and legal all have to be ready on the same day. Each of these functions can be "almost done" on its own schedule. The launch is what converts six separate "almost done" states into a single pass-or-fail.
The failure mode is that each function reports green against its own definition and no one checks the seams. Engineering ships the capability. Marketing has the campaign. Support has not been told the capability exists, so the first wave of confused customers hits a team that cannot answer them. Or legal has not cleared a claim the campaign is built around, and the campaign has to be pulled the morning of. None of these is a hard problem in isolation. They become a problem because launch is the only date that forces all of them to be true together, and most planning happens inside functions rather than across them.
The fix is unglamorous: one launch owner who does not own any of the individual functions, whose only job is the path from find to use working end to end, and a readiness review where each function reports against the same definition of done. The owner's real value is asking the seam questions. Does support have the answer to the question marketing's headline will generate. Can the buy button actually charge a card in the regions the press release will reach.
You do not have to announce and ship on the same day, and often you should not. The first iPhone is the clearest example. Apple revealed it on January 9, 2007, and did not sell one until June 29, almost six months later. Part of the reason was practical: Apple had to file with the FCC, and Steve Jobs did not want the product's existence leaking out of a regulatory filing before he could reveal it on his own terms. The announce-early choice traded months of buildable hype against the risk of telling competitors what was coming and the risk of customers holding off on buying current products.
That trade-off generalizes. Announcing before you ship buys you demand, press, and time to seed reviewers, at the cost of tipping your hand and starting a clock you now have to beat. Announcing at ship, the more common pattern, keeps things tight but forfeits the runway. There is no universal right answer, but there is a wrong one: announcing on a date you then cannot ship by, which converts your hype into impatience and your early fans into people writing "where is it" under every post.
Related is the choice between a soft launch and a big-bang launch. A soft launch, limited regions, a waitlist, a percentage rollout, lets you test the find-buy-receive-use path with real customers while the blast radius is small. You give up the concentrated attention of a single big moment, but you get to find the broken step before it breaks at scale. For anything where you are unsure the path holds, soft launch first and save the big moment for when you know it works.
A committed launch date is one of the most useful tools a team has. It organizes the whole program, it forces decisions that would otherwise drift, and it gives every function a shared deadline to plan against. Take the date away and work expands to fill all available time. This is the honest case for picking a date early and defending it.
The same date becomes a risk the moment it has no slack. A plan where every milestone has to land exactly on time for the launch to hold is a plan that assumes nothing goes wrong, and something always goes wrong. The way through is milestone gating: the launch date depends on earlier gates, like "feature complete," "support trained," "legal cleared," each with its own date and each able to slip a little without the launch slipping. When a gate misses, you learn it weeks out, while you still have moves, rather than the night before. A date with gates and slack is a forcing function. A date with neither is a countdown to an incident.
Define the metric that matters before launch, because afterward you will be tempted to pick whichever number went up. The trap is vanity metrics: page views, sign-ups, press mentions, downloads. These feel like success and mostly measure attention, which fades. The metric you want measures whether the product is doing its job. For software that usually means activation, the share of sign-ups who reach the point of getting real value, and then retention, whether they come back. For physical goods it often means sell-through, the share of units that actually sell rather than ship to a shelf, and the return rate. Pick the one or two numbers that would still matter in three months, write them down, and write down the threshold that would count as the launch working. Doing this before launch is the only way to keep the post-launch story honest, because the temptation to declare victory off a traffic spike is strong and the spike tells you almost nothing.
Three patterns account for most launch failures I have seen, and none of them is bad luck. Scope creep is the first: the launch keeps absorbing "just one more" thing, the date holds, and quality and readiness get squeezed to make room. The second is marketing writing checks operations cannot cash, a promise of same-day shipping the warehouse cannot meet, a capability the demo implied that the product does not have, so the gap between expectation and reality lands as a flood of disappointment right when attention peaks. The third, underneath both, is no clear owner of the end-to-end path, so every function optimizes its own piece and the seams between them belong to no one.
A launch is a process, and the day is where you find out whether the process was sound. Plan the path, not the announcement. Pick a date, then give it gates and slack. Decide what winning means before you can be tempted to redefine it. Do that and the day mostly takes care of itself, which is the goal: a launch day that feels uneventful because all the work that mattered already happened.
A launch is the day you tell the world the product exists and ask people to buy it. The day before, someone has to make sure there is a product to hand over, that it works when the buyer turns it on, and that it can actually reach them. That second job is operations, and it is the one most likely to be treated as a detail right up until it becomes the reason the whole thing slips.
Marketing runs on dates. A campaign has a calendar, a set of beats, a moment everything points toward. Operations runs on lead times and physical limits, and those do not move because a date was promised in a meeting. When the two disagree, the launch breaks along the operations line almost every time, because you can rewrite a press release the night before and you cannot conjure ten thousand finished units the night before. A launch fails if there is nothing to ship or what ships is broken. Everything else is downstream of that.
Whatever you are building, some of it comes from outside your walls, and the parts that come from outside arrive on their own schedule. In supply chain terms, lead time is the total elapsed time from placing an order to having the thing in hand, including the supplier's order processing, their production, and transport. For a stock screw it is days. For a custom display, a specialized sensor, or a chip built on a constrained process, it can run months, and it is common for the longest single lead time in a program to exceed the entire marketing runway. The campaign you can plan in eight weeks may depend on a part that takes twenty to arrive.
That math is why the long-lead component, not the launch date, sets the real start of the program. If your slowest part takes five months, ordering it the day you finalize the design already puts shipping five months out, and no effort downstream buys that time back. The first job of sourcing is finding which part is slowest before it finds you.
The second job is asking how many suppliers can make each part. A single-source component, one that only one qualified vendor produces, gives you tight integration and often the best version of the part. It also means a fire, a flood, a labor dispute, or a quality hold at that one supplier stops your line cold. Teams reduce that exposure by second-sourcing, qualifying a second vendor for the same part. Qualification is not a phone call. It means confirming the second source meets the same electrical, mechanical, and environmental specs, running their parts through your tests, and often re-validating the assembly that uses them. It is slow and it costs money, which is exactly why people skip it and then regret it when the single source fails.
A change to the design is cheap when the design is still on a screen. It gets more expensive at every step after that, and the curve is steep. Once you have cut tooling, the molds and fixtures and test rigs built specifically for one version of the part, changing the part can mean cutting new tooling, and tooling has long lead times of its own. Once you have qualified suppliers and validated the assembly, a change can force re-qualification and re-validation of everything the change touches.
The original iPhone is the clean example. Weeks before the June 2007 launch, Steve Jobs decided the screen had to be glass instead of plastic, and Apple pushed Corning to industrialize a strengthened glass it had developed decades earlier and shelved. It worked, and it gets told as a triumph, but the honest reading is that it worked because Apple could put serious money and its own engineers behind a supplier who already had the technology, with just barely enough calendar left. For most teams the lesson runs the other way. Decide the things that touch your suppliers and your tooling early, because a change that costs a meeting in design costs a quarter in late ramp.
Hardware teams move a product to volume through a sequence of gates, and the vocabulary is worth knowing because the gates exist to stop you from shipping a problem you have not found yet. The common sequence is EVT, DVT, PVT, then MP. EVT, engineering validation test, is where the first real prototypes confirm the thing works at all and meets the spec it was designed to. DVT, design validation test, widens the question from individual parts to the whole product, checking that the design as a system hits its functional and performance targets and can be built. PVT, production validation test, builds units on the actual production line with production tooling and process, to prove that the line, not just the engineers, can make a good unit at rate. Only then does the program move to MP, mass production, where multiple lines run in parallel and output climbs to full volume.
The reason you cannot safely compress these phases is that each one tests a different failure mode, and skipping a phase means shipping with that failure mode untested. EVT finds design bugs. DVT finds system interactions. PVT finds the problems that only appear when a real line with real operators and real tooling tries to make the part repeatedly. Yield, the fraction of units that pass, starts low on a new line and climbs as you find and fix process problems by building, failing, and learning. Compressing the phases does not make the line faster. It moves the discovery of the problems from your factory to your customers.
Software has its own version of this. The analog to a careful ramp is staged rollout, releasing to a small fraction of users first and widening as the metrics hold, backed by load testing against expected traffic and capacity provisioning so the servers do not fall over on launch day. A digital launch can fail the same way a hardware launch fails, by going to everyone at once before anyone confirmed the system survives everyone at once.
You need enough units in the channel to sell on launch day, with stock in stores and distribution centers before the first customer shows up. That is channel fill, and getting it wrong is costly in both directions. Build too few and you sell out, lose the launch-day momentum, and hand the moment to a competitor or to resellers marking up scarce stock. Build too many of a product that flops and you are left with inventory you have to discount or write down, having spent cash on parts and assembly for units nobody wanted.
The hard part is that a new product has no sales history to forecast from. You are estimating demand for something the market has never seen, under deep uncertainty, and the usual statistical tools that lean on past demand have nothing to lean on. Teams fall back on analogs to similar products, pre-orders, channel commitments, and judgment, and they stay honest about the error bars. The reasonable move is usually to stage the build so you can ramp output as real demand data arrives, rather than committing the whole bet to a number guessed before launch.
First-pass yield, the percentage of units that pass inspection with no rework or repair, is the headline quality number on a line, and a low first-pass yield early in a ramp is normal. But passing in the factory is not the same as surviving in the field. The lab tests what you thought to test, for as long as you had patience to run it. The field tests everything, in conditions you did not imagine, for far longer, across far more units.
That gap is where the expensive failures live. A part can pass every bench test and then fail after months of thermal cycling in a hot car, or after a kind of drop the test plan did not cover, or only on the small fraction of units that sit at the edge of a tolerance the lab sample happened to miss. These are the failures that pass the lab and surface as field returns once the population is large and the clock has run. You cannot fully close the gap before launch, but you narrow it with margin, with accelerated life testing, and with the honesty to treat the lab result as suggestive rather than final.
Having a working unit in a warehouse is not the same as a customer using the product. Between those two states sits logistics and everything that has to work for the buyer. Distribution has to put units where people can get them. Activation, the steps to set the product up and get it running, has to work on the buyer's first try, because a product that ships fine and then fails to activate is, to the customer, broken. Returns will happen, and the path to handle them has to exist on day one rather than be improvised under load. Support has to be staffed and trained before the questions arrive, not after.
These are easy to defer because they sit at the end of the line and feel like they can be sorted out later. They cannot, because launch day is exactly when all of them get hit at once, at the highest volume the product will ever see relative to how ready the support organization is.
The single point that ties all of this together is about who holds the launch date. A date set by marketing without operations buy-in is a wish. The parts arrive when they arrive, the line yields what it yields, and the field finds what it finds, none of which care about the campaign calendar. The only safe launch date is one that operations has signed off on, derived from the longest lead time, the ramp the phases actually allow, and the inventory the build plan can deliver.
This is unglamorous, and it is the constraint. You can move a launch date before you announce it. You cannot move it after, and you cannot ship what you did not build. Let operations readiness set the date, and let marketing point its calendar at that date rather than the other way around.
A launch is the moment a product stops being yours and starts being something a stranger can buy, use, and judge. Everything before it is internal. The schematic, the roadmap, the demo that wowed the board, none of it counts until someone outside the building hands over money and gets value back. Getting a product into hands is its own discipline, separate from building it, and teams that are good at one are often bad at the other.
This is about that discipline: how you price, where you sell, how you create the demand, what you actually say to the buyer, who closes the deal, and how you know afterward whether any of it worked. I will lean on the original iPhone because it is well documented and because Apple got some of this exactly right and one piece of it visibly wrong, which is more instructive than a clean win.
Most teams treat pricing as arithmetic. Add up the cost, mark it up, ship it. That is cost-plus pricing, and its main virtue is that it is easy. Its flaw is that it ignores the buyer entirely. You can be profitable on every unit and still leave money on the table, or price yourself into a market you did not intend to enter. Cost-plus answers the question "what do I need to charge" and never asks "what is this worth to the person buying it."
The two alternatives both start from outside the building. Competitive pricing sets your number against what rivals already charge, which is a shortcut, because the market has validated the range and buyers will not flinch. The cost is that it signals parity. You are telling the buyer you are one more option in a known category. Value-based pricing starts from what the product is worth to the customer for the job it does, and it is the only one of the three that lets you charge a premium on purpose. It also demands that you actually understand the buyer, which is harder than reading a competitor's price sheet.
The part teams miss is that the number is read as a claim. A high price says "this is the good one." A low price says "this is the affordable one," and you cannot easily say both. Pricing is the most compressed piece of positioning you have, a single figure the buyer decodes before they have read a word of your marketing. Pick it to match where you want to sit, not just to clear your costs.
Once you set a launch price, cutting it fast is dangerous, and the first iPhone is the textbook case. It shipped on June 29, 2007 at 599 dollars for the 8-gigabyte model. On September 5, barely two months later, Apple dropped that to 399 dollars, a 200-dollar cut. Early buyers were furious. They had paid a premium to be first and watched the premium evaporate before the phone was a season old.
Steve Jobs posted an open letter within days. He apologized, gave every full-price iPhone buyer a 100-dollar Apple store credit, and defended the move in the same breath, writing that "there is always change and improvement, and there is always someone who bought a product before a particular cutoff date." That credit cost Apple real money and bought back some goodwill, but the cleaner lesson is that the situation should not have existed. A price cut this soon after launch tells your most loyal customers, the ones who paid first, that their loyalty was a mistake. The goodwill you spend repairing it is worth more than the demand the cut creates. If you think you will need to cut, you probably launched too high, and it is cheaper to start lower than to retreat.
Where you sell shapes what you can control. Selling direct, your own store or site, gives you the customer relationship, the data, the margin, and the message, at the cost of reach and the work of building distribution yourself. Selling through retail or partners buys reach fast and hands away some of the relationship: the partner owns the shelf, often the data, and a slice of the margin, and they can bury you next to a competitor.
Apple's 2007 carrier deal is a sharp example of refusing the usual trade. Carriers at the time dictated the phones on their networks, loaded them with their own software, and put their brand on the hardware. Jobs would not accept that. He went to Verizon first and the asks were too steep, so Verizon walked. AT&T took the deal: a multi-year exclusive in exchange for letting Apple keep control of the device, the software, and the brand, with Apple even selling music and apps to iPhone users without sharing that revenue. Apple used a partner for reach onto a cellular network it could never build, and still kept the things that made the product Apple's. That is the move worth copying. Decide which parts of the relationship are load-bearing for you and refuse to trade those, even when a partner offers reach for them.
The iPhone was announced in January 2007 and shipped at the end of June. That announce-then-ship gap was deliberate. It let anticipation build, gave the press six months of material, and turned launch day into an event with lines outside stores. The lines are theater, and theater is not worthless. It manufactures scarcity, gives the press an image, and signals to everyone still deciding that other people already decided. What it does not do is create durable demand on its own. A line is a spike. The business is what happens in the months after the cameras leave.
The split that matters here is earned versus paid. Earned media is the coverage and word of mouth you do not buy, and it carries credibility paid media cannot. Paid media is the advertising you control and can turn up at will. Launch-day press is mostly earned, which is why it is so valuable and so hard to repeat: it is a one-time event, and once the spike fades, paid media is what keeps users arriving. The mistake is treating the announcement as the finish. The announcement is the easy part. Sustaining demand after the novelty wears off is the actual job, and it runs on a clear message far more than on a long list of capabilities. Nobody outside your company has memorized what your product can do.
The clearest way to think about that message comes from Clayton Christensen's jobs-to-be-done idea: customers do not buy products, they hire them to do a job. His example was a fast-food chain trying to sell more milkshakes. Research on the milkshake itself, sweeter, thicker, more flavors, did nothing. When the team watched who actually bought them, they found morning commuters hiring the milkshake to make a long, boring drive bearable, something thick enough to last and easy to hold one-handed. The job was the commute, not the dessert. Once you see the job, the product and the pitch both change.
April Dunford makes the same point for the business buyer in her book "Obviously Awesome." Positioning, in her framing, is choosing the market context where your product's strengths are the obvious answer, and most products are positioned badly not because they are weak but because they are described against the wrong alternative. A spec sheet describes the product. Positioning describes the situation the buyer is in and why this is the thing that resolves it. Lead with the job and the situation, and the specs become supporting evidence rather than the argument.
Awareness is not a sale, and the gap between them is where launches quietly fail. For a consumer product the closing motion is often the product itself: the buyer sees the ad, walks into the store or to the checkout page, and the experience does the selling. The funnel is short, so the work is removing friction, making the path from "I want this" to "I own this" as direct as possible.
For a B2B product the funnel is long and a human closes it. That salesperson needs material the marketing never had to produce: the answer to the procurement question, the security review, the comparison against the specific competitor the buyer already uses, the case for why now rather than next year. This is sales enablement, and it is unglamorous and decisive. A launch that generates demand the sales team cannot convert is a launch that produced expensive noise. The funnel from awareness to a buyer who is using the product has to be aligned end to end, and the handoffs between marketing and sales are where it usually breaks.
The numbers a launch generates are not all worth the same. Press impressions, page views, and pre-orders are easy to celebrate and easy to mistake for success. The numbers that tell you whether you have a business are sell-through (units that actually reached end customers, not units shipped to a channel), activation (buyers who set the thing up and used it), and retention (buyers still using it weeks later). A launch spike with no activation behind it is a product people bought once and abandoned, and you will see it eventually in returns and silence.
The failure that recurs across launches has two shapes, and they are mirror images. In one, marketing writes a check operations cannot cash: demand arrives and there is no supply, no support, no working signup, and the goodwill burns off in a week. In the other, the product is genuinely good and nobody can find it or understand what it is for. A good launch is the narrow case where the product is ready, the price says the right thing, the buyer can find it, the message names the job, and somebody is there to close and to count what happened honestly. Getting it built is half the work. Getting it into hands is the other half, and it is the half more teams underestimate.
Rick Rubin has produced records for Johnny Cash, the Beastie Boys, Slayer, and Adele, and he often says he is not sure what a producer does. In The Creative Act: A Way of Being, published in 2023, he describes his real work as reduction. He listens to a song and tries to find the version of it that already exists underneath the arrangement, then removes whatever is covering it up. For the first of Cash's American Recordings sessions, that meant a man, a guitar, and the songs, recorded in a living room. No strings. No backing vocals. The job was to get out of the way.
This is the idea I want to open the folder with. Reduction is the practice of finding the essential core of a thing by removing everything that is not essential. It sounds simple. It is one of the harder things to do well, in any field, and the rest of this folder is about why.
Antoine de Saint-Exupéry was a pilot before he was a writer, and the most quoted sentence he wrote is about machines, not poetry. In Terre des Hommes (1939), translated into English as Wind, Sand and Stars, he is writing about the slow refinement of the airplane. The French reads: "Il semble que la perfection soit atteinte non quand il n'y a plus rien à ajouter, mais quand il n'y a plus rien à retrancher." Perfection is reached not when there is nothing more to add, but when there is nothing left to take away.
He meant it literally. Decades of engineering, he writes, end in a shape so plain that its guiding principle is simplicity, as if the design had been worn down to its essentials by use. The line gets borrowed for everything now, often without the source. It is worth keeping the original context. A man who flew unreliable aircraft over the Andes had concrete reasons to admire a design with nothing extra on it.
Dieter Rams ran design at Braun for decades and shaped how a calculator, a radio, and a record player were allowed to look. His phrase for the whole approach was "Weniger, aber besser," less but better. It appears as the last of his ten principles for good design: good design is as little design as possible. The point is not minimalism as a style. It is that a good product concentrates on what is essential and is not burdened with the rest.
Rams is a large enough subject that a later article in this folder takes him on his own. I mention him here because his phrase is the cleanest statement of the discipline. Less, but better, holds two things at once. Less is the cut. Better is the reason for the cut. You are not removing things to be spare. You are removing things so the remaining thing works.
There is good evidence that humans are bad at this by default, and the reason is not laziness. In 2021, Gabrielle Adams, Benjamin Converse, Andrew Hales, and Leidy Klotz published a paper in Nature with a blunt title: "People systematically overlook subtractive changes." Across eight experiments, when people were asked to improve something, they reliably reached for adding before considering removing, even when removing was the better move and sometimes the obvious one.
One experiment used a Lego structure that needed stabilizing. People added bricks far more often than they took one away, though taking one away solved it. The results suggest subtraction is cognitively harder. It tends not to come to mind unless people are prompted, given more than one try, or freed from time pressure and mental load. When the researchers paid participants to consider both options, subtractive solutions went up. So the bias is real, and it is also movable.
That finding matters for everything else in this folder. If addition is the default move, then reduction is the move that takes effort and attention. It will not happen on its own. Someone has to decide to look for what to remove.
What I find interesting is how consistent this is across fields that otherwise share almost nothing. An architect cutting a wall, a composer dropping a melody line, a painter scraping back a passage, an editor killing a paragraph that the writer loved, a product team removing a feature that took three months to build. Different materials, same underlying act. Find the part that carries the work and protect it. Remove what competes with it.
The folder walks through these one at a time. There are pieces on architecture, on music, on visual art, on film and the cutting room, on writing, and on product work. Each domain has its own vocabulary for the same discipline, and its own famous failures, the projects that drowned in additions. Reading them together, the pattern is hard to miss.
Here is my honest position. More is almost always the easy choice, and it usually looks like progress. Adding a feature is something you can point to in a meeting. Adding an instrument fills a silence that felt risky. Adding a sentence covers a thought you had not finished. Addition produces visible output, and visible output feels like work getting done.
Reduction produces less, and it produces it slowly, because deciding what is essential means understanding the thing well enough to know what it can lose. That is the harder skill and the more valuable one. "Just enough" is a worse pitch and a better outcome.
The word I keep coming back to is editing. Reduction is editing, and editing is something you practice, not something you feel your way into. It has methods. Cut and see if anything breaks. Remove the part you are proudest of and check whether the work survives, because pride is a bad signal of necessity. Ask what a thing is for, then keep only what serves that.
None of this is mystical, whatever Rubin's calmer passages might suggest. The Nature paper shows the bias is measurable and that prompting reduces it. So the rest of this folder is, in part, a set of prompts. Reduce, do not produce, is the instruction. The articles that follow are about how people in different fields have learned to follow it.
The Farnsworth House sits on a wooded site near Plano, Illinois, finished in 1951. Ludwig Mies van der Rohe designed it for Edith Farnsworth as a weekend retreat, and what he built was close to nothing: a single rectangular room of glass, raised on slender steel columns above the floodplain of the Fox River. Three horizontal planes, the terrace, the floor, and the roof, hold the whole thing in the air. There are no interior walls to speak of, only a freestanding core for the bathrooms and kitchen. The structure is painted white. The glass is everything else.
Stand inside it and you understand why reduction is hard. There is nothing to hide behind. Every weld, every proportion, every column flange has to be right, because there is no ornament, no trim, no second material to distract the eye from a mistake. When people quote Mies saying "God is in the details," this is what they mean. Strip a building down to structure, glass, and proportion, and the details stop being decoration. They become the building.
Mies is the architect most tied to "less is more," and he used it often enough that it reads like his personal motto. He did not coin it. The line comes from Robert Browning's 1855 poem "Andrea del Sarto," where the painter says "Well, less is more, Lucrezia." Browning meant it as a remark about restraint in art, spoken by a painter who was technically flawless and creatively timid. Mies took the phrase and gave it a structural meaning it never had in the poem. That is worth holding onto, because it shows reduction as an idea older than modernism, borrowed and sharpened rather than discovered.
His earlier work makes the point even more clearly than the Farnsworth House. The Barcelona Pavilion, built for the 1929 International Exposition, was a small structure with almost no program at all, no rooms, no function beyond being walked through. Mies filled it with rich materials, Italian travertine, green marble, a wall of onyx, chrome columns, and let planes of stone slide past each other under a thin flat roof. Reduction here did not mean cheapness or poverty. It meant removing everything that was not space, surface, and proportion, then making what remained as fine as he could. The pavilion was demolished after the exposition closed and rebuilt on the original site in the 1980s, which tells you something on its own. A building with so little in it survived as an idea, precise enough to reconstruct from drawings and photographs, because the idea was the proportions and the materials, not any single physical wall.
The argument that ornament itself was the problem came earlier, from the Viennese architect Adolf Loos. His essay "Ornament and Crime" is usually dated to 1908, though the record is messier than that: he seems to have first delivered it as a lecture around 1910 and published the French text, "Ornement et crime," in 1913. The date matters less than the claim, which was blunt. Loos argued that the evolution of culture moves in step with the removal of ornament from everyday objects. To carve, gild, and decorate was, to him, a waste of labor and material, and a sign of a culture that had not grown up.
Some of Loos's reasoning has aged badly, including the way he linked ornament to so-called primitive peoples, and it is fair to read him skeptically there. But the architectural core of the argument held. He saw that decoration dates an object, that a thing covered in the fashionable patterns of one decade looks tired in the next, and that a plain wall outlives a busy one. Reduction, for Loos, was partly a way of building things that would not embarrass you in twenty years. That is a quieter and more practical idea than "less is more," and it sits underneath the whole modern movement.
The German industrial designer Dieter Rams put the principle into its sharpest words. Born in 1932, he ran design at Braun for decades, and somewhere along the way he wrote down ten principles of good design. The last one is the one people remember: good design is as little design as possible. His own phrasing for the whole approach was "Weniger, aber besser," less, but better. Not less for its own sake, but less so that what remains can be better.
Rams worked mostly in products rather than buildings, and his influence on later product design, including a famous line of computers and phones, belongs to another piece. What carries over to architecture is the qualifier in his phrase. "Less, but better" is not the same as "less." The word "better" is doing the work. It says the point of removing things is to improve what is left, and that a reduction which makes a thing worse has failed, however clean it looks. You can apply the same test to a building. Take a thing away and ask whether the room is better for losing it. If it is, the thing was inessential. If it is not, you have confused empty with simple, and those are different.
Japanese aesthetics arrived at related ideas by a different road. The concept of "ma" names the interval, the gap, the negative space between things. The character combines a gate with the sun shining through it, light in an empty doorway. In Japanese art and architecture, that emptiness is not leftover space. It is the active part of the composition, the pause that gives the rest its shape. One common description calls ma the silence between the notes that makes the music.
Tadao Ando builds with this. His Church of the Light, finished in 1989 in Ibaraki near Osaka, is a small concrete box, around 113 square meters, with a cross cut through the wall behind the altar. Ando uses three things: concrete, glass, and light. The room is dark and bare. The cross of daylight that falls across the front is the whole event. Nothing else competes with it, which is exactly why it lands. The English architect John Pawson, who spent time in Japan in his twenties and absorbed this thinking, has built a long career on the same principle: large quiet spaces that push you toward the few essential forms left in them.
Reduction in architecture carries a specific risk. When you take things away, whatever survives has to carry the entire load of the experience: proportion, material, light, the joint where two surfaces meet. There is no slack. A richly detailed building can absorb a clumsy corner because the eye is busy elsewhere. A reduced one cannot. The Farnsworth House works because the proportions and the steelwork are exact. Change them slightly and you have a cold glass shed.
This is why so much that calls itself minimal is just empty. Removing ornament is easy. Anyone can specify white walls, gray concrete, and no skirting boards. What is hard is making the remaining elements good enough to hold attention on their own, and that takes more care, not less. When the care is missing, reduction collapses into a style, a look you can buy, bare rooms that feel like a showroom rather than a place. Minimalism becomes its own kind of decoration, an ornament of absence. The honest version is harder won. It removes the inessential so the essential can actually be felt, the light on the concrete, the weight of the floor, the length of a room. Less, when it works, is not emptier. It is clearer.
In 1994 Johnny Cash released an album with almost nothing on it. American Recordings was his voice and an acoustic guitar, recorded mostly in his living room in Hendersonville, Tennessee and at Rick Rubin's house in Los Angeles. No band. No backing vocals. No reverb to hide behind. By then Cash was sixty-two, dropped by his label, treated by the industry as a museum piece. Rubin had seen him perform at Bob Dylan's thirtieth-anniversary concert and thought there was something in that voice that no one was recording.
What they made was an album that works precisely because of what is missing. With no production around it, you hear every crack in the voice, every breath before a line, every silence between phrases. The reduction is the point. Strip a thing back far enough and you find out whether there was anything holding it up in the first place. Cash's voice held it up.
Rubin built his reputation as a producer who removes rather than adds. He did it across genres, pulling Red Hot Chili Peppers back toward the song under the funk, sitting with artists until the arrangement stopped competing with what the singer was actually saying. He is not a trained musician and does not play an instrument well. His instrument is taste, and taste here mostly means knowing what to cut.
In his 2023 book The Creative Act, Rubin describes creativity less as invention than as reception. The work, in his telling, already exists in some form, and the artist's job is to notice it and clear away what obscures it. I would not put words in his mouth that he did not write, and the idea is older than any one phrasing of it, but the practical version is plain enough in how he works. You subtract until you reach the thing that cannot be subtracted, and then you stop. The stopping is as much a decision as anything you played.
The line most often attached to Miles Davis is that it is not the notes you play, it is the notes you don't play. The attribution is hard to pin to a single documented source, and similar phrasings have been hung on other musicians, so I would treat the exact wording as folklore rather than transcript. What is not folklore is the playing. Davis built solos out of space. He let a phrase land and then left a gap where a lesser player would have crowded in more notes.
Kind of Blue, from 1959, is the clearest case. The record is built on modal jazz, which means fewer chords and longer stretches over a single scale. Instead of racing through dense chord changes, the players get room. The harmony moves slowly, so a soloist can breathe, repeat, wait. Davis treated the empty bars as part of the music, not as time to fill. That is reduction working at the level of the bar and the beat.
There is a sentence that floats around music writing, attributed to Claude Debussy, that music is the space between the notes. It is a lovely line and I have no clean source for it. It gets pinned on Debussy, sometimes on Davis, occasionally on Mozart, which usually means no one really knows. I quote it the way you quote a proverb, for what it says rather than who said it.
What it says is true to how listening works. A note means something because of the silence that frames it. Remove the silence and you remove the shape. This is the same principle as a held breath in speech or a margin around a paragraph. The blank is not nothing. It is doing the work that makes the marks legible.
John Cage took the idea to its end. On August 29, 1952, the pianist David Tudor walked onstage at the Maverick Concert Hall in Woodstock, New York, sat at the piano, and played nothing for four minutes and thirty-three seconds. He marked the three movements by opening and closing the keyboard lid and timed them with a stopwatch. The score for 4'33" instructs the performer to make no intentional sound.
The piece is not really about silence, because there is no silence. Cage's point was that once you stop adding intended sound, you start hearing everything else. He recalled the wind outside during the first movement, rain on the roof during the second, and the audience's own restless noises during the third. The reduction went all the way to zero, and what remained was the room. 4'33" is the frame with the picture removed, and it turns out the frame was the picture.
The minimalists reduced in a different direction. Rather than stripping a song to its core, they started from very little material and let it run. Terry Riley's In C, from 1964, hands the players a single page of short figures over a steady pulse and lets them move through it at their own pace. Steve Reich's It's Gonna Rain, from 1965, takes a recorded fragment of a preacher's voice and plays two copies of it slightly out of step, so the phrase slowly drifts against itself. Philip Glass worked with small repeating cells that grow by adding a note at a time.
The discipline here is restraint of a particular kind. You commit to a tiny amount of musical material and refuse to rescue it with more. The interest has to come from the process and from your attention sharpening on small changes you would miss in a busier piece. It asks the listener to do some of the work, which is what reduction usually asks.
Brian Eno gave this way of working a name. Ambient 1: Music for Airports, recorded in 1978, was built from a few looped notes and melodic fragments chosen so they would never clash, layered at lengths that keep them from lining up the same way twice. He wanted music that could be ignored or listened to closely without demanding either. The composition is mostly a decision about how little to use.
Eno also made a habit of asking subtractive questions. The deck of cards he made with the painter Peter Schmidt, Oblique Strategies, is full of prompts that push you to remove rather than add, including ones in the spirit of asking what you would not do. The hardest part of arranging is not coming up with parts. It is hearing which parts are hiding the song and being willing to mute them. Most arrangement is subtraction, and the skill that takes years is knowing what to leave out so the essential thing still carries on its own.
Run these together and a single craft shows up under different names. Rubin clears away what hides the song. Davis leaves the bar half empty so a phrase can ring. Cage removes the notes and lets the room speak. The minimalists keep almost nothing and trust it. Eno asks what not to play. In each case the reduction is not a shortcut or a failure to fill the space. It is the work itself, the slow business of finding what cannot be removed and trusting that it is enough.
Between December 1945 and January 1946, Pablo Picasso went to the print shop of Fernand Mourlot in Paris and drew a bull on a lithographic stone. Then he drew it again. Over eleven states, all pulled from the same stone, the animal changes. The first state is a heavy, almost academic bull, modeled with shadow and weight, the kind of beast you could lean against. By the middle states Picasso has cut the body into planes, mapping the muscles like a butcher's diagram. And by the eleventh state the bull is a few continuous lines: a horn, a back, the slung curve of the belly, the small comic detail of the genitals, and almost nothing else. You read it instantly as a bull. There is no shading, no mass, no decoration left to take away.
What makes the series matter is the order. Picasso did not start simple and add. He started full and removed, and Mourlot recorded that to reach the pure linear bull the artist had to pass through all the intermediate stages first. The simplicity at the end was earned by going through the complexity, not around it. The series became a teaching tool partly for that reason. Apple used the eleven states in its design training to show new designers how a product gets refined toward its essence by stripping away, and the example holds because it is honest about the cost. The last bull only works because the first ten happened.
Sculptors talk about reduction more literally than anyone, because their medium gives them no choice. You cannot add marble. You can only take it off, and once it is off it stays off. Michelangelo built a working philosophy around this. In his sonnets and letters he returns to the idea that the form the sculptor wants is already present in the block, and the work is to release it by removing the surplus stone. One sonnet puts it as the best artist having no concept that the marble does not already contain within its excess.
The line most people quote, "I saw the angel in the marble and carved until I set him free," is almost certainly not his. It reads like a later paraphrase, and you should treat it as folklore rather than a documented quotation. What is documented is the practice. Michelangelo often left work unfinished, and the half-emerged figures of his Prisoners, struggling out of rough stone, are the clearest picture we have of his method. Art historians call this the non-finito. Whether he meant those pieces to stay unfinished is debated, but they show the act of reduction frozen partway through, the figure arriving as the stone leaves.
Constantin Brancusi spent four decades making the same bird. The Bird in Space series, begun around 1924 in polished bronze and marble, has more than thirty versions, and none of them has wings or feathers or a beak you could name. What Brancusi kept was the upward thrust, the sense of flight rather than the anatomy of a bird. He polished the bronze until it nearly disappeared into its own reflection. A 1926 version even ended up in a United States customs dispute over whether it counted as art or as a taxable manufactured metal object, which tells you how far he had carried the reduction past anything obviously bird-shaped.
Brancusi described the goal carefully. Simplicity, he said, is not an end in art, but one arrives at simplicity in spite of oneself in approaching the real sense of things. I find that the most precise thing anyone has said about this. You do not aim at simple. You aim at the true sense of the thing, and simplicity is the residue you are left holding when you get there. Aiming at simplicity directly tends to produce something thin instead.
Henri Matisse arrived at his own reduction late and by necessity. Surgery in 1941 left him largely bedridden, unable to stand at an easel for long, so he had assistants paint sheets of paper in flat color and he cut shapes from them with scissors. He called it drawing with scissors, and described it in a 1952 interview as one movement linking line with color, contour with surface. The book Jazz, published in 1947, collected twenty of these cut-paper compositions. The Snail, finished in 1953 and now at the Tate, is close to three meters square and is built from torn-edged blocks of color spiraling loosely around a center.
You would not guess a snail from the shapes alone. The reduction here is not toward outline, the way Picasso's bull reduces, but toward pure color and the relationships between flat areas of it. Matisse threw out modeling, perspective, the brushstroke, almost the whole apparatus of painting, and kept the thing he had been chasing his entire life, which was color doing the work of structure. He got to it by removing everything that had been carrying color before.
Some artists pushed reduction until almost nothing was left, and the question becomes whether anything survives the cut. Kazimir Malevich showed Black Square in 1915, a black square on a white field, and called it the zero point of painting. He meant it as a beginning rather than an ending, a cleared ground from which a new kind of image could start. Decades later Ad Reinhardt made his black paintings, large canvases so close to uniform black that you have to stand and wait before your eyes find the faint cross or grid divisions inside them. Reinhardt was blunt that he was painting the last paintings anyone could make, removing color, gesture, subject, and incident one at a time.
Agnes Martin worked the same edge from a gentler direction. Her canvases are pale fields ruled with faint pencil grids and washes of color you can barely register from across a room. Up close the hand-drawn lines waver slightly, and that small human unsteadiness is most of what is there to see. Donald Judd removed even the hand, fabricating plain boxes in metal and plywood and insisting they were objects rather than compositions. These artists disagreed about why they were reducing, but they shared the conviction that you could keep taking away and still have a work left.
East Asian ink painting reached a related place much earlier, and from a different philosophy. In sumi-e, painting in black ink on white paper, the unpainted ground is not background waiting to be filled. The empty paper is part of the image, often the larger part, and a few brushstrokes have to imply the mountain or the bamboo while the blankness does the rest. The enso, the circle drawn in a single brushstroke in Zen practice, takes this to one mark. The circle is usually left slightly open, the ink running dry at the end, and the gap and the bare paper around it carry as much meaning as the stroke. What is left out is doing work, the same way Brancusi's polished void around the Bird matters as much as the bronze.
Here is the trap, and it is worth being plain about. All of this can be faked, and the fake looks almost identical from the outside. A blank canvas and a Reinhardt black painting are both nearly empty. A child's single brush squiggle and a master's enso are both one stroke. The difference is that reduction is a destination reached from somewhere fuller, and the work carries evidence of the journey even after the journey is hidden. Picasso's final bull contains the ten earlier bulls in the way a single line lands exactly where mass used to be. The enso is loose because the hand drew thousands first.
Emptiness arrived at without that mastery is just emptiness. It does not read as essential, it reads as unfinished or as nothing, because there is no removed weight pressing against the simplicity to give it tension. This is why reduction is a discipline and a risk rather than a shortcut. You earn the right to leave things out by knowing exactly what you are leaving out and why. The few lines at the end are expensive. They cost you everything you cut to get there, which is the whole point of the folder this piece sits in. The essential core is what remains when you have done the hard work of removing the rest, and you cannot get to it by starting with little.
In most films you would see the man cross the room, reach the door, turn the handle, and step through. Robert Bresson would cut almost all of that. A hand on a latch. The door. Feet on the floor outside. He trusted you to assemble the rest, because the brain fills gaps faster than the camera can show them. The action survives the cutting. What dies is the padding.
That instinct, to remove until only the load-bearing parts remain, runs through a strain of filmmaking that treats subtraction as the main creative act. The director and the editor are both asking the same question all day. What can I take out and still be understood? This piece is about the people who answered that question well, and about why the answer is harder than it sounds.
Robert Bresson published his "Notes sur le cinematographe" in 1975, a thin book of aphorisms collected over decades of work on films like "Diary of a Country Priest" (1951) and "Pickpocket" (1959). He drew a hard line between filmed theater, which he disliked, and what he called cinematography, an art made of moving images and recorded sound rather than borrowed stage performance.
The practice followed the theory. Bresson stopped using trained actors and worked instead with non-professionals he called models, asking them to deliver lines flat, drained of the inflection an actor would add. He stripped out scoring, stripped out expressive faces, stripped out most of what an audience expects to feel guided by. The aim was to remove the performer's interpretation so that something truer could show through. One of his notes reads, "Make visible what, without you, might never have been seen." The instruction is reductive and demanding at once. Clear the screen of the obvious so the unseen thing has room to appear.
This is the risk written plainly. Take away the music, the acting, the camera moves, and you can be left with nothing. Bresson got away with it because under the bareness there was control of rhythm, framing, and sound that most directors never reach. Restraint without that mastery is just an empty room.
Alfred Hitchcock is usually credited with the line, "Drama is life with the dull bits cut out." The attribution holds, though the wording came to us through a journalist. In 1956 the syndicated columnist Leonard Lyons quoted Hitchcock, around the release of "The Man Who Knew Too Much," saying that drama is life with the dull bits cut out. The phrasing has drifted slightly in later quote books, so treat the exact words as approximate and the idea as firmly his.
The idea governed how he built films. Hitchcock showed you what the story needed and almost nothing else. A character travels across a country in a few seconds of screen time because the journey is dull and the arrival is not. He spent his economy where tension lived and refused to spend it anywhere else. Drama, in his hands, was ordinary time with the slack removed.
Yasujiro Ozu reduced a different element. Where Bresson cut performance, Ozu cut motion and incident. In "Tokyo Story" (1953), an aging couple visits their grown children in the city and finds them too busy to care. Ozu shot most of it from a low position, roughly the eye level of a person kneeling on a tatami mat, with the camera almost always locked in place. Accounts of the film note only a single moving shot in the whole picture.
He also removed the events a conventional drama would build toward. The big confrontation happens off screen or not at all. A death is registered in glances and silences rather than scored breakdowns. By taking out the camera movement and the manufactured turning points, Ozu left the small things visible, a daughter-in-law's patience, a father's quiet, the train pulling away. The restraint is the meaning, not a frame around it.
Editing is where removal becomes the literal job. Walter Murch, who cut "Apocalypse Now" and "The Conversation," wrote about this in "In the Blink of an Eye," first published in 1992 and widely read in the 1995 American edition. A film, in his account, is made as much by what gets dropped as by what gets shot. Every cut is a decision to end one idea and begin another, and the footage left on the floor is part of the result.
Murch is also useful on how to choose. He ranked the things a cut should honor and put emotion at the top, assigning it roughly half the weight, well above story, rhythm, and the spatial concerns that beginners worry about first. His rule was blunt. If you have to give something up, never give up emotion before story. That gives the editor a way to decide what is essential and what is merely present. Keep the cut that makes the audience feel the right thing. Lose the rest, however well it was filmed.
Screenwriting carries the same discipline under the old advice to show rather than tell. Exposition is the writing equivalent of the dull bits. A character who announces how he feels has stolen work from the image that could have shown it. Cutting that line trusts the audience the way Bresson trusted them to finish the man walking through the door.
Hemingway described the principle for prose as the iceberg, the claim that what a writer leaves out still presses on the part the reader sees. On screen the omission works the same way. An unspoken grievance, a fact withheld until the right moment, a scene that starts after the argument has already begun, all of these register precisely because they are not stated. The common editing maxim, enter late and leave early, is the same instinct applied to scene length. Arrive after the setup, leave before the wind-down, and let the audience carry the connective tissue.
Constraint can be formalized. In 1995 Lars von Trier and Thomas Vinterberg published the Dogme 95 manifesto with its Vow of Chastity, a set of rules that banned added music, special lighting, props brought to location, and other comforts of conventional production. Von Trier was open that the rules were somewhat arbitrary. The point was to force invention by removing the easy options. Whether the films justified the vow is a separate argument, but the logic is the reductive logic in its purest form. Take the tools away and see what the work has to find instead.
Removal is painful because you rarely cut your worst material. You cut good scenes, strong shots, lines an actor nailed, because they slow the whole or pull attention from what matters more. The skill is holding the finished film in mind and judging each piece against it rather than against its own quality. A beautiful shot that serves nothing is the first thing a good editor loses.
The same skill carries the same danger. Strip too much, or strip without knowing what the work is for, and you get emptiness sold as discipline. Bresson and Ozu survived radical subtraction because their control of everything left on screen was total. The lesson is not that less is automatically more. The lesson is that you cannot find the essential core of a thing until you are willing to remove everything that is not it, and that you had better understand the thing first. Reduction is a method, not a guarantee.
William Strunk Jr. wrote the most famous instruction in American prose in a single paragraph, and you can read it in the time it takes to drink cold coffee. "Omit needless words. Vigorous writing is concise. A sentence should contain no unnecessary words, a paragraph no unnecessary sentences, for the same reason that a drawing should have no unnecessary lines and a machine no unnecessary parts." That is Rule 17 of The Elements of Style, the rule E.B. White, Strunk's former student, kept and amplified when he revised the little book in 1959. The instruction is itself an example of the thing it asks for. Strunk does not say that the writer should make all sentences short. He says every word should tell.
The reason to start here, in a collection about finding the core of a thing by taking away what is not essential, is that writing is the one craft where the subtraction is most visible on the page. A sculptor removes stone you never see. A writer crosses out a sentence that was, a moment ago, the thing you were proud of. The discipline of writing is mostly the discipline of removal, and most of the names worth knowing in it earned them by cutting.
Ernest Hemingway gave the clearest theory of why omission works. In chapter sixteen of Death in the Afternoon, published in 1932, he compared good prose to an iceberg. "The dignity of movement of an iceberg is due to only one-eighth of it being above water." The part below the surface is the part the writer knows and does not write down. His claim is precise and worth quoting: "If a writer of prose knows enough about what he is writing about, he may omit things that he knows, and the reader, if the writer is writing truly enough, will have a feeling of those things as strongly as though the writer had stated them."
The condition in that sentence is the whole of it. The writer has to know the thing before leaving it out. The reader can feel the difference between a silence the author chose and a silence the author could not fill. The famous six-word story often hung on Hemingway, "For sale: baby shoes, never worn," is usually offered as proof of his method. There is no good evidence he wrote it. The earliest versions of the idea predate him, and researchers have never found it in his papers, so treat it as a story about the theory rather than an example of it.
The hardest case to think about is Raymond Carver, because the spareness most readers think of as his was largely the work of his editor, Gordon Lish. When What We Talk About When We Talk About Love appeared in 1981, it set Carver's reputation as the leading writer of what critics called minimalism. The manuscripts, opened to scholars later, show that Lish cut the collection by roughly half. Several of the best known stories lost between fifty and almost eighty percent of their words. Lish did more than trim. He changed endings, renamed characters, and wrote new lines. He removed the warmth and sentiment in Carver's drafts and left the bleakness exposed.
Carver was grateful and then anguished. He wrote to Lish asking him to stop the book, afraid he would be exposed as a fraud who had not written his own style. After Carver died, his widow published some of the stories in their fuller original form so readers could compare. You can read both versions and decide which you prefer. What the case proves is that reduction is a real force on the page, strong enough to make a writer's name and strong enough to be a kind of theft. The cut was not neutral. Someone chose what was essential, and it was not always the author.
Elmore Leonard turned the same instinct into rules a working novelist could use. His list of ten, printed in The New York Times in 2001, ends with the only one he said really mattered: "Try to leave out the part that readers tend to skip." He meant the thick descriptive paragraphs, the second pass at the weather, the writing that announces itself as writing. His summary of the whole list was blunter still. "If it sounds like writing, I rewrite it." The goal was to disappear, to keep the reader inside the story and not standing back admiring the prose.
This is reduction aimed at a different target than word count. Leonard is cutting the apparatus, the visible effort, the parts that exist to impress rather than to carry the story. A sentence can be short and still sound like writing. The test he proposes is closer to the read-aloud test than to the word count: does the sentence call attention to itself, and if so, out it goes.
The most quoted line about this is also the most misattributed. "Murder your darlings" comes from Arthur Quiller-Couch, who said it in a lecture at Cambridge, published in 1916 in On the Art of Writing. His full advice is gentler and more exact than the slogan: "Whenever you feel an impulse to perpetrate a piece of exceptionally fine writing, obey it, whole-heartedly, and delete it before sending your manuscript to press." The phrase gets handed to William Faulkner, to Stephen King, to Oscar Wilde, to Mark Twain. None of them coined it. Quiller-Couch did, and almost no one credits him, partly because his name has faded and the others sell better as authorities.
The point survives the confusion over who said it. The sentence you most want to keep is often the one to cut, because the pleasure you take in it is your pleasure, not the reader's. You wrote it to show what you could do. That is exactly why it does not belong.
Reduction in writing runs along two lines. One is the haiku tradition, where the form itself enforces a small space and the poet works inside it. Matsuo Basho built whole scenes in seventeen syllables, trusting the reader's mind to supply the rest, which is Hemingway's iceberg arriving three centuries early in another country. The other line is the removal of the machinery around the words. Cormac McCarthy stripped his punctuation to almost nothing, no quotation marks, no semicolons, colons only to introduce a list. He told Oprah Winfrey that James Joyce was his model and that there was no reason to "blot the page up with weird little marks." Write clearly enough, McCarthy held, and the punctuation becomes unnecessary, though he also warned that writing dialogue this way takes more care, not less, so the reader always knows who is speaking.
Anton Chekhov gave the structural version of the same idea. If a rifle hangs on the wall in the first act, it has to be fired by the last, and if nothing will fire it, take it down. Anything that does not do work should not be there. That is a rule about removal as much as about setup, and it scales from a prop to a paragraph to a clause.
None of this means shorter is always better. Spareness without substance is just thin. A page can be cut until it has nothing left to say, and the reader feels that emptiness as surely as they feel padding. Hemingway's condition holds the whole theory together: you may omit only what you know. The Carver case cuts both ways, because some readers find Lish's versions sharper and others find them cold, hollowed out where the author had put something real. Reduction is a judgment about what is essential, and the judgment can be wrong.
So the work is revision, and revision is mostly subtraction. You write more than you need, then you take away what does not tell, then you read it aloud and take away what sounds like writing. What remains, if you knew the thing in the first place, is the core. That is the whole of the method, and it is harder than it sounds, because the words you have to remove are usually your own and usually the ones you liked best.
When Steve Jobs came back to Apple in 1997, the company was selling so many overlapping computers that even Jobs could not tell which one to recommend to a friend. The story that has been told many times since, and that holds up across the accounts, is that he walked to a whiteboard, drew a two-by-two grid, and wrote "Consumer" and "Pro" across the top and "Desktop" and "Portable" down the side. Four boxes. One great product for each. Everything that did not fit a box was canceled. Apple lost over a billion dollars in fiscal 1997 and turned a profit the next year, and while the four-quadrant grid was not the only reason, it is the cleanest illustration of what the turnaround was actually about. It was about subtraction.
Jobs explained the thinking at Apple's developer conference that same year. "People think focus means saying yes to the thing you've got to focus on," he said. "But that's not what it means at all. It means saying no to the hundred other good ideas that there are. You have to pick carefully. I'm actually as proud of the things we haven't done as the things I have done. Innovation is saying no to a thousand things." That last line is the one people quote. The line before it is the one that matters. He was proud of what Apple did not ship.
I work on products, and I can tell you the default direction is always more. More is easy. Adding a feature feels like progress, it satisfies the person who asked for it, and it ships without anyone having to defend a cut. Saying no requires a reason, a conversation, and usually a disappointed stakeholder. So features accumulate, not because anyone decided the product should be larger, but because no single addition was worth the fight to refuse.
The cost of that drift is real and it keeps arriving long after the feature ships. Every feature is code that has to be maintained, a path that has to be tested, a question support has to answer, and one more thing the user has to look at and skip over to find what they came for. The user pays the largest share of that bill, and they pay it on every visit, while the team paid once. Reduction is mostly the discipline to count that bill before adding to it, and the courage to cut scope when the honest answer is that something is not earning its place.
The clearest modern statement of this comes from Dieter Rams, who ran design at Braun from the 1960s and wrote ten principles for good design in the late 1970s. The tenth is the one everyone remembers: good design is as little design as possible. Rams did not say "less is more," which still implies more would be better. He said "less, but better," which means strip the product to its essentials so it can do those well. His Braun radios, calculators, and record players were the working proof.
Jony Ive, who led design at Apple for two decades, has called Rams's work "beyond improvement," and the lineage is not subtle. The original iPod owes a clear debt to Braun's T3 pocket radio, and the iPhone calculator app sits next to Braun's ET66 calculator like a portrait of its parent. Ive's own definition is worth keeping. "Simplicity is not the absence of clutter," he has said. "That's a consequence of simplicity. Simplicity is somehow essentially describing the purpose and place of an object and product." A clean surface is the result, not the goal. The goal is understanding the thing well enough to know what it is for, and then removing whatever does not serve that.
You can see the method in the products as much as the quotes. The original iPod was a music player with most of the controls of a music player removed, leaving a wheel. The original iPhone took the existing smartphone, a thing with a keyboard and a stylus, and deleted both. Each was a reduction of a category that already existed, and the reduction was the product. John Maeda put the principle plainly in his 2006 book "The Laws of Simplicity." His first law: the simplest way to achieve simplicity is through thoughtful reduction. The word doing the work there is "thoughtful." Reduction without judgment is just deletion.
Software arrived at this independently, decades earlier. Doug McIlroy, who ran the research center at Bell Labs where Unix was built and who invented the pipe, summed up the Unix philosophy in a sentence that programmers still recite: write programs that do one thing and do it well. The corollary was that programs should work together, each small and sharp, rather than one program trying to do everything. Ken Thompson and Dennis Ritchie built the system around that idea. The reason a Unix command line still composes cleanly fifty years later is that nobody tried to make any single tool do too much.
The same instinct shows up in how some companies build products. Jason Fried and David Heinemeier Hansson at 37signals, the makers of Basecamp, wrote two books, "Getting Real" and "Rework," that read almost entirely as arguments for doing less: build half a product instead of a half-built one, say no by default, treat every feature request as something that has to earn its way in rather than something that gets added because it was asked for. Their stated aim is the smallest set of features that does the job, and their products are deliberately smaller than their competitors'.
Eric Ries gave the startup version of the idea a name in his 2011 book "The Lean Startup." The minimum viable product is the smallest thing you can ship to learn whether you are building something people want. The point is to ship the essential first and find out, rather than spend a year polishing features that turn out not to matter. Fried has pushed back on the framing, arguing that you should build something to ship rather than something to test, but the disagreement is about emphasis. Both of them are telling you to stop adding and start with the core.
There is a trap in all of this worth naming. You cannot always make a thing simpler by removing parts, because some of the complexity is real and belongs to the problem itself. Larry Tesler, who worked on the early interactive systems at Xerox PARC, described this as the conservation of complexity, now usually called Tesler's Law. Every system has some irreducible complexity. The only question is who absorbs it: the user, the application developer, or the platform underneath.
Tesler's argument was that the engineer should take the complexity so the user does not have to. "If a million users each waste a minute a day dealing with complexity that an engineer could have eliminated in a week," he wrote, "you are penalizing the user to make the engineer's job easier." This is the honest version of reduction. You are not making complexity disappear. You are deciding to carry it yourself so the person using the product never sees it.
Reduction can be taken too far, and pretending otherwise would be dishonest. The goal is the essential, not the fewest possible. When Apple removed the headphone jack from the iPhone, plenty of people genuinely needed it, and the removal solved Apple's problem more than the user's. Killing a feature that a real group of users depends on is not discipline. It is a different mistake wearing the costume of the first. The test is not "can this be removed" but "does the thing still do what it is for once this is gone." If the answer is no, you have cut into the bone.
Antoine de Saint-Exupéry, writing in 1939 about the slow refinement of aircraft, put the standard better than any design manual has since. Perfection is reached, he wrote, not when there is nothing more to add, but when there is nothing left to take away. The sentence gets quoted in engineering circles because it names the actual target. Not emptiness. The point at which everything that remains is load-bearing.
The pattern across all of these, Jobs and Rams and Ive, Maeda and McIlroy and Fried, is that they treated saying no as the real work and saying yes as the easy part. Finding the essential core of a product means doing the harder thing on purpose: refusing good ideas, carrying complexity yourself so the user does not have to, and stopping exactly when the thing does what it is for and not one feature later. More is the lazy default. Just enough is a decision you have to make again every time.
These are the books, essays, recordings, and papers behind the articles in this folder. I have linked each to a legitimate source: a publisher, a museum, an author's own site, a journal, or Project Gutenberg for anything in the public domain. A few public-domain works are also hosted in this folder so you can read them here. Where a work is still in copyright, the link goes to a place you can buy or borrow it, not to a copy I have reproduced.
Rick Rubin, The Creative Act: A Way of Being (Penguin Press, 2023). The central text on reducing rather than producing. Publisher page.
Dieter Rams, the ten principles of good design, including "good design is as little design as possible." Maintained by Vitsoe, where Rams worked. vitsoe.com.
Antoine de Saint-Exupery, Wind, Sand and Stars (Terre des Hommes, 1939), the source of "perfection is reached not when there is nothing more to add, but when there is nothing left to take away." Background.
Gabrielle Adams, Benjamin Converse, Andrew Hales, and Leidy Klotz, "People systematically overlook subtractive changes," Nature (2021). The research that explains why reduction takes effort. Nature. See also Klotz, Subtract: The Untapped Science of Less (Flatiron, 2021).
Robert Browning, "Andrea del Sarto" (1855), the poem that coined "less is more." Hosted in this folder as a reading page, and at the Poetry Foundation.
Adolf Loos, "Ornament and Crime" (lecture, c. 1910; published 1913). Overview and sources.
Mies van der Rohe, the Farnsworth House (1951) and the Barcelona Pavilion (1929). Farnsworth House, Barcelona Pavilion.
Tadao Ando, Church of the Light (1989). Pritzker laureate page. John Pawson on minimalism: johnpawson.com.
Johnny Cash, American Recordings (1994), produced by Rick Rubin: voice and guitar, almost nothing else. Background.
Miles Davis, Kind of Blue (1959), modal jazz and space. Background.
John Cage, 4'33" (1952). The John Cage Trust.
Brian Eno, Ambient 1: Music for Airports (1978), and the Oblique Strategies deck made with Peter Schmidt. EnoShop. On minimalism, see Steve Reich and Terry Riley's In C (1964).
Pablo Picasso, The Bull (Le Taureau), eleven lithographic states, 1945–46. Museum of Modern Art.
Constantin Brancusi, the Bird in Space series. Centre Pompidou.
Michelangelo, the unfinished Prisoners (Slaves), Galleria dell'Accademia, Florence. Accademia.
Henri Matisse, the cut-outs, including The Snail (1953). Tate. On the empty ground, see sumi-e ink painting and the enso.
Robert Bresson, Notes on the Cinematograph (1975). NYRB Classics.
Yasujiro Ozu, Tokyo Story (1953). Criterion.
Walter Murch, In the Blink of an Eye (1995), on editing as removal. Book listing.
Alfred Hitchcock, "drama is life with the dull bits cut out." Wikiquote, with sourcing. The Dogme 95 manifesto (1995): overview.
William Strunk Jr., The Elements of Style (1918), "omit needless words." Hosted in this folder as a PDF, and at Project Gutenberg. The expanded edition revised by E. B. White is the one in print.
Arthur Quiller-Couch, On the Art of Writing (1916), the source of "murder your darlings." Hosted in this folder (the full PDF, and the "On Style" lecture as a reading page), and at Project Gutenberg.
Ernest Hemingway's iceberg theory, set out in Death in the Afternoon (1932). Explainer.
Raymond Carver and his editor Gordon Lish: D. T. Max, "The Carver Chronicles," The New York Times Magazine (1998). NYT.
Elmore Leonard, "Easy on the Adverbs, Exclamation Points and Especially Hooptedoodle," The New York Times (2001). NYT.
Steve Jobs on focus, "saying no to a thousand things," from his 1997 return to Apple. Wikiquote, with sourcing.
Jony Ive and the lineage from Dieter Rams. See the Vitsoe page above and this Braun-to-Apple comparison.
John Maeda, The Laws of Simplicity (MIT Press, 2006). MIT Press.
The Unix philosophy, "do one thing and do it well" (Doug McIlroy). The Art of Unix Programming.
Jason Fried and David Heinemeier Hansson, Getting Real (free to read) and Rework (2010). Getting Real, Rework.
Eric Ries, The Lean Startup (2011), the minimum viable product. theleanstartup.com. Larry Tesler's Law of conservation of complexity: Laws of UX.
Arthur Quiller-Couch · Lecture XII, On the Art of Writing (1916)
A public-domain reading from this library. This is the twelfth and final lecture, "On Style," from Arthur Quiller-Couch's On the Art of Writing (1916), delivered at Cambridge. It is the original source of the advice now usually shortened to "kill your darlings." His own phrasing, near the end, is "Murder your darlings." The writing article in this folder discusses it. Text from Project Gutenberg; the lecture is in the public domain.
Wednesday, January 28, 1914
Should Providence, Gentlemen, destine any one of you to write books for his living, he will find experimentally true what I here promise him, that few pleasures sooner cloy than reading what the reviewers say. This promise I hand on with the better confidence since it was endorsed for me once in conversation by that eminently good man the late Henry Sidgwick; who added, however, 'Perhaps I ought to make a single exception. There was a critic who called one of my books "epoch-making." Being anonymous, he would have been hard to find and thank, perhaps; but I ought to have made the effort.'
May I follow up this experience of his with one of my own, as a preface or brief apology for this lecture? Short-lived as is the author's joy in his critics, far-spent as may be his hope of fame, mournful his consent with Sir Thomas Browne that 'there is nothing immortal but immortality,' he cannot hide from certain sanguine men of business, who in England call themselves 'Press-Cutting Agencies,' in America 'Press-Clipping Bureaux,' and, as each successive child of his invention comes to birth, unbecomingly presume in him an almost virginal trepidation. 'Your book,' they write falsely, 'is exciting much comment. May we collect and send you notices of it appearing in the World's Press? We submit a specimen cutting with our terms; and are, dear Sir,' etc.
Now, although steadily unresponsive to this wile, I am sometimes guilty of taking the enclosed specimen review and thrusting it for preservation among the scarcely less deciduous leaves of the book it was written to appraise. So it happened that having this vacation, to dust--not to read--a line of obsolete or obsolescent works on a shelf, I happened on a review signed by no smaller a man than Mr Gilbert Chesterton and informing the world that the author of my obsolete book was full of good stories as a kindly uncle, but had a careless or impatient way of stopping short and leaving his readers to guess what they most wanted to know: that, reaching the last chapter, or what he chose to make the last chapter, instead of winding up and telling 'how everybody lived ever after,' he (so to speak) slid you off his avuncular knee with a blessing and the remark that nine o'clock was striking and all good children should be in their beds.
That criticism has haunted me during the vacation. Looking back on a course of lectures which I deemed to be accomplished; correcting them in print; revising them with all the nervousness of a beginner; I have seemed to hear you complain--'He has exhorted us to write accurately, appropriately; to eschew Jargon; to be bold and essay Verse. He has insisted that Literature is a living art, to be practised. But just what we most needed he has not told. At the final doorway to the secret he turned his back and left us. Accuracy, propriety, perspicuity--these we may achieve. But where has he helped us to write with beauty, with charm, with distinction? Where has he given us rules for what is called _Style_ in short?--having attained which an author may count himself set up in business.'
Thus, Gentlemen, with my mind's ear I heard you reproaching me. I beg you to accept what follows for my apology.
To begin with, let me plead that you have been told of one or two things which Style is _not_; which have little or nothing to do with Style, though sometimes vulgarly mistaken for it. Style, for example, is not--can never be--extraneous Ornament. You remember, may be, the Persian lover whom I quoted to you out of Newman: how to convey his passion he sought a professional letter-writer and purchased a vocabulary charged with ornament, wherewith to attract the fair one as with a basket of jewels. Well, in this extraneous, professional, purchased ornamentation, you have something which Style is not: and if you here require a practical rule of me, I will present you with this: 'Whenever you feel an impulse to perpetrate a piece of exceptionally fine writing, obey it --whole-heartedly--and delete it before sending your manuscript to press. _Murder your darlings._'
But let me plead further that you have not been left altogether without clue to the secret of what Style is. That you must master the secret for yourselves lay implicit in our bargain, and you were never promised that a writer's training would be easy. Yet a clue was certainly put in your hands when, having insisted that Literature is a living art, I added that therefore it must be personal and of its essence personal.
This goes very deep: it conditions all our criticism of art. Yet it conceals no mystery. You may see its meaning most easily and clearly, perhaps, by contrasting Science and Art at their two extremes--say Pure Mathematics with Acting. Science as a rule deals with things, Art with man's thought and emotion about things. In Pure Mathematics things are rarefied into ideas, numbers, concepts, but still farther and farther away from the individual man. Two and two make four, and fourpence is not ninepence (or at any rate four is not nine) whether Alcibiades or Cleon keep the tally. In Acting on the other hand almost everything depends on personal interpretation--on the gesture, the walk, the gaze, the tone of a Siddons, the _ruse_ smile of a Coquelin, the exquisite, vibrant intonation of a Bernhardt. 'English Art?' exclaimed Whistler, 'there is no such thing! Art is art and mathematics is mathematics.' Whistler erred. Precisely because Art is Art, and Mathematics is Mathematics and a Science, Art being Art can be English or French; and, more than this, must be the personal expression of an Englishman or a Frenchman, as a 'Constable' differs from a 'Corot' and a 'Whistler' from both. Surely I need not labour this. But what is true of the extremes of Art and Science is true also, though sometimes less recognisably true, of the mean: and where they meet and seem to conflict (as in History) the impact is that of the personal or individual mind upon universal truth, and the question becomes whether what happened in the Sicilian Expedition, or at the trial of Charles I, can be set forth naked as an alegebraical sum, serene in its certainty, indifferent to opinion, uncoloured in the telling as in the hearing by sympathy or dislike, by passion or by character. I doubt, while we should strive in history as in all things to be fair, if history can be written in that colourless way, to interest men in human doings. I am sure that nothing which lies further towards imaginative, creative, Art can be written in that way.
It follows then that Literature, being by its nature personal, must be by its nature almost infinitely various. 'Two persons cannot be the authors of the sounds which strike our ear; and as they cannot be speaking one and the same speech, neither can they be writing one and the same lecture or discourse.' _Quot homines tot sententiae._ You may translate that, if you will, 'Every man of us constructs his sentence differently'; and if there be indeed any quarrel between Literature and Science (as I never can see why there should be), I for one will readily grant Science all her cold superiority, her ease in Sion with universal facts, so it be mine to serve among the multifarious race who have to adjust, as best they may, Science's cold conclusions (and much else) to the brotherly give-and-take of human life.
_Quicquid agunt homines, votum, timor, ira, voluptas..._ Is it possible, Gentlemen, that you can have read one, two, three or more of the acknowledged masterpieces of literature without having it borne in on you that they are great because they are alive, and traffic not with cold celestial certainties, but with men's hopes, aspirations, doubts, loves, hates, breakings of the heart; the glory and vanity of human endeavour, the transience of beauty, the capricious uncertain lease on which you and I hold life, the dark coast to which we inevitably steer; all that amuses or vexes, all that gladdens, saddens, maddens us men and women on this brief and mutable traject which yet must be home for a while, the anchorage of our hearts? For an instance:--
Here lies a most beautiful lady, Light of step and heart was she: I think she was the most beautiful lady That ever was in the West Country. But beauty vanishes, beauty passes, However rare, rare it be; And when I crumble who shall remember That lady of the West Country?
(Walter de la Mare.)
Or take a critic--a literary critic--such as Samuel Johnson, of whom we are used to think as of a man artificial in phrase and pedantic in judgment. He lives, and why? Because, if you test his criticism, he never saw literature but as a part of life, nor would allow in literature what was false to life, as he saw it. He could be wrong-headed, perverse; could damn Milton because he hated Milton's politics; on any question of passion or prejudice could make injustice his daily food. But he could not, even in a friend's epitaph, let pass a phrase (however well turned) which struck him as empty of life or false to it. All Boswell testifies to this: and this is why Samuel Johnson survives.
Now let me carry this contention--that all Literature is personal and therefore various--into a field much exploited by the pedant, and fenced about with many notice-boards and public warnings. _'Neologisms not allowed here,' 'All persons using slang, or trespassing in pursuit of originality....'_
Well, I answer these notice-boards by saying that, literature being personal, and men various--and even the "Oxford English Dictionary" being no Canonical book--man's use or defiance of the dictionary depends for its justification on nothing but his success: adding that, since it takes all kinds to make a world, or a literature, his success will probably depend on the occasion. A few months ago I found myself seated at a bump-supper next to a cheerful youth who, towards the close, suggested thoughtfully, as I arose to make a speech, that, the bonfire (which of course he called the 'bonner') being due at nine-thirty o'clock, there was little more than bare time left for 'langers and godders.' It cost me, who think slowly, some seconds to interpret that by 'langers' he meant 'Auld Lang Syne' and by 'godders' 'God Save the King.' I thought at the time, and still think, and will maintain against any schoolmaster, that the neologisms of my young neighbour, though not to be recommended for essays or sermons, did admirably suit the time, place, and occasion.
Seeing that in human discourse, infinitely varied as it is, so much must ever depend on _who_ speaks, and to _whom_, in what mood and upon what occasion; and seeing that Literature must needs take account of all manner of writers, audiences, moods, occasions; I hold it a sin against the light to put up a warning against any word that comes to us in the fair way of use and wont (as 'wire,' for instance, for a telegram), even as surely as we should warn off hybrids or deliberately pedantic impostors, such as 'antibody' and 'picture-drome'; and that, generally, it is better to err on the side of liberty than on the side of the censor: since by the manumitting of new words we infuse new blood into a tongue of which (or we have learnt nothing from Shakespeare's audacity) our first pride should be that it is flexible, alive, capable of responding to new demands of man's untiring quest after knowledge and experience. Not because it was an ugly thing did I denounce Jargon to you, the other day: but because it was a dead thing, leading no-whither, meaning naught. There is _wickedness_ in human speech, sometimes. You will detect it all the better for having ruled out what is _naughty_.
Let us err, then, if we err, on the side of liberty. I came, the other day, upon this passage in Mr Frank Harris's study of 'The Man Shakespeare':--
In the last hundred years the language of Moliere has grown fourfold; the slang of the studios and the gutter and the laboratory, of the engineering school and the dissecting table, has been ransacked for special terms to enrich and strengthen the language in order that it may deal easily with the new thoughts. French is now a superb instrument, while English is positively poorer than it was in the time of Shakespeare, thanks to the prudery of our illiterate middle class.[1]
Well, let us not lose our heads over this, any more than over other prophecies of our national decadence. The "Oxford English Dictionary" has not yet unfolded the last of its coils, which yet are ample enough to enfold us in seven words for every three an active man can grapple with. Yet the warning has point, and a particular point, for those who aspire to write poetry: as Francis Thompson has noted in his Essay on Shelley:--
Theoretically, of course, one ought always to try for the best word. But practically, the habit of excessive care in word-selection frequently results in loss of spontaneity; and, still worse, the habit of always taking the best word too easily becomes the habit of always taking the most ornate word, the word most removed from ordinary speech. In consequence of this, poetic diction has become latterly a kaleidoscope, and one's chief curiosity is as to the precise combinations into which the pieces will be shifted. There is, in fact, a certain band of words, the Praetorian cohorts of Poetry, whose prescriptive aid is invoked by every aspirant to the poetic purple.... Against these it is time some banner should be raised.... It is at any rate curious to note that the literary revolution against the despotic diction of Pope seems issuing, like political revolutions, in a despotism of his own making;
and he adds a note that this is the more surprising to him because so many Victorian poets were prose-writers as well.
Now, according to our theory, the practice of prose should maintain fresh and comprehensive a poet's diction, should save him from falling into the hands of an exclusive coterie of poetic words. It should react upon his metrical vocabulary to its beneficial expansion, by taking him outside his aristocratic circle of language, and keeping him in touch with the great commonalty, the proletariat of speech. For it is with words as with men: constant intermarriage within the limits of a patrician clan begets effete refinement; and to reinvigorate the stock, its veins must be replenished from hardy plebeian blood.
In diction, then, let us acquire all the store we can, rejecting no coin for its minting but only if its metal be base. So shall we bring out of our treasuries new things and old.
Diction, however, is but a part of Style, and perhaps not the most important part. So I revert to the larger question, 'What is Style? What its [Greek: to ti en einai], its essence, the law of its being?'
Now, as I sat down to write this lecture, memory evoked a scene and with the scene a chance word of boyish slang, both of which may seem to you irrelevant until, or unless, I can make you feel how they hold for me the heart of the matter.
I once happened to be standing in a corner of a ball-room when there entered the most beautiful girl these eyes have ever seen or now--since they grow dull--ever will see. It was, I believe, her first ball, and by some freak or in some premonition she wore black: and not pearls--which, I am told, maidens are wont to wear on these occasions--but one crescent of diamonds in her black hair. _Et vera incessu patuit dea._ Here, I say, was absolute beauty. It startled.
I think she was the most beautiful lady That ever was in the West Country. But beauty vanishes, beauty passes....
She died a year or two later. She may have been too beautiful to live long. I have a thought that she may also have been too good.
For I saw her with the crowd about her: I saw led up and presented among others the man who was to be, for a few months, her husband: and then, as the men bowed, pencilling on their programmes, over their shoulders I saw her eyes travel to an awkward young naval cadet (Do you remember Crossjay in Meredith's "The Egoist"? It was just such a boy) who sat abashed and glowering sulkily beside me on the far bench. Promptly with a laugh, she advanced, claimed him, and swept him off into the first waltz.
When it was over he came back, a trifle flushed, and I felicitated him; my remark (which I forget) being no doubt 'just the sort of banality, you know, one does come out with'--as maybe that the British Navy kept its old knack of cutting out. But he looked at me almost in tears and blurted, 'It isn't her beauty, sir. You saw? It's--it's--my God, it's the _style_!'
Now you may think that a somewhat cheap, or at any rate inadequate, cry of the heart in my young seaman; as you may think it inadequate in me, and moreover a trifle capricious, to assure you (as I do) that the first and last secret of a good Style consists in thinking with the heart as well as with the head.
But let us philosophise a little. You have been told, I daresay often enough, that the business of writing demands _two_--the author and the reader. Add to this what is equally obvious, that the obligation of courtesy rests first with the author, who invites the seance, and commonly charges for it. What follows, but that in speaking or writing we have an obligation to put ourselves into the hearer's or reader's place? It is _his_ comfort, _his_ convenience, we have to consult. To _express_ ourselves is a very small part of the business: very small and almost unimportant as compared with _impressing_ ourselves: the aim of the whole process being to persuade.
All reading demands an effort. The energy, the good-will which a reader brings to the book is, and must be, partly expended in the labour of reading, marking, learning, inwardly digesting what the author means. The more difficulties, then, we authors obtrude on him by obscure or careless writing, the more we blunt the edge of his attention: so that if only in our own interest--though I had rather keep it on the ground of courtesy--we should study to anticipate his comfort.
But let me go a little deeper. You all know that a great part of Lessing's argument in his "Laokoeon", on the essentials of Literature as opposed to Pictorial Art or Sculpture, depends on this--that in Pictorial Art or in Sculpture the eye sees, the mind apprehends, the whole in a moment of time, with the correspondent disadvantage that this moment of time is fixed and stationary; whereas in writing, whether in prose or in verse, we can only produce our effect by a series of successive small impressions, dripping our meaning (so to speak) into the reader's mind--with the correspondent advantage, in point of vivacity, that our picture keeps moving all the while. Now obviously this throws a greater strain on his patience whom we address. Man at the best is a narrow-mouthed bottle. Through the conduit of speech he can utter--as you, my hearers, can receive--only one word at a time. In writing (as my old friend Professor Minto used to say) you are as a commander filing out his battalion through a narrow gate that allows only one man at a time to pass; and your reader, as he receives the troops, has to re-form and reconstruct them. No matter how large or how involved the subject, it can be communicated only in that way. You see, then, what an obligation we owe to him of order and arrangement; and why, apart from felicities and curiosities of diction, the old rhetoricians laid such stress upon order and arrangement as duties we owe to those who honour us with their attention. '_La clarte,_' says a French writer, '_est la politesse._' [Greek: Charisi kai sapheneia thue], recommends Lucian. Pay your sacrifice to the Graces, and to [Greek: sapheneia]--Clarity--first among the Graces.
What am I urging? 'That Style in writing is much the same thing as good manners in other human intercourse?' Well, and why not? At all events we have reached a point where Buffon's often-quoted saying that 'Style is the man himself' touches and coincides with William of Wykeham's old motto that 'Manners makyth Man': and before you condemn my doctrine as inadequate listen to this from Coventry Patmore, still bearing in mind that a writer's main object is to _impress_ his thought or vision upon his hearer.
'There is nothing comparable _for moral force_ to the charm of truly noble manners....'
I grant you, to be sure, that the claim to possess a Style must be conceded to many writers--Carlyle is one--who take no care to put listeners at their ease, but rely rather on native force of genius to shock and astound. Nor will I grudge them your admiration. But I do say that, as more and more you grow to value truth and the modest grace of truth, it is less and less to such writers that you will turn: and I say even more confidently that the qualities of Style we allow them are not the qualities we should seek as a norm, for they one and all offend against Art's true maxim of avoiding excess.
And this brings me to the two great _paradoxes_ of Style. For the first (1),--although Style is so curiously personal and individual, and although men are so variously built that no two in the world carry away the same impressions from a show, there is always a norm somewhere; in literature and art, as in morality. Yes, even in man's most terrific, most potent inventions--when, for example, in "Hamlet" or in "Lear" Shakespeare seems to be breaking up the solid earth under our feet--there is always some point and standard of sanity--a Kent or an Horatio--to which all enormities and passionate errors may be referred; to which the agitated mind of the spectator settles back as upon its centre of gravity, its pivot of repose.
(2) The second paradox, though it is equally true, you may find a little subtler. Yet it but applies to Art the simple truth of the Gospel, that he who would save his soul must first lose it. Though personality pervades Style and cannot be escaped, the first sin against Style as against good Manners is to obtrude or exploit personality. The very greatest work in Literature--the "Iliad," the "Odyssey," the "Purgatorio," "The Tempest," "Paradise Lost," the "Republic," "Don Quixote"--is all
Seraphically free From taint of personality.
And Flaubert, that gladiator among artists, held that, at its highest, literary art could be carried into pure science. 'I believe,' said he, 'that great art is scientific and impersonal. You should by an intellectual effort transport yourself into characters, not draw _them_ into _yourself_. That at least is the method.' On the other hand, says Goethe, 'We should endeavour to use words that correspond as closely as possible with what we feel, see, think, imagine, experience, and reason. It is an endeavour we cannot evade and must daily renew.' I call Flaubert's the better counsel, even though I have spent a part of this lecture in attempting to prove it impossible. It at least is noble, encouraging us to what is difficult. The shrewder Goethe encourages us to exploit ourselves to the top of our bent. I think Flaubert would have hit the mark if for 'impersonal' he had substituted 'disinterested.'
For--believe me, Gentlemen--so far as Handel stands above Chopin, as Velasquez above Greuze, even so far stand the great masculine objective writers above all who appeal to you by parade of personality or private sentiment.
Mention of these great masculine 'objective' writers brings me to my last word: which is, 'Steep yourselves in _them_: habitually bring all to the test of _them_: for while you cannot escape the fate of all style, which is to be personal, the more of catholic manhood you inherit from those great loins the more you will assuredly beget.'
This then is Style. As technically manifested in Literature it is the power to touch with ease, grace, precision, any note in the gamut of human thought or emotion.
But essentially it resembles good manners. It comes of endeavouring to understand others, of thinking for them rather than for yourself--of thinking, that is, with the heart as well as the head. It gives rather than receives; it is nobly careless of thanks or applause, not being fed by these but rather sustained and continually refreshed by an inward loyalty to the best. Yet, like 'character' it has its altar within; to that retires for counsel, from that fetches its illumination, to ray outwards. Cultivate, Gentlemen, that habit of withdrawing to be advised by the best. So, says Fenelon, 'you will find yourself infinitely quieter, your words will be fewer and more effectual; and while you make less ado, what you do will be more profitable.'
[Footnote 1: 'An oration,' says Quintilian, 'may find room for almost any word saving a few indecent ones (_quae sunt parum verecunda_).' He adds that writers of the Old Comedy were often commended even for these: 'but it is enough for us to mind our present business--_sed nobis nostrum opus intueri sat est._']
INDEX
Abelard 203, 205, 212 Abercrombie, Lascelles 18 Addison, Joseph 124, 172 Alcuin 199, 200, 204, 205 Alfred, King 186 Aristophanes 192 Aristotle 128, 203, 227 Arnold, Matthew 35, 76, 139, 186, 202 "Arte of Rhetorique," Wilson's 118 Ascham, Roger 121, 188 Augustine 199
Bacon, Lord 6, 7, 10, 220, 231 Bagehot, Walter 216 "Ballata" 45 Barbour, John 112 Barrie, Sir James Matthew 17, 135 Bede 204 Beerbohm, Max 222 Belisarius 175 Bentham, Jeremy 97 "Beowulf" 159-165 Beranger, Pierre-Jean de 45 Berners, Lord 108-110,120 Bible, The: Authorised Version 53, 97, 110, 122 et seq., 141, 143, 190 Revised Version 131-133 Blair, Wilfred 80 Blake, William 12 Boccaccio 184 Boethius 203 Bologna, University of 200-1, 206 Borneil, Giraud de 181 Boswell, James 238 Bridges, Robert 19 Brooke, the Rev. Stopford A. 159 Brougham, Ld 47, 101 Browne, Sir Thomas 10, 51, 124, 168, 232 Browning, Robert 39, 186 Buffon 245 Bunyan, John 124 Burke, Edmund 27, 28, 46, 47-52, 101 Burns, Robert 45 Butler, Arthur John 20
Caedmon 163 Cambridge 201 _et seqq._ Campion, Thomas 185, 188 Carducci, Giosue 154-5 Carlyle, Thomas 18, 103, 245 Cellini, Benvenuto 41 Cervantes 7, 25 Chadwick, Professor H. M. 163 Chair of English Literature, University Ordinance 7 Chambers, E. K. 199 Champeaux, William of 205 Chaucer, Geoffrey 10, 110-111, 163, 183, 184, 219 Chesterton, Gilbert K. 233 Chichester, Richard of 211 Cicero 28, 49 Clare, John 39 Coleridge, Samuel Taylor 41, 64, 65 Conington, John 171-2 Courthope, W. J. 13, 158, 184, 199 Coverdale, Miles 124 Cowley, Abraham 185 Cowper, William 186 Crewe, Ld Chief Justice 7 Cynewulf 163
Daniel, Samuel 185, 188 Dante 77, 184 Darwin, Charles 221 Defoe, Daniel 61, 75. Dekker, Thomas 65 De La Mare, Walter 237 De Quincey, Thomas 54 Desiderius, Archbishop 199 Dionysius of Halicarnassus 28 Donne, John 102, 106, 185 Dryden, John 172, 186, 227 "Duchess of Malfy," Webster's 99 Dunbar 10
'Eliot, George' 11 Emerson, Ralph Waldo 11
Falconer, William 79 Falkner, J. Meade 168-9 Fenelon 248 FitzGerald, Edward 97 Flaubert, Gustave 247 Fletcher, John 13 Fowler, W. H. and F. G. 90, 137 Freeman, Professor E. A. 158, 160, 174-179, 186 "Froissart," Berners' 108 Froude, James Anthony 78 Fuller, Thomas 206
Gibbon, Edward 124, 216 Gildas 175 Goethe 103, 247 Gray, Thomas 11, 16, 136, 157-8, 162 Green, J. R. 158 Green, T. H. 8 Gregory the Great, Pope 199 Grierson, Professor H. J. C. 185
Hamilton, Sir William 213 Hardy, Thomas 18 Harris, Frank 240 Harvey, Gabriel 185, 216-7 Heine, Heinrich 45 Herbert, George 133 "Hero and Leander," Marlowe's 98 Herodotus 44, 63 Homer 25, 64, 69, 76-78, 80, 81, 161, 190, 228 Horace 171-2 Housman, Professor A. E. 222
Ibsen 96 Irnerius 206 Isaiah 130-133
Jackson, Dr Henry 213 Johnson, Samuel 11, 37, 69, 121, 172, 238 Jonson, Ben 129, 146, 185, 219, 220 Jowett, Benjamin 29 Jusserand, J. J. 182 Juvenal, 172
Keats, John 16, 39, 186 Kempis, Thomas a 15 Ker, Professor W. P. 160, 199 Kipling, Rudyard 61
Lamb, Charles 41 Lessing 81, 227, 244 Lindsay, the Rev. T. M., D.D. 118 Lloyd George, the Right Hon. David 137-8 Lucian 6, 160, 192, 228, 245 Lucretius 193
Malory, Sir Thomas 107-110, 120 Marlowe, Christopher 98-9, 185, 220 Marvell, Andrew 185 Mason, William 157 Masson, David 12 McKenna, the Right Hon. Reginald 137-8 Meredith, George 243, 247 Milton, John 1, 10, 16, 43, 56-62, 74-76, 124, 152, 185, 195, 238 Minto, Professor William 245 Moore, Thomas 45 Morris, William 188 Mullinger, J. Bass 205, 219 Murray, Professor Gilbert 193
Nashe, Thomas 120 Newman, Cardinal 5, 30, 31-2, 115, 134, 144, 147, 234 Newton, Sir Isaac 221 Noyes, Alfred 78 "Nut-Brown Maid, The" 111
Oates, Captain 42 Origen 195, 202 Oxford 201 _et seq._
Paris, University of 200, 205 Pater, Walter 77, 222 Patmore, Coventry 245 Payne, E. J. 100-103 "Pervigilium Veneris" 151, 194 Pheidias 14 Philosophy and Poetry 1 Piers Plowman 163, 182 "Pilgrimage to Parnassus, The" 217-220 Plato 1-4, 150, 205 Pliny 152-3 Podsnap (_see_ Freeman) Poggio 205 Pope, Alexander 157, 162 Powell, F. York 159 Provencal Song 181-183 Pythagoras 208
Quintilian 29, 140, 240
Raleigh, Professor Sir Walter 9 Rashdall, Hastings 208-213 Remigius 206 Renan 1 Reynolds, Sir Joshua 23-25
Sainte-Beuve, Charles Augustus 20 Saintsbury, Prof. George 55, 56, 187 Salamanca, University of 200 Scott, The Antarctic Expedition 42 Severus, Sulpicius 199 Shakespeare, William 15, 41, 50, 51-2, 97-100, 113, 129, 185, 190, 197, 219, 229, 246 Shaw, George Bernard 72 Shelley 40 Shirley, James 106 Sidgwick, Henry 232 Sidney, Sir Philip 41-2 Skeat, Walter W. 12 "Sonata" 45 South, Robert 102 Spenser, Edmund 185, 206, 217, 219 Stevenson, Robert Louis 133 Stubbs, Bishop W. 44 'Student's Handbook, The' 72-3 Swift, Jonathan 61 Swinburne, Algernon 196
Taylor, Jeremy 68-9 Tennyson, Lord 75, 186 Tertullian 195, 198, 202 Thackeray, William Makepeace 124 Thompson, Francis 241 Thomson, James 39 Toulouse, University of 208 Tyndale, William 122, 126, 127
Vacarius 206 Ventadour, Bernard de 181 "Venus and Adonis" 98-9 Verrall, Dr A. W. 7 Vigfusson, Gudbrand 159 Virgil 25, 80, 194, 200 Voltaire 192
Waller, Edmund 85 Walpole, Horatio 173 Walton, Isaak 70-1, 124, 201 Warton, Thomas 158 Watson, E. J. 155 Watson, William 16 Webster, John 99 Wendell, Barrett 97 Whistler, James McNeill 236 Whitman, Walt 53, 56 "Widsith" 60 Wolfe, General 134 Wood, Anthony 184 Wordsworth, William 11, 12, 55, 67, 68, 129, 146, 186, 204, 210 Wright, Aldis 12 Wyat, Sir Thomas 115-118, 184 Wyclif, John 124, 127
Yeats, William Butler 143 Young, Arthur 171
Cambridge: Printed by J. B. Peace, M.A., at the University Press.
CALLED "THE FAULTLESS PAINTER"
A public-domain reading from this library. Robert Browning published "Andrea del Sarto" in his 1855 collection Men and Women. The painter of the title, looking back on a technically perfect but uninspired career, is the source of the line "Well, less is more," which Mies van der Rohe later borrowed for architecture. The architecture article in this folder traces that path. Text from Project Gutenberg; the poem is in the public domain.
1855
But do not let us quarrel any more,
No, my Lucrezia; bear with me for once:
Sit down and all shall happen as you wish.
You turn your face, but does it bring your heart?
I'll work then for your friend's friend, never fear,
Treat his own subject after his own way,
Fix his own time, accept too his own price,
And shut the money into this small hand
When next it takes mine. Will it? tenderly?
Oh, I'll content him--but to-morrow. Love!
I often am much wearier than you think,
This evening more than usual, and it seems
As if--forgive now--should you let me sit
Here by the window with your hand in mine
And look a half-hour forth on Fiesole,
Both of one mind, as married people use,
Quietly, quietly the evening through,
I might get up to-morrow to my work
Cheerful and fresh as ever. Let us try.
To-morrow, how you shall be glad for this!
Your soft hand is a woman of itself,
And mine the man's bared breast she curls inside.
Don't count the time lost, neither; you must serve
For each of the five pictures we require:
It saves a model. So! keep looking so--
My serpentining beauty, rounds on rounds!
--How could you ever prick those perfect ears,
Even to put the pearl there! oh, so sweet--
My face, my moon, my everybody's moon,
Which everybody looks on and calls his,
And, I suppose, is looked on by in turn,
While she looks--no one's: very dear, no less.
You smile? why, there's my picture ready made,
There's what we painters call our harmony!
A common grayness silvers everything--
All in a twilight, you and I alike
--You, at the point of your first pride in me
(That's gone you know)--but I, at every point;
My youth, my hope, my art, being all toned down
To yonder sober pleasant Fiesole.
There's the bell clinking from the chapel-top;
That length of convent-wall across the way
Holds the trees safer, huddled more inside;
The last monk leaves the garden; days decrease,
And autumn grows, autumn in everything.
Eh? the whole seems to fall into a shape--
As if I saw alike my work and self
And all that I was born to be and do,
A twilight-piece. Love, we are in God's hand.
How strange now, looks the life he makes us lead;
So free we seem, so fettered fast we are!
I feel he laid the fetter: let it lie!
This chamber for example--turn your head--
All that's behind us! You don't understand
Nor care to understand about my art,
But you can hear at least when people speak:
And that cartoon, the second from the door
--It is the thing. Love! so such things should be--
Behold Madonna!--I am bold to say.
I can do with my pencil what I know,
What I see, what at bottom of my heart
I wish for, if I ever wish so deep--
Do easily, too--when I say, perfectly,
I do not boast, perhaps: yourself are judge,
Who listened to the Legate's talk last week,
And just as much they used to say in France.
At any rate 'tis easy, all of it!
No sketches first, no studies, that's long past:
I do what many dream of, all their lives,
--Dream? strive to do, and agonize to do,
And fail in doing. I could count twenty such
On twice your fingers, and not leave this town,
Who strive--you don't know how the others strive
To paint a little thing like that you smeared
Carelessly passing with your robes afloat--
Yet do much less, so much less. Someone says,
(I know his name, no matter)--so much less!
Well, less is more, Lucrezia: I am judged.
There burns a truer light of God in them,
In their vexed beating stuffed and stopped-up brain,
Heart, or whate'er else, than goes on to prompt
This low-pulsed forthright craftsman's hand of mine.
Their works drop groundward, but themselves, I know,
Reach many a time a heaven that's shut to me,
Enter and take their place there sure enough,
Though they come back and cannot tell the world.
My works are nearer heaven, but I sit here.
The sudden blood of these men! at a word--
Praise them, it boils, or blame them, it boils too.
I, painting from myself and to myself,
Know what I do, am unmoved by men's blame
Or their praise either. Somebody remarks
Morello's outline there is wrongly traced,
His hue mistaken; what of that? or else,
Sightly traced and well ordered; what of that?
Speak as they please, what does the mountain care?
Ah, but a man's reach should exceed his grasp,
Or what's a heaven for? All is silver-gray
Placid, and perfect with my art: the worse!
I know both what I want and what might gain,
And yet how profitless to know, to sigh
"Had I been two, another and myself,
Our head would have o'erlooked the world!" No doubt.
Yonder's a work now, of that famous youth
The Urbinate who died five years ago.
('Tis copied, George Vasari sent it me.)
Well, I can fancy how he did it all,
Pouring his soul, with kings and popes to see,
Reaching, that heaven might so replenish him,
Above and through his art--for it gives way;
That arm is wrongly put--and there again--
A fault to pardon in the drawing's lines,
Its body, so to speak: its soul is right,
He means right--that, a child may understand.
Still, what an arm! and I could alter it:
But all the play, the insight and the stretch--
Out of me, out of me! And wherefore out?
Had you enjoined them on me, given me soul,
We might have risen to Rafael, I and you!
Nay, Love, you did give all I asked, I think--
More than I merit, yes, by many times.
But had you--oh, with the same perfect brow,
And perfect eyes, and more than perfect mouth,
And the low voice my soul hears, as a bird
The fowler's pipe, and follows to the snare--
Had you, with these the same, but brought a mind!
Some women do so. Had the mouth there urged
"God and the glory! never care for gain.
The present by the future, what is that?
Live for fame, side by side with Agnolo!
Rafael is waiting: up to God, all three!"
I might have done it for you. So it seems:
Perhaps not. All is as God over-rules.
Beside, incentives come from the soul's self;
The rest avail not. Why do I need you?
What wife had Rafael, or has Agnolo?
In this world, who can do a thing, will not;
And who would do it, cannot, I perceive:
Yet the will's somewhat--somewhat, too, the power--
And thus we half-men struggle. At the end,
God, I conclude, compensates, punishes.
'T is safer for me, if the award be strict,
That I am something underrated here,
Poor this long while, despised, to speak the truth.
I dared not, do you know, leave home all day,
For fear of chancing on the Paris lords.
The best is when they pass and look aside;
But they speak sometimes; I must bear it all.
Well may they speak! That Francis, that first time,
And that long festal year at Fontainebleau!
I surely then could sometimes leave the ground,
Put on the glory, Rafael's daily wear,
In that humane great monarch's golden look--
One finger in his beard or twisted curl
Over his mouth's good mark that made the smile,
One arm about my shoulder, round my neck,
The jingle of his gold chain in my ear,
I painting proudly with his breath on me,
All his court round him, seeing with his eyes,
Such frank French eyes, and such a fire of souls
Profuse, my hand kept plying by those hearts--
And, best of all, this, this, this face beyond,
This in the background, waiting on my work,
To crown the issue with a last reward!
A good time, was it not, my kingly days?
And had you not grown restless . . . but I know--
'T is done and past; 't was right, my instinct said,
Too live the life grew, golden and not gray,
And I'm the weak-eyed bat no sun should tempt
Out of the grange whose four walls make his world.
How could it end in any other way?
You called me, and I came home to your heart.
The triumph was--to reach and stay there; since
I reached it ere the triumph, what is lost?
Let my hands frame your face in your hair's gold,
You beautiful Lucrezia that are mine!
"Rafael did this, Andrea painted that;
The Roman's is the better when you pray,
But still the other's Virgin was his wife--"
Men will excuse me, I am glad to judge
Both pictures in your presence; clearer grows
My better fortune, I resolve to think.
For, do you know, Lucrezia, as God lives,
Said one day Agnolo, his very self,
To Rafael's . . . I have known it all these years . . .
(When the young man was flaming out his thoughts
Upon a palace-wall for Rome to see,
Too lifted up in heart because of it)
"Friend, there's a certain sorry little scrub
Goes up and down our Florence, none cares how,
Who, were he set to plan and execute
As you are, pricked on by your popes and kings,
Would bring the sweat into that brow of yours!"
To Rafael's!--And indeed the arm is wrong.
I hardly dare . . . yet, only you to see,
Give the chalk here--quick, thus the line should go!
Ay, but the soul! he's Rafael! rub it out!
Still, all I care for, if he spoke the truth,
(What he? why, who but Michel Agnolo?
Do you forget already words like those?)
If really there was such a chance, so lost--
Is, whether you're--not grateful--but more pleased.
Well, let me think so. And you smile indeed!
This hour has been an hour! Another smile?
If you would sit thus by me every night
I should work better, do you comprehend?
I mean that I should earn more, give you more.
See, it is settled dusk now; there's a star;
Morello's gone, the watch-lights show the wall,
The cue-owls speak the name we call them by.
Come from the window, love--come in, at last,
Inside the melancholy little house
We built to be so gay with. God is just.
King Francis may forgive me: oft at nights
When I look up from painting, eyes tired out,
The walls become illumined, brick from brick
Distinct, instead of mortar, fierce bright gold,
That gold of his I did cement them with!
Let us but love each other. Must you go?
That Cousin here again? he waits outside?
Must see you--you, and not with me? Those loans?
More gaming debts to pay? you smiled for that?
Well, let smiles buy me! have you more to spend?
While hand and eye and something of a heart
Are left me, work's my ware, and what's it worth?
I'll pay my fancy. Only let me sit
The gray remainder of the evening out,
Idle, you call it, and muse perfectly
How I could paint, were I but back in France,
One picture, just one more--the Virgin's face,
Not yours this time! I want you at my side
To hear them--that is, Michel Agnolo--
Judge all I do and tell you of its worth.
Will you? To-morrow, satisfy your friend.
I take the subjects for his corridor,
Finish the portrait out of hand--there, there,
And throw him in another thing or two
If he demurs; the whole should prove enough
To pay for this same Cousin's freak. Beside,
What's better and what's all I care about,
Get you the thirteen scudi for the ruff!
Love, does that please you? Ah, but what does he,
The Cousin! what does he to please you more?
I am grown peaceful as old age to-night.
I regret little, I would change still less.
Since there my past life lies, why alter it?
The very wrong to Francis!--it is true
I took his coin, was tempted and complied,
And built this house and sinned, and all is said.
My father and my mother died of want.
Well, had I riches of my own? you see
How one gets rich! Let each one bear his lot.
They were born poor, lived poor, and poor they died:
And I have labored somewhat in my time
And not been paid profusely. Some good son
Paint my two hundred pictures--let him try!
No doubt, there's something strikes a balance. Yes,
You loved me quite enough, it seems to-night.
This must suffice me here. What would one have?
In heaven, perhaps, new chances, one more chance--
Four great walls in the New Jerusalem,
Meted on each side by the angel's reed,
For Leonard, Rafael, Agnolo and me
To cover--the three first without a wife,
While I have mine! So--still they overcome
Because there's still Lucrezia--as I choose.
Again the Cousin's whistle! Go, my Love.
This is a working reading list for someone trying to learn how hardware companies scale. It is biased toward Apple because Apple is the most studied case, but the principles transfer. The list is organized in a rough order: start with the operations and product books, then the company histories, then the technical and supply-chain titles, then the podcasts and ongoing reading.
Each entry has a one-line note on what it is for. The notes are mine. The list is opinionated.
Build: An Unorthodox Guide to Making Things Worth Making, Tony Fadell, Harper Business, 2022.
The most directly useful book on this list for someone new to hardware. Fadell built the iPod and Nest. He writes plainly about the practical work of going from idea to shipped product, and the parts about hiring, prototyping, and managing engineering teams are unusually concrete. Read this first.
High Output Management, Andrew S. Grove, Vintage (Random House), 1995 (orig. 1983).
Grove was Intel's CEO. The book is the canonical text on managing technical organizations. The framework on the manager's output ("the output of your organization plus the output of the organizations under your influence") and the discussion of one-on-ones are the parts that stay with you. Required reading for anyone who runs a team that ships things.
Only the Paranoid Survive, Andrew S. Grove, Currency Doubleday, 1996.
Grove's second book. About strategic inflection points: the moments when the rules of an industry change and the companies that do not adapt do not survive. The framing is more useful than the case studies, and the framing is genuinely good.
The Goal, Eliyahu Goldratt and Jeff Cox, North River Press, 1984.
A novel about a factory manager who is trying to save his plant. The narrative format obscures that this is the standard reference on the Theory of Constraints. The lesson, which is hard to over-stress, is that the throughput of any system is determined by the bottleneck. Optimization anywhere else is wasted work.
The Toyota Way, Jeffrey K. Liker, McGraw-Hill, 2004 (2nd ed. 2020).
The reference text on the Toyota Production System. Long. Worth working through if you want to understand where modern manufacturing operations came from. Most of what is now called "lean" started here.
Working Backwards, Colin Bryar and Bill Carr, St. Martin's Press, 2021.
The Amazon book. The PR/FAQ method, the bar-raiser hiring practice, and the section on input metrics versus output metrics are the parts that travel. Useful as a contrast to Apple. Different culture, different operating model, same level of operational seriousness.
After Steve: How Apple Became a Trillion-Dollar Company and Lost Its Soul, Tripp Mickle, William Morrow, 2022.
The most thoroughly reported recent book on the Cook era. The thesis is in the subtitle. The reporting on Jony Ive's departure and the design organization's evolution is the part to read carefully. The thesis is contested. The reporting is solid.
Tim Cook: The Genius Who Took Apple to the Next Level, Leander Kahney, Portfolio (Penguin Random House), 2019.
The closest thing to a Cook biography. Hagiographic in places. Useful for the pre-Apple Cook chapters and for the operational history of the early Cook years.
Becoming Steve Jobs, Brent Schlender and Rick Tetzeli, Crown Business, 2015.
The Jobs biography to read if you only read one. Schlender knew Jobs well over decades, and the book has access to the post-NeXT, post-return Jobs in a way the Walter Isaacson biography does not. The Jobs you want to study is the one in this book, who learned how to run a company over the second half of his career.
Inside Apple, Adam Lashinsky, Business Plus (Hachette), 2012.
The first serious reported book on how Apple actually operates. Older than the others, but the description of the DRI ("directly responsible individual") system, the secrecy compartments, and the executive operating cadence is still accurate.
Creative Selection: Inside Apple's Design Process During the Golden Age of Steve Jobs, Ken Kocienda, St. Martin's Press, 2018.
A first-person account of building the iPhone keyboard and Safari. The book's value is the description of how design decisions actually got made: through demos, in front of decision-makers, with iteration cycles measured in days. The "creative selection" framing (a hill-climbing analogy borrowed from evolutionary biology) is the most useful single concept in the book.
Chip War: The Fight for the World's Most Critical Technology, Chris Miller, Scribner, 2022.
The reference text on the geopolitics of semiconductors. Won the Financial Times Business Book of the Year. The history is well-told and the strategic framing is the standard one. Required reading if you want to understand why TSMC and the Taiwan question are the most consequential industrial topics of the next decade.
The Hardware Hacker, Andrew "bunnie" Huang, No Starch Press, 2017.
A collection of essays from someone who has spent his career building hardware in Shenzhen and writing about how the supply chain there actually works. The chapters on factories, on counterfeit components, and on the practical realities of small-volume hardware manufacturing are unique. There is no other book like this.
The Hardware Startup: Building Your Product, Business, and Brand, Renee DiResta, Brady Forrest, Ryan Vinyard, O'Reilly Media, 2015.
A practical primer on the hardware startup path. Less philosophical than Fadell's Build, more procedural. Good for the chapters on contract manufacturing, certifications, and the operational specifics that nobody teaches you anywhere else.
The Innovator's Dilemma, Clayton M. Christensen, Harvard Business School Press, 1997.
The book that introduced disruption theory. The framework has been over-applied by everyone for thirty years. Read it anyway. The hard-disk-drive case study is the most carefully argued example of the theory and the part that has held up best.
Crossing the Chasm, Geoffrey A. Moore, HarperBusiness, 1991 (3rd ed. 2014).
The book on the gap between early adopters and the early majority. The framework is dated in places. The core observation, that hardware products fail more often at the chasm than at any other transition, is still accurate.
Loonshots, Safi Bahcall, St. Martin's Press, 2019.
A book on how organizations protect speculative R&D from the operational pressure of the rest of the business. The "Bush-Vail rules" framework (named after Vannevar Bush and Theodore Vail) is the part to take. The case studies are uneven.
The Everything Store: Jeff Bezos and the Age of Amazon, Brad Stone, Little Brown, 2013.
The Amazon biography to read for the operational history. Read alongside Working Backwards if you want to understand how an operationally serious company gets built from scratch.
Acquired (Ben Gilbert and David Rosenthal). The reference podcast for company histories. Multi-hour episodes, well-researched. Most relevant episodes for this curriculum:
The Talk Show With John Gruber. Apple-focused podcast since 2007. The WWDC live shows, when Apple executives appear, are the relevant episodes. Phil Schiller (2015), Eddy Cue and Craig Federighi (2016), Schiller and Federighi (2016 and 2017).
Big Technology Podcast (Alex Kantrowitz). Regular Apple coverage. The April 2026 episode "Tim Cook Steps Down. And Who Is John Ternus? With Joanna Stern" is the best post-announcement listen.
Stratechery (Ben Thompson, stratechery.com). The standard reference for technology strategy analysis. The essays on Apple's vertical integration, on aggregation theory, and on the Apple Silicon transition are the relevant ones for this list. Specifically:
Asymco (Horace Dediu, asymco.com). The reference for Apple operations and capital strategy analysis. The "Bank of Apple" framework comes from here. Specific posts to read:
Daring Fireball (John Gruber, daringfireball.net). Apple commentary since 2002. The signal-to-noise on day-to-day Apple news is high, and the longer essays are usually worth reading.
Power On (Mark Gurman, Bloomberg). Weekly Sunday Apple newsletter. The most reliable source for advance reporting on Apple's product roadmap and corporate decisions. Subscribe.
Hardware FYI (Benji Chia, hardwarefyi.substack.com). Weekly newsletter on hardware engineering and manufacturing. Useful for the supply-chain reporting and the technical primers. Hosts the Kinetic conference.
Embedded Artistry (embeddedartistry.com). Field-manual-style reference for embedded engineering, with clean glossary entries on NPI, FATP, EVT/DVT/PVT, and the rest of the hardware vocabulary. Good when you need to look something up.
Instrumental Build Better Handbook (instrumental.com). A glossary and reference resource on hardware manufacturing terms, written for engineers and PMs.
Apple's Penn 2024 commencement speech, John Ternus, May 18, 2024. Available on YouTube as part of the University of Pennsylvania School of Engineering and Applied Science commencement ceremony. The clearest public statement of Ternus's operating philosophy.