Strengths & Weaknesses

Simultaneously holds optimistic and pessimistic views. Authored "Machines of Loving Grace" (utopian) AND "The Adolescence of Technology" (cautionary) within 15 months. Neither cancels the other. Rare capacity to hold genuine complexity.

PhD in biophysics, co-invented RLHF, co-led GPT-2/GPT-3 development. Not a business-leader-who-learned-AI but a scientist-who-became-a-leader. Can discuss scaling laws, interpretability, and Constitutional AI with genuine depth.

Pledged 80% of ~$7B wealth. Held red lines against Pentagon pressure at enormous cost. Disclosed model failures publicly (deception, blackmail attempts). "No one elected us. Honestly, no one." Rare willingness to name uncomfortable truths about his own industry and position.

Built $19B ARR company with 2.1x more revenue per compute dollar than main competitor. $78B projected compute spend vs OpenAI's $235B. Efficiency as a genuine strategic advantage.

Despite Meta offering $100-500M to individuals, lost only 2 researchers. All 7 co-founders remain. Culture as a moat that money can't buy. "Technology can be bought, but culture can't."

Essays like "Machines of Loving Grace" (15K words) and "Adolescence of Technology" (20K words) demonstrate ability to build complex arguments that shape public discourse. The DVQ meetings (biweekly hour-long deep dives) show this translates internally.

The ASL system provides graduated, principled approach to risk vs. binary safe/unsafe. Bioweapons classifier. Constitutional AI. Interpretability investment. Systematic rather than reactive.

Self-describes as "accidental CEO of an accidental business" who "had no idea how we would make money." Growth has been rapid, but positioning, pricing, and go-to-market have sometimes lagged capabilities.

Pentagon conflict spiraled into supply chain risk designation. Leaked internal memos calling OpenAI "mendacious." Deleted Facebook post calling Trump a "feudal warlord." Multiple PR crises despite correct underlying positions.

Endorsed Kamala Harris, then needed to work with Trump administration. Pentagon standoff escalated to court threats. Lacks the political flexibility of competitors who maintained relationships across the aisle.

Safety commitment sometimes slows product cycles. While competitors ship fast, Anthropic's careful approach means features arrive later. The market doesn't always wait for the most responsible option.

Building the thing you're warning about. Critics note contradiction between "AI could be existentially dangerous" and raising hundreds of billions to build it faster. Zvi Mowshowitz identified contradictions between stated confidence and conservative business decisions.

Amazon ($8B), Google ($3B), and others create complex dependency web. Investors may eventually prioritize returns over safety mission.

"I'm deeply uncomfortable with these decisions being made by a few companies, by a few people." While admirable, this discomfort may translate to hesitation when decisive action is needed. The person who's uncomfortable with power may be the right person to hold it, but they may also be slower to wield it.