Benjamin Mann

Benjamin Mann, co-founder of Anthropic and tech lead for product engineering, discusses the pace of AI progress, transformative AI and the Economic Turing Test, job displacement, and why he left OpenAI to found a safety-first lab. The conversation also covers constitutional AI, existential risk estimates, and how Anthropic operationalizes alignment in its products and team culture.

01Key learnings

• Mann argues progress is accelerating, not plateauing: model releases now come every one to three months, and scaling laws continue to hold across many orders of magnitude.
• Define transformative AI by the Economic Turing Test: if an agent passes as a human hire for a role over a month or three, track what share of money-weighted jobs it can cover.
• Use AI tools ambitiously and persistently; Mann says retrying a task from a fresh start succeeds far more often than repeatedly patching a failed attempt.
• Expect near-term productivity expansion rather than immediate replacement, but recognize lower-skill and less-headroom roles face significant displacement pressure over time.
• Constitutional AI trains models using written principles drawn from sources like human rights declarations, letting the model critique and rewrite its own outputs, reducing reliance on human raters.
• Mann sees alignment and capability as increasingly complementary: his team's safety work produced the character and personality that made Claude popular and less sycophantic.
• Publicly sharing concerning model behaviors and testing them in lab settings builds policymaker trust and surfaces real risks before they appear in the wild.
• Product and non-research roles matter for safety: revenue, distribution, and influence fund safety research, so Mann encourages people in any function to contribute to the mission.

02Frameworks mentioned

Summary and takeaways written by PM Atlas; quotes are short excerpts. © the original author.