AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Anthropic Co-founder Ben Mann on AGI Timelines, AI Safety, and the Talent Wars

· Lenny's Podcast

Benjamin Mann, co-founder of Anthropic and tech lead for product engineering, discusses why he and other safety-focused researchers left OpenAI in 2020, arguing that safety was not the top priority there. He predicts a 50% chance of superintelligence by 2028, citing accelerating scaling laws, the transition from pre-training to reinforcement learning scaling, and continued exponential progress in compute. Mann emphasizes that the narrative of AI hitting plateaus is false—progress is actually accelerating but feels slower due to compressed release cycles and benchmark saturation. On economic impact, Mann argues people are poor at modeling exponential progress and points to concrete examples already happening: 82% automated resolution rates at Intercom and 95% of code written by Claude Code at Anthropic (meaning humans write 10-20x more code, not that jobs disappear). He envisions a 'transformative AI' milestone defined by an economic turing test—AI passing turing tests for 50% of money-weighted jobs—rather than the vague AGI concept. Long-term, he questions whether capitalism itself will survive a world of labor abundance. On safety, Mann details Anthropic's approach including Constitutional AI (where models self-critique against natural language principles), RLAIF, and evidence of deceptive alignment observed in lab settings. He estimates 0-10% probability of an x-risk outcome and frames the work as critical 'on the margin' given fewer than 1,000 people worldwide focus on AI safety versus $300B in annual industry capex. He describes Anthropic's three-world theory of alignment difficulty and explains why the company's mission focus insulates it from Meta's $100M talent poaching efforts.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.