Richard Socher on recursive self-improvement: AI research systems beating humans in two days
Latent Space · youtube · 2026-09-14
In a Latent Space interview, Richard Socher (You.com, Recursive) lays out his bet on recursive self-improvement and 'The Eureka Machine': AI that automates AI research itself. Highlights include Recursive's research system outperforming humans on optimization tasks in under two days, GPU kernel improvements discovered without CUDA experts, his critique of constitutional AI and intelligence-level regulation, reward hacking risks, physical/economic constraints on hard takeoff, and skepticism about world models.
More from AGI Musings
- Two Years Ago Twitter Seriously Debated If LLMs Could Plan or Do Arithmetic — maksym_andr · 2026-09-15
- Rob Leclerc: Safety Orgs' High p(doom) Means They'll Sandbag and Veto Every Model Release — robleclerc · 2026-09-15
- SF Takes an Abstract Noble-Savage View of Enterprise Workflows, While New Yorkers Feel It Viscerally — willcb · 2026-09-15
- Ten Years Ago Musk, Hassabis, Bostrom and Others Warned About AI Risks on One Panel — iruletheworldmo · 2026-09-15
- Why Couldn't AI Be Conscious? Reddit Thread Challenges Intuition-Based Objections — Successful-Lie1603 · 2026-09-15
- Reddit: Without continuous learning, models can't be called AGI — Monochrome21 · 2026-09-15