Summary
Richard Socher — co-founder of Recursive and former Salesforce Chief Scientist — joins the Latent Space podcast to discuss his forthcoming book and the company he built around it. The central concept is the “Eureka Machine”: a superintelligence designed to pursue open-ended scientific discovery, capable of being given any goal or reward environment and inventing solutions that humanity would ask for. Socher argues this is the most consequential technology humanity could build, with transformative upside for physics, chemistry, biology, and economics.
The conversation covers where AI optimists go wrong (blind faith without clear-eyed downside analysis), Socher’s view that regulating AI at the model level is analogous to regulating intelligence itself — and why application-level regulation is the right frame instead. He draws parallels to internet governance debates and argues that constraints on GPU compute would produce authoritarian outcomes worse than the harms they target.
Deeper in the episode, Socher and the hosts dig into metacognition as an underexplored dimension of machine intelligence, the economic incentives that currently prevent models from developing autonomous goal structures, and the inadequacy of existing intelligence benchmarks — IQ, Elo ratings — for measuring systems that may surpass human-level performance. Socher shares early thinking on a new unit of intelligence he’s developing, calling current definitions “entropic” in that they cap potential by defining intelligence relative to humans. The episode is a substantive long-form discussion for anyone tracking AGI timelines and the emerging superintelligence startup landscape.
📺 Source: Latent Space · Published September 14, 2026
🏷️ Format: Podcast
