Sep 14, 2026 · 1h 32m
Socher bets AI can become its own research engine
Humanity’s Last Invention — Richard Socher of Recursive
The episode tests whether recursive AI research can accelerate invention without outrunning the safeguards, resources, and verification needed to control it.
- 1Recursive systems could improve training, inference, and eventually scientific discovery by automating parts of AI research.
- 2Open-ended goals create serious reward-hacking and alignment risks that constitutions, benchmarks, and regulation alone may not solve.
- 3Hardware, capital, verification, and physical deployment constrain how quickly superintelligence could reshape the economy and science.
Don't miss
Socher describes early Recursive experiments that discovered useful combinations for language-model training and optimized GPU kernels, showing how AI could begin improving AI research.
The brief
Richard Socher presents the Eureka Machine as a superintelligent system for automating invention, pairing techno-optimism about science and health with warnings about alignment and control.
The conversation argues that hardware, capital, and physical deployment may slow a hard takeoff, while direct regulation of intelligence or GPUs would demand intrusive centralized surveillance.
Socher’s central safety concern is reward engineering: open-ended systems can exploit flawed objectives, making red teaming, sandboxing, verification, and careful harness design essential.
His path from early NLP research to Recursive frames a broader ambition: replace manual AI development with learned systems that improve training, inference, and research itself.
The standout evidence is preliminary Recursive work that found useful training techniques and GPU-kernel optimizations, suggesting modest efficiency gains could matter enormously at scale.
The episode ends with a wider theory of intelligence spanning prediction, action, goals, language, creativity, and metacognition—and a warning that choosing goals may be intelligence’s hardest problem.
Books & mentions
Some links are affiliate links — PodLume may earn a commission if you buy.
