Jeff Dean and Fei-Fei Li's latest conversation: all about RSI non-consensus

What happened: In a Stanford HAI conversation with Fei-Fei Li, Jeff Dean discussed limits on recursive self-improvement (RSI) and described his new company Discovery Loop, which automates experiment loops for machine learning, science, and engineering. Key numbers: Discovery Loop Why it matters: Dean argued RSI cannot run infinitely fast because training, compute, data storage, and evaluation still take time. He said he underestimated how deeply coding ability would affect models' problem solving, and he values humans being able to extract reusable methods from machine generated proofs.

In a Stanford HAI conversation with Fei-Fei Li, Jeff Dean discussed limits on recursive self-improvement (RSI) and described his new company Discovery Loop, which automates experiment loops for machine learning, science, and engineering.

Dean argued RSI cannot run infinitely fast because training, compute, data storage, and evaluation still take time. He said he underestimated how deeply coding ability would affect models' problem solving, and he values humans being able to extract reusable methods from machine generated proofs.