AI researchers debate how close we are to recursive self-improvement
The item is a new podcast episode from Dwarkesh Patel featuring John Schulman, Beren Millidge, and Charlie O’Neill. Patel says he wanted to hear details of what is actually happening at the frontier and what comes next, and he notes the guests are at somewhat open-ish labs so they can speak on record. The guests are Beren Millidge, CTO of Zyphra, which develops open source models; John Schulman, chief scientist at Thinking Machines and a former OpenAI co-founder who led the RLHF work that led to ChatGPT; and Charlie O’Neill, head of model training at Baseten.
The episode opens with a question about a 2036 scenario in which billions of crazy superintelligences have not radically transformed the world. Its timestamped topics include steelmanning the case against RSI, what is driving Chinese labs’ progress, how automated AI researchers will be trained, whether long-horizon RL will elicit AGI, the sim-to-real gap, how much progress is explained by data, why RL is working so well, Move 37 and entropy collapse, and rapid-fire timelines. The episode is available on YouTube, Apple Podcasts, or Spotify.
The show notes also include sponsor and announcement segments. Antithesis is presented as a way to test and verify code as agents generate more software, with Ron Minsky of Jane Street saying it helped shake out bugs in software that had already undergone heavy review. Grok Bot is described as a task-handoff tool used as a producer: it opens an interview transcript on its own computer, matches notes to exact moments, uses a preferences file to suggest edits, and sends top clip candidates for phone review.
Jane Street’s competition asks entrants to design a protocol-emulator ASIC: a general-purpose, reprogrammable design that can connect to a chip and simulate realistic traffic across multiple protocols and remain useful as new ones emerge. The most novel submissions will be taped out, winners will receive a physical copy, and the competition is open until January 18, 2027, with teams encouraged. The discussion matters because it brings notable researchers together to debate how close recursive self-improvement is, what bottlenecks remain, and how automated AI research and Chinese lab progress fit into the timeline.