AI researchers debate how close we are to recursive self-improvement

Summary — In this panel discussion hosted by Dwarkesh Patel, AI researchers John Schulman (Chief Scientist at Thinking Machines, co-founder of OpenAI), Beren Millidge (CTO of Zaphra), and Charlie O’Neil (Head of Model Training at Baseten) explore the technical hurdles and timelines surrounding recursive self-improvement (RSI) and automated AI R&D. They debate reinforcement learning sample efficiency, synthetic environments, distillation versus frontier training, and timeline estimates for human-level automated remote workers and superintelligence.

What is shown

• [00:13] Introduction of panelists Beren Millidge, John Schulman, and Charlie O’Neil.
• [06:12] Graphic overlay comparing chess Elo and human win rates over time against AI benchmarks.
• [17:24] Sponsor feature highlighting Antithesis software verification and an interview with Yaron Minsky (Co-Head of Technology at Jane Street).
• [40:02] Sponsor segment outlining Jane Street’s competition to design a protocol emulator ASIC on an open-source 130nm process node.
• [59:32] Sponsor segment demonstrating Grokbot for automated transcript editing and video clip suggestions in Slack.
• [62:26] Discussion referencing Patel’s blog post and graphs dissecting pre-training efficiency gains from data versus architectural improvements.
• [78:14] Visual references to research posts by the panelists, including John Schulman’s “LoRA Without Regret”, Dwarkesh Patel’s RL efficiency post, and Beren Millidge’s blog on RL information efficiency.

Claims & numbers

• Charlie O’Neil suggests that once an agent is even 0.1% better than all humans at AI research, the ability to run hundreds of thousands or millions of instances in parallel could drive rapid takeoff via recursive self-improvement [31:12].
• Dwarkesh Patel notes empirical data from 2019 to 2025 showing data improvements accounted for roughly 9x–12x compute efficiency gains, while model/architecture improvements accounted for approximately 3x–3.7x gains [65:56].
• Charlie O’Neil predicts a drop-in remote worker capable of a month of complex autonomous tasks within roughly 1 year (or 2–3 years if full arbitrary web/browser interaction is mandated) [92:28]. Beren Millidge estimates 3 years for full generality [93:32].
• For a 10x productivity uplift for AI researchers, John Schulman and Beren Millidge both estimate around 2 years [93:02, 93:06], whereas Charlie O’Neil estimates 5 to 10 years [92:31].
• For artificial superintelligence dominating top human experts across all cognitive domains, John Schulman predicts 3 to 4 years [94:45], Charlie O’Neil estimates 5 to 10 years [95:58], and Beren Millidge estimates around 5 years [96:25].
• Jane Street’s ASIC emulator contest accepts submissions until January 18, 2027, for tape-out on a 130nm node [41:11].

Notable quotes

• [22:20] Charlie O’Neil: “All thinking can do is like update your posterior basically based on like, you know, the bits that you’ve gotten since you formed your prior. Like, you can’t gain any new bits from like just thinking.”
• [37:41] John Schulman: “I think a lot of what people call taste is something about behavior that works in the long run and that people have realized works in the long run.”
• [94:45] John Schulman: “I would say like three or four years.” (on the arrival of AI dominating top human experts across every cognitive field)

Assessment This is a long-form roundtable technical interview and debate format featuring active AI researchers, interspersed with clearly designated sponsor segments. The discussion represents expert perspectives, qualitative hypotheses, and theoretical debate rather than a single company product demo.

Described by gemini-3.8-flash on 2026-10-18 from the video’s audio and frames.