OpenAI’s chief research officer, Mark Chen, said AI is “maybe 80% of the way” along what he called the AGI spectrum, but “somewhere between, let’s say, 25% and 50% there” on recursive self-improvement (RSI), in the first episode of a16z’s Working Theory series, published on Tuesday, 17 November. The same day, OpenAI said GPT-6.2 Astra was the first model it has released at its High rating for AI self-improvement.
Our number: Chen’s RSI range sits 30 to 55 points below his own AGI estimate, by our arithmetic. Two measured results in OpenAI’s system card land above the top of his range. GPT-6.2 Astra reached 59% of the best human result on NanoGPT. It scored 29.01% on Vals AI’s modified RSI Index, where 50% marks human level, or 58% of that mark, by our arithmetic. These tests measure different things from his estimate.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on why products built on the same model differ. It teaches the general idea only and does not discuss OpenAI, Thinking Machines or the episode.
The key insight: OpenAI’s research chief put recursive self-improvement 30 to 55 points behind his own AGI estimate, by our arithmetic, while the system card he works under shows research tests at 58% to 59% of human level on two measures that are not like-for-like with his range.
What Chen Said
Chen, speaking on a video call, said: “I do believe that we’ve built AI research interns today”. He said OpenAI measures that by “the amount of compute that we are using through Codex managed by Codex end-to-end” to run experiments or automate parts of the research pipeline.
He defined his RSI range as “how effective are the models at feeding back and discovering ideas”. On measuring progress, he said pre-training has one long-standing yardstick, perplexity, while evaluations further down the pipeline have “shorter and shorter shelf life”, so much of OpenAI’s work there is creating and automating harder benchmarks.

What Schulman Said
John Schulman, a co-founder of OpenAI and now co-founder and chief scientist of Thinking Machines Lab, described RSI as “just using AI to assist with this research process and speed it up or make it work better”.
He described his own workflow with Anthropic’s Claude Fable: “I’ll go and write a complicated prompt to Fable to go and work on this idea”. After a few days, he said, “the end result won’t be that good”. When he steers the work himself, he said, he will “ask a lot of questions that the AI wouldn’t have thought to ask”.
On research intuition, he said models can read the literature, but “memorization can be a blessing and a curse”. On alignment, he said: “I’d rather have this be like a public science that a lot of people understand well”.
Alignment, in Their Words
Chen said the most effective tool for aligning models “is chain of thought monitorability today”, and that the future of alignment “has to be stronger models aligning other models”. Schulman said training objectives could push models to “report on their internal states that they otherwise didn’t verbalize”, and named OpenAI’s Confessions research.
What OpenAI’s Own Data Shows
OpenAI’s GPT-6.2 system card, as we reported earlier on Tuesday, says GPT-6.2 Astra saved 42.87% of baseline training time on NanoGPT against 72.38% for the best human solution, below an 80% indicative cutoff, and cleared one of three second-stage cutoffs, by our count.
The card says Vals AI “concluded that GPT-6.2 Astra was unlikely to fully automate AI research, but could provide a larger productivity boost to researchers than previous models”. Epoch AI scores the model at 174, as we also reported.
The Business Lever
The lever is research throughput per unit of compute. Chen said OpenAI counts its research interns by the compute Codex runs end to end, so the measure he gave rises with the compute budget. Sarah Wang, the a16z general partner who hosted, wrote that “research taste” remains “harder to teach”.
The Structural Read
As we read the episode, both researchers describe RSI today as AI assisting human researchers, not replacing them: Chen’s research interns take scoped tasks, and Schulman said his own long AI-run projects went further once he steered them.
As we read the figures, Chen’s range is a spoken estimate with no stated method, while the NanoGPT and Vals numbers are single tests, so neither confirms the other.
John Schulman, Thinking Machines Lab, a16z Working Theory, 17 November 2026
“we might see lots of minor things going wrong, but we would learn something new each time”
Three Implications
FOR OPENAI Chen said OpenAI counts its research interns by the compute Codex runs end to end, and rated GPT-6.2 Astra High for AI self-improvement the same day, from our earlier reporting.
FOR THINKING MACHINES Schulman said he would rather alignment be a public science that a lot of people understand well, and said people should have more control over the values of the AIs they use.
FOR RESEARCH TEAMS Schulman described long AI-run projects that fell short until he steered them himself, asking questions the AI would not have thought to ask.
The Business Engineer Lens
This story maps onto the Business Engineer framework The Agentic Harness War.
The framework says: “The surface, not the model, governs autonomy.”
As we read it, Schulman’s account of the same model going further once he steered it fits the framework’s point that the surface around a model sets how much it can do alone. The framework predates this episode and does not discuss RSI, Chen or Schulman.
What Is Not Established
Chen gave his percentages as rough estimates and did not describe a method. The NanoGPT and Vals figures come from OpenAI’s system card and measure specific tests, not the spectrum Chen described, so the comparison is not like-for-like.
The quotations are from our machine transcription of the episode audio, checked against the video; a16z produced and edited the episode. We matched speakers to the picture. We did not contact OpenAI, Thinking Machines or a16z.
The Bottom Line
OpenAI’s research chief put AI at 25% to 50% of the way to self-improvement, 30 to 55 points behind his own AGI estimate, by our arithmetic, on the day OpenAI gave GPT-6.2 Astra its first High self-improvement rating. Schulman said his own long AI-run projects still need his steering to get far.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Quotations are from our machine transcription of a16z’s Working Theory episode 1, published 17 November 2026, checked against the video; speakers were matched to the picture. Titles are from a16z’s post. NanoGPT, Vals AI and Epoch figures are from OpenAI’s GPT-6.2 system card and release post, as in our earlier coverage. The 30-to-55-point gap and the 59% and 58% ratios are our arithmetic; the measured tests are not like-for-like with Chen’s estimates. We did not contact OpenAI, Thinking Machines or a16z.
Sources: a16z, Working Theory ep. 1: How Close Is Self-Improving AI? (YouTube), 17 Nov 2026 · @a16z on X, 17:44 UTC, 17 Nov 2026 · @sarahdingwang on X, 17:52 UTC, 17 Nov 2026









