Jeff Dean had just described about 20 automated improvement loops already running at his new company, Discovery Loop, when Fei-Fei Li asked him about the physical limits of recursive self-improvement. His answer was almost mundane. Someone still has to wait for the experiment to finish at 2 a.m., he said, and then wait until 8 a.m. to look at the results. Training, compute, saving data, and evaluation all still take wall-clock time. RSI, per his remarks, cannot be infinitely fast.
Last week this newsletter covered agent liability and open-weight economics. This week the thread is different: the RSI rate debate, and the gap between what the engineering evidence supports and what the market is paying for.
Dean's framing is not a hedge. It is a description of what his own company actually does. Discovery Loop has built an agent harness designed to scale to tens or hundreds of thousands of agents, not hundreds of people. Its stated target is a 20-to-50x productivity gain for domain experts. The company ran 400 interviews in three weeks, many at Cafe Venetia on University Avenue, to assemble a team where about 10% of employees have won an ACM computing prize and 10% have IOI medals.
That is a serious bet on automation. It is also a bet on compression, not explosion.
That is a serious bet on automation. It is also a bet on compression, not explosion. Dean described the goal as shortening each iteration of the science-and-engineering loop — propose, run, observe, decide — and letting systems learn from prior experiments so the next one is better. He traced the lineage back through neural architecture search, which he called an early form of recursive self-improvement: let the system find the architecture. What changed is that you can now specify the search in natural language to a capable model, rather than in brittle Python.
Then he drew the distinction that most public RSI discourse blurs. Self-improvement, as he defined it, is giving a system a measurable target — say, MRI interpretation accuracy — and letting it try different approaches. Recursive self-improvement is applying that same method to the system that proposes the improvements. The first is already routine. The second is the thing everyone is pricing.
The most useful recent evidence on how the second one actually pays off comes from a Salesforce AI Research preprint, arXiv 2611.08453, posted November 5. The paper compares on-policy distillation against GRPO, the reinforcement-learning method that rewards verifiably correct answers, on a single scoreboard: held-out accuracy per unit of training compute.
The headline result is about method, not magic. Train separate GRPO experts per domain, then distill them into one student. On a math-plus-code mix, that matched joint GRPO training with 42% fewer total FLOPs. Two experts cost 10.7 EFLOP; the merge added 0.55 EFLOP, about 5.2%. At four domains, distillation's share rose to 8.0% and total savings came to 31.6%.
The merge step stays cheap as the expert lineup grows. That is the scalable part. The expensive part is training the specialists, which is exactly the work you cannot skip.
The paper's counterexample is the more instructive half. When distillation used an external, post-trained teacher rather than the student's own RL checkpoint, it generally lost to plain RL on accuracy per FLOP. On LiveCodeBench, GRPO hit 46.4% at 0.6 EFLOP. Distillation from an external 32B teacher peaked at 28.7% after 68.7 EFLOP — roughly 114 times the compute for 17.7 points lower. External teachers also inflated response length.
Where distillation did pay was when the teacher was the student's own earlier checkpoint. On math, OPD recovered that checkpoint's accuracy with up to 50x fewer distillation FLOPs than GRPO needed to reach the same accuracy. The authors are explicit that on a single domain, distilling adds to total compute, because the RL checkpoint is already the desired model. The saving comes only from consolidation across domains.
Read together, these are not results about intelligence exploding. They are results about the shape of the research loop. Specialists, then merge. Supervised agent loops that run experiments and learn from the results. Compute-per-accuracy as the metric that matters, not headline capability.
Now put that next to the measurement. METR spent months inside Anthropic, looking at internal metrics and interviewing employees, and concluded that the pace of capability improvement on well-specified tasks was likely around twice as fast in October 2026 as it was between March 2024 and October 2025. METR believes a majority of that acceleration is likely driven by AI being used in AI R&D. That is a real, dated, quantified change in slope.
It is also not full RSI. METR states plainly that AI is not yet able to make highly significant and fully autonomous contributions to the most important parts of the model training process. Anthropic employees described "signs of life" suggesting such contributions could occur soon. And Anthropic redacted a majority of the evidence METR found most informative, citing commercial sensitivity and the risk of handing out a roadmap — including material suggesting the rate of growth could be undergoing continuous, hyperbolic acceleration.
So the most consequential public measurement of RSI-adjacent progress is, by the measurer's own account, mostly withheld. That is the pricing gap in miniature. Investors are underwriting a capability curve whose slope its own builders dispute, using evidence the builders declined to publish.
Anthropic's revised S-1/A, filed November 2, sharpened the ambiguity rather than resolving it. The filing says internal estimates and expectations regarding future models, not current ones, indicate the company may imminently meet its Responsible Scaling Policy's definition of "dramatically accelerated progress attributable to AI." It also says no current model has crossed that threshold, and commits to keeping humans in the R&D process. Whether that language is disclosure or marketing is a capital-markets question, not an engineering one. It is a follow-on to the open thread from 11/02, and it belongs to a different desk.
What the engineering desk can say is narrower and more useful. The recursion premium, as currently priced, assumes an unbounded version of a mechanism whose published behavior is bounded and specific. The leverage is in per-domain experts merged into a single student at a fraction of joint-training cost. It is in self-checkpoint teachers recovering accuracy at a fiftieth of the distillation compute. It is in roughly 20 supervised loops running at a startup that still has to wait for the data to save.
None of that is a small thing. A 42% FLOP reduction on a matched result is the kind of number that changes training budgets. A doubled rate of improvement on well-specified tasks, majority AI-driven, is the kind of number that changes timelines. Dean's 20-to-50x expert productivity target, if even partially met, is the kind of number that changes org charts.
But every one of those numbers is a compression of a loop that still contains a human, a GPU, and a clock. The market is pricing the version where the loop closes on itself. The people building the loops are publishing the version where it does not — yet.