On October 12, 2026, Anthropic co-founder Jack Clark moved up his forecast for when AI could develop its own successor models and repeat that improvement cycle, known as "recursive self-improvement" (RSI). He had previously pointed to 2028; he now points to 2027. One input to that judgment was TasteVal, a study that evaluates how well AI selects research experiments and interprets their results. The study found that AI's ability to run experiments efficiently within limited compute is improving rapidly, while the final performance it reaches is improving at a comparatively gentle pace. Running experiments efficiently is not the same as inventing new techniques and carrying out research and development autonomously from start to finish. To judge how plausible Clark's 2027 forecast is, we need to separate the progress that has been confirmed from the challenges that remain.
Why Anthropic's co-founder moved his recursive self-improvement forecast up by a year
In the newsletter Import AI 455, published May 4, 2026, Clark estimated a better than 60% chance that, by the end of 2028, AI would be advanced enough to develop successor models without human help. In Import AI 476, published October 12, he revised his view: full recursive self-improvement is now more likely to arrive in 2027 than in 2028. He did not give a new probability. This is his personal forecast, not an announcement by Anthropic of a product schedule or technical target.
Since May, Clark has distinguished between engineering progress, which incrementally improves existing techniques, and creative research, which produces new methods. Continually refining existing methods can raise AI capabilities on its own, but adding new inventions would speed up improvement further. For the October revision, he cited what appear to be instances of AI showing creativity in mathematics, along with results from the research-capability benchmark TasteVal. He did not name specific mathematical results.
TasteVal was developed by Oliver Jaffe and Dane Sherburn of P-Zero Research. Their preprint, arXiv:2610.06824v1, published October 5, 2026, tests whether AI can pick promising experiments for a research problem and interpret the results. The study frames this kind of judgment as a researcher's "taste." However, what is evaluated is the ability to decide what to try for a given problem, not the ability to discover research topics autonomously. That difference matters when linking more efficient experimentation to the automation of AI research as a whole.
Experiment efficiency doubles every three months, but final performance lags far behind
TasteVal compared 20 AI models released between 2023 and 2026 across eight AI research problems. Each model was assigned to choose the next experiment and interpret its results, while a common AI model, "Opus 4.8" (reasoning setting: high), handled writing the code and running the experiments. Using the same model for implementation means the evaluation measures differences in the judgment of the side setting research direction, not implementation skill.
Each trial was given one NVIDIA H100 GPU, with a cap of 40 GPU hours or 120 hours of wall-clock time. Twenty-four humans also took part, with at least two working on each problem. They received the same implementation support environment and compute budget as the AI models, and the best-performing human trial on each problem served as the benchmark. The AI was therefore compared not with the average of the 24 participants but with the best human result selected for each problem.
Sections 2.5 and 3.1 and Table 2 of the paper evaluate progress in research capability with two metrics: compute efficiency and final performance. Compute efficiency compares how much experimental compute is needed to reach the same result, and for models released since December 2025 it showed a trend of doubling roughly every 3.0 months. Final performance, measured when the budget is exhausted, doubled roughly every 14.6 months across 2023 to 2026, with no statistically significant inflection point toward acceleration. The two metrics differ not only in what they measure but also in the release periods of the models used to estimate their doubling times.
| Metric | What it evaluates | Opus 5.5 max result [95% CI] | Estimated doubling time [95% CI] |
|---|---|---|---|
| Compute multiplier | Experimental compute needed to reach the same score, compared with the human expert baseline | 2.30x [1.15–4.37] | 3.0 months [1.7–5.0 months] since December 2025; 14.0 months before that |
| Performance multiplier | Final score normalized so that a naive baseline is 0 and the human expert baseline is 1 | 1.14x [1.01–1.31] | 14.6 months [8.4–29.8 months] for 2023–2026; no significant inflection point toward acceleration |
To simplify the compute multiplier: if a human reaches a certain score using 40 GPU hours and an AI reaches it in 20 GPU hours, the multiplier is 2x. In the actual evaluation, compute is compared using the lower of the two final scores as the reference, and comparisons between model generations are chained together. Because results are also aggregated with a geometric mean across random seeds and research problems, the 2.30x figure in the table cannot be read as a 2.30-fold improvement in GPU performance or R&D cost.
Opus 5.5 and Fable 5.1 also each declined to work on one of the eight problems. The paper treats this as missing data rather than failure. When the results are recalculated on the seven problems common to all models and humans, Opus 5.5 max's compute multiplier is 2.25x, with a 95% confidence interval of 1.16 to 4.29. The overall trend is largely unchanged, but the 2.30x value is not the result of completing all eight problems.
Most important, the 3.0-month doubling time was estimated by comparing AI models released at different times. A single AI model did not run while improving itself and double its experimental efficiency three months later. What was observed is that, across generations of models, the compute efficiency needed to reach the same research result is improving rapidly. That progress cannot be treated as a demonstration of recursive self-improvement.
The study also found large differences in experiment costs. The average cost per trial was $282 for Opus 5.5 and $9,123 for humans. The calculation includes API fees for the AI used in each role, $2.50 per hour for GPU use, and $119 per hour for human labor. This is a comparison under specific conditions that include labor costs, and it does not show that the overall development costs of an AI lab could fall by the same proportion.
No "inventions" found among 540 research results
Another key question for the autonomy of AI research is the ability to create new techniques. The TasteVal authors analyzed 30 formal submissions from each of 18 models, 540 in total. They did not find a single one classified as an "invention," meaning one containing an original new structural element.
The classification used "Fable 5.0," which can access literature search tools, with up to 100 processing turns. Research results were sorted into four categories: "tuning," "combination," "modification," and "invention." Work was judged an invention if it contained structural elements not found in existing literature. Because both the literature survey and the classification judgment were performed by AI, the results should be interpreted within the limits of this evaluation method.
Notably, none of the formal submissions by human researchers were classified as inventions either. The authors also examined about 1,600 experiments proposed in the top 10% of trials, including failed or rejected ones, and found none that qualified as an invention. This result therefore does not show that only AI lacks the ability to invent. Under these research problems and this classification method, neither humans nor AI were found to have achieved results through the invention of new structural elements.
This matters for assessing the recursive self-improvement Clark predicts. If the ability to tune existing techniques, combine multiple methods, and refine them improves, that alone may speed up research and development. But such progress cannot be treated as evidence that invention has become possible. Conversely, it would not be appropriate to generalize from the absence of inventions in this evaluation to a claim that AI has no creativity.
The research overview published by the TasteVal authors also states that the humans in the comparison did not include world-class researchers. The research problems were fixed in advance, and the experiments were limited to small-scale ones using a single H100 GPU. Progress confirmed in an environment where results come relatively quickly and vary little will not necessarily appear in long-running research or in training large AI models. In addition, the evaluation problems are not public, to prevent contamination of training data, so the published paper alone does not allow a complete replication. Given these constraints, TasteVal's results alone cannot support a conclusion that AI has surpassed the world's best researchers.
What AI needs to develop its own successor models autonomously
The automation of AI research and development is already advancing in real research settings. In a survey Anthropic published on its internal R&D as of August 2026, 26% of work was "AI-led," meaning humans give broad instructions, AI carries out most of the work, and humans supervise. Work at the stage of AI-human collaboration or beyond exceeded 90%. At the same time, fully autonomous operation was not confirmed in any of the R&D areas surveyed.
The 26% figure, however, is neither a share of researchers nor a share of working hours done by AI. The survey assesses AI involvement by classifying weighted tasks. Anthropic also used its own AI models in the assessment, and no common measurement method has been established across the industry. Even so, distinguishing work in which humans direct detailed implementation from work in which they give only broad direction and let the AI proceed is a useful reference for considering how much of R&D can now be entrusted to AI.
What TasteVal showed is improvement in the ability to choose the next experiment for a given research problem and interpret its results. This ability is an important element in AI taking the lead in more R&D work. Yet a large gap remains between being able to select and interpret experiments and being able to discover research topics, make long-term plans, run large-scale training, and complete a successor model. In particular, how much human oversight is needed must still be tested, including who judges whether research results are valid and who changes course when experiments stall.
To judge whether the 2027 recursive self-improvement Clark predicts will materialize, better experiment efficiency alone is not enough. It will matter whether AI can select research topics autonomously, reproduce results from small-scale experiments in large-scale model training, and develop improved successor models while reducing human supervision and intervention.
TasteVal captured rapid progress in how efficiently AI can run research experiments, but it did not demonstrate new inventions or the autonomy of R&D as a whole. The focus going forward will be not only how well AI performs in individual experiments but also how far it can replace human judgment, from drafting a research plan to completing a successor model.