Self-improving AI in 2027: Jack Clark’s forecast and the limits of TasteVal

AI already helps people write code and analyze experiments. But what happens if it can conduct research to create the next, more capable version of itself? Anthropic co-founder Jack Clark believes this may become possible sooner than he expected.

Recursive self-improvement in plain language

Imagine a cycle: researchers ask AI to improve how models are trained. The system proposes hypotheses, conducts experiments and helps create a new model. That model, in turn, can research more effectively and help create the next one. This repeated process is called recursive self-improvement, or RSI.

This is a broader idea than simply “AI wrote code for AI.” A full cycle requires choosing a useful problem, planning experiments, verifying results, training a new system and demonstrating that it is better. It also requires data, servers, energy, budgets and safety mechanisms. Performing one part of the chain does not mean the entire system works autonomously.

What Clark said — and did not say

In his October 12 issue of Import AI Clark explained that he had previously considered creativity a major constraint on RSI. He noted progress on mathematical tasks and TasteVal, a test of how models plan experiments and evaluate results. On that basis he changed his forecast: in his view full RSI could become possible in 2027, earlier than he had expected.

This is not a deadline or a new official Anthropic date. In the October text Clark did not give a new numerical probability. He co-founded the company whose model performed well on the test, so his expertise matters, but his potential interest should also be considered. The October 25 article in THE DECODER, which prompted our investigation, covers the October forecast.

What TasteVal tested

P-Zero Research described the test in the public TasteVal preprint. The researchers describe research taste as the ability to select important problems, plan experiments and draw conclusions. But in this test people selected the problem in advance: the model was evaluated on the next two steps, experiment planning and interpretation of results.

There were eight AI research tasks, such as data selection, model training and fine-tuning, and resistance to malicious prompts. A researcher model proposed experiments, and a separate fixed agent implemented and ran them. The human baseline used results from 24 experts, at least two per task; the best human attempt was selected for each task.

2.3 times and three months: reading the numbers

The authors estimated that Opus 5.5 achieved a compute efficiency factor of 2.30 relative to the human baseline. A simple hypothetical example: if a person achieves a result using 23 GPU hours, a factor of 2.3 would mean reaching the same level in about 10 hours. This illustrates the metric; it is not a specific task from the study.

The result has uncertainty: the authors reported a 95% confidence interval of 1,15–4,37. They also estimated that an average model attempt cost about 30 times less than an average expert attempt under their methodology. The comparison included estimated human labor, API and compute costs under test conditions; it cannot be generalized to salaries or the cost of an entire research team.

The authors estimated that this particular compute-efficiency metric for leading models had doubled roughly every three months since December 2025. The estimate’s confidence interval was 1.7 to 5 months. Another metric, final result quality at a given budget, showed a different trend. Saying “AI intelligence doubles every quarter” would therefore be incorrect.

Why this is not yet proof of autonomous AI creation

• People chose the problem. The test does not measure whether a model can identify the most important research direction itself.
• There were eight tasks. They are useful for comparison but do not cover the full complexity of years of laboratory work.
• Conditions differ from a real laboratory. The authors note that the tasks are relatively easy to verify and run on one H100; results could either overstate or understate progress.
• The findings come from one preprint. The tasks themselves are not public to avoid test contamination, limiting independent replication. The trend does not guarantee future continuation.

A useful comparison from a team’s real work: in the Atria Dawn model-development study, humans retained most final decisions about methods and parameters even when AI proposed options. This is a separate study of one project, not universal statistics for all AI labs. Atria Dawn study ↗

What this means for working with AI now

For a developer, designer or animator, the practical lesson is not to wait for an exact RSI date, but to build a verifiable workflow. A modern agent can suggest options, but a confident answer does not guarantee correct code, rights to materials or a quality final result.

1. Set a measurable task. For example: “Improve scene loading time without changing its appearance.”
2. Ask for several hypotheses and ways to test them. For each, specify expected outcome, risks, time and success criteria.
3. Test on a copy of the project. Compare with the original under the same conditions; do not give the agent unnecessary access.
4. Leave the final decision to a person. Check functionality, copyright, visual quality and the ability to roll back changes.

This is an editorial example for independent testing, not a TasteVal test or guarantee of results. See also our guide to checking AI agents’ work.

Risks researchers discuss

If future systems accelerate research, people may find it harder to check every experiment and the consequences of changes. Access to compute, independent evaluation and control of models will become more important. The test itself does not prove a loss of control or an inevitable intelligence explosion. Scenarios depend on assumptions that still need testing.

Sources and methodology

• Jack Clark, Import AI 476, October 12, 2026 — his October assessment and context.
• P-Zero Research: TasteVal — preprint methodology, results and limitations.
• P-Zero Research: test description — metric definition and task design.
• Atria Dawn: study of agent use in model development — a separate example of how decisions were distributed.
• OpenAI, October 26, 2026 — primary account of three AI-generated results, materials and external review process.
• The Verge, October 27, 2026 — journalistic context for examining the openness of verification.
• THE DECODER, October 25, 2026 — the starting point for our investigation.

We checked the forecast against Clark’s text and the figures against the authors’ study. We conducted no experiments of our own and have no independent confirmation of the forecast. The tool above is our editorial heuristic, not a scientific reliability scale. The Mia cover is a generated editorial illustration, not a photograph of an event.

Discuss the news and your projects in the Ukrainian Vibe Coding community on Telegram.