Jack Clark predicts AI could start improving itself as early as 2027

What is recursive self-improvement?

Recursive self-improvement occurs when AI systems create their own next versions, accelerating development with minimal human intervention.

The TasteVal test and its results

A new benchmark called TasteVal recently tested AI models’ ability to plan experiments and interpret results across eight AI development challenges. Each model was given a predefined problem, suggested experiments, and a uniform programming agent executed these suggestions, returning the results for analysis.

In the test, the Opus 5.5 model matched human benchmark results using 2.3 times less computation. This efficiency led the authors to calculate that frontier model performance has doubled every three months since December 2025, compared to the previous 14-month doubling cycle.

Researchers compared 20 models with the best attempts of 24 experts from companies and research centers like OpenAI, Google DeepMind, Nvidia, and Microsoft Research. Additionally, the average cost per attempt for Opus 5.5 was about one-thirtieth of what was spent on human experts.

Why the forecast changed to 2027

Clark attributes the revision of his estimate to recent advances in mathematics and evidence that models can plan and evaluate experiments more efficiently. Based on these signs, he now believes 2027 is the most likely year for a system to train its own successor, moving up his previous May projection by one year.

What challenges remain

TasteVal only assessed part of the research process. The models didn’t need to choose which scientific problem to investigate, train a new AI, or decide how to modify their own architecture—key steps for a complete self-improvement cycle.

The authors highlight additional limitations: the tasks were designed to produce quick, easily verifiable results, and the expert group used for comparison may not include the most experienced researchers. Furthermore, while computational efficiency has grown every three months, final performance improvements still follow a doubling rate estimated at 14.6 months.

In summary, while efficiency gains are impressive, they don’t equate to full autonomous research capability. For AI to improve itself, further advances are needed to connect experimental planning with the independent creation and modification of new models.

Future outlook

Recent progress suggests AI is nearing the ability to conduct larger portions of its own development. However, achieving full recursive self-improvement still requires solving technical, methodological, and governance challenges that go beyond the computational efficiency demonstrated in TasteVal.