Anthropic's Jack Clark moves self-improving AI forecast to 2027 on TasteVal results

Anthropic co-founder Jack Clark now believes recursive self-improvement (RSI), where AI systems develop their own successors, could become possible as early as 2027. He had previously expected 2028. Clark, who writes the Import AI newsletter, bases the revision on TasteVal, a new benchmark from P-Zero Research. According to Clark, the capability it measures has roughly doubled every three months since December 2025.

What TasteVal measures

P-Zero Research calls the capability "research taste": picking the right experiments and drawing the right conclusions from them. It defines the skill as three abilities: identifying worthwhile research questions, designing experiments to answer them, and interpreting results. TasteVal tests only the last two. The research goal is fixed in advance.

The benchmark has eight tasks drawn from AI model development, each with a fixed goal and a measurable score. They cover training data selection, pretraining and fine-tuning language models, training on human preferences, and defending against adversarial prompts. The model chooses which experiment to run next and uses the results to guide its following step.

To keep coding skill from skewing results, a fixed coding agent runs all experiments on a single Nvidia H100 GPU. Each run ends after 40 GPU-hours of compute or 120 hours of wall-clock time.

The human baseline comes from 24 experts who had recently worked at places such as OpenAI, Google DeepMind, Nvidia and Microsoft Research. At least two people worked on each task, and the best human attempt per task served as the benchmark.

Results

P-Zero Research converts research taste into compute efficiency, which it calls the compute multiplier. A model that matches expert scores with half the compute has twice the research taste. The lab bases this on its "AI Futures Model," which predicts that experiment compute becomes the biggest bottleneck once coding is automated.

• Top model: Anthropic's Opus 5.5 reached a compute multiplier of 2.3, hitting expert-level scores with less than half the compute.
• Uncertainty: the statistical range runs from 1.15 to 4.37. At the low end, the model barely leads.
• Cost: a single run cost roughly one-thirtieth of an average expert attempt.
• Trend: the compute multiplier has roughly doubled every three months since December 2025, with a reported doubling time between 1.7 and 5.0 months. Growth was much slower in prior years and steepened around the turn of 2025 into 2026.

The authors say their results may overstate how quickly research taste is improving. The tasks are quick and cheap to verify, which is the kind of task AI labs find easiest to optimize for, and the human group included no top-tier researchers.

Clark's revised timeline

In a May essay, Clark put the odds of a system that independently trains a stronger successor at roughly 60 percent by the end of 2028 and 30 percent by 2027. Earlier this year he wrote that models had not shown creativity, which he has long called the key factor in self-improvement, at any large scale. He now points to progress in mathematical breakthroughs and benchmarks like TasteVal as signs that this is changing. He co-founded the company whose model tops the benchmark, which the source notes adds a layer of bias to the forecast.

AI researcher Herbie Bradley had pushed back on the May essay. He argued that models could plausibly take over junior researchers' work, but that research taste, creativity and coherent long-term research planning were nowhere in sight.

A study led by a team at Fudan University, which analyzed the development of the team's own model with AI agents, found that agents contributed up to 55 percent of methodology proposals. Humans still made over 85 percent of final decisions.

In June, Anthropic wrote that RSI has not been achieved yet but "could come sooner than most institutions are prepared for." CEO Dario Amodei now calls for a controlled slowdown in AI development. OpenAI acknowledges that fully autonomous RSI is not happening yet and says it should be pursued only if human oversight is guaranteed.

What this means

TasteVal does not show that AI systems are already improving themselves. It covers a narrow slice of scientific work, and it leaves out the choice of research direction, which is often considered the core of scientific intuition. It does test experiment design and interpretation, the area Bradley had flagged as out of reach.

The main uncertainty is statistical: the 2.3 multiplier and the three-month doubling time both carry wide ranges, and the authors flag possible overstatement. If the trend holds in other tests, the source notes that AI development could become faster and cheaper, while a growing share of research would be harder for humans to verify because models would be planning and evaluating their own experiments.