The Anthropic co-founder now says AI capable of training its own successor is more likely by 2027 than 2028 a full year earlier than his forecast just five months ago.
Who revised the forecast? Jack Clark, co-founder of Anthropic and author of the Import AI newsletter. What changed? Clark brought forward his timeline for recursive self-improvement (RSI), a stage in which AI systems can conduct the research needed to develop their own successors. When did he make the revision? On October 12, 2026, in issue 476 of Import AI, moving his projected timeline from late 2028 to 2027. What evidence informed the change? Clark pointed to a new benchmark called TasteVal and recent AI-assisted mathematical breakthroughs that he described as genuinely creative. Why does it matter? Recursive self-improvement could represent a significant milestone in AI development, with potentially far-reaching benefits and risks. How confident is he? Clark did not provide a revised numerical probability in this update. However, in May 2026, he estimated a 30% chance of RSI by 2027 and roughly 60% by the end of 2028.
In May 2026, Jack Clark put the odds of AI systems being able to train their own successors at 30% by 2027, rising to about 60% by the end of 2028. He called it, in his own words, a “Rubicon into an almost unpredictable future,” but treated 2028 as the more likely year.
That changed on October 12, 2026. In Import AI 476, Clark wrote a short but consequential line: he now thinks it’s more likely full recursive self-improvement becomes possible in 2027, rather than 2028. He did not attach a new numerical probability to the update, this is a directional shift in his thinking, not a formally revised forecast.
Recursive self-improvement describes a loop in which an AI system does the research work needed to train a stronger successor and that successor, in turn, is capable of doing the same for the generation after it. If the loop holds, capability gains compound across generations largely without new human research insight driving each step.
This is distinct from an AI model simply getting incrementally better with more training data or compute. RSI specifically means the AI itself is doing meaningful parts of the research identifying promising directions, designing experiments, and interpreting results the kind of work currently done by human AI researchers.
TasteVal: The Benchmark Behind the Shift
Clark’s revised forecast leans heavily on a new benchmark called TasteVal, introduced by a group called P-Zero Research. TasteVal measures what researchers call “research taste” a composite of three abilities: identifying which research questions are actually worth pursuing, designing experiments capable of answering them, and correctly interpreting the results those experiments produce.
Clark has long argued that creativity, not raw technical skill, is the real bottleneck standing between current AI systems and genuine self-improvement. TasteVal is notable to him because it’s one of the first benchmarks attempting to directly measure that creative judgment rather than narrower technical competence.
What the benchmark does and doesn’t show
- It captures a narrow slice of scientific research work, not full end-to-end AI development
- Anthropic’s own Claude Opus 5.5 reportedly outperformed human experts from rival labs on relevant tasks, while being markedly more cost-effective
- It does not prove AI systems are already improving themselves — it suggests the underlying skill required for that loop is advancing quickly
The Evidence Clark Is Pointing To
Beyond TasteVal, Clark’s broader case for pulling the timeline forward rests on a pattern of rapid gains across several existing benchmarks. On SWE-Bench, a test of real-world software engineering tasks pulled from GitHub repositories, the best models went from solving roughly 2% of cases at the end of 2023 to around 94% by spring 2026. On CORE-Bench, which tests whether an AI agent can faithfully reproduce the results of a scientific paper including environment setup, code execution, and analysis Clark has described the benchmark as effectively “closed,” with modern agents scoring around 95.5%.
He also points to a capability that’s roughly doubled every three months since December 2025, and to recent AI-generated mathematical results he characterizes as showing real creative insight rather than pattern-matching. Separately, independent forecasting group Vals AI has published its own RSI Index, which currently estimates full automation of machine learning engineering work at Anthropic could arrive as early as July to August 2027 — a dated, checkable prediction that roughly lines up with Clark’s revised timeline.
Why This Matters Beyond Anthropic
If Clark’s revised timeline holds, the practical implications extend well past Anthropic’s own roadmap. A shift from “self-improving AI by 2028” to “by 2027” compresses the window governments, safety researchers, and competing labs have to prepare policy, safety evaluations, and deployment guardrails around a technology most AI researchers already consider one of the highest-stakes thresholds in the field.
It also raises a harder question that Clark himself has flagged before: what happens to human oversight and control once AI systems start developing successors faster than human teams can meaningfully review. That tension between capability gains and the pace of human oversight is likely to shape how seriously labs, regulators, and the public treat 2027 as a milestone worth watching closely.
Key Talking Points
- Jack Clark moved his recursive self-improvement forecast from “more likely in 2028” to “more likely in 2027” on October 12, 2026.
- The shift is based partly on TasteVal, a new benchmark measuring AI research taste and creative scientific judgment.
- Supporting evidence includes rapid gains on SWE-Bench and CORE-Bench, plus a research-judgment capability Clark says has doubled roughly every three months since December 2025.
- This is Clark’s personal forecast published in Import AI, not an official Anthropic company position.
- Independent forecaster Vals AI separately estimates full ML engineering automation at Anthropic could arrive around July to August 2027.
- Clark’s position as an Anthropic co-founder, whose model tops the cited benchmark, is a relevant conflict of interest to weigh alongside the forecast.