Jack Clark Pulls RSI Timeline to 2027 as TasteVal Shows AI Research Skill Doubling Every 3 Months

Anthropic co-founder Jack Clark has revised his estimate for when AI systems will achieve full recursive self-improvement, shifting his most-likely year from 2028 to 2027. The revision, published in Import AI issue 476 on October 13, is driven by two converging data points: a new benchmark called TasteVal showing that frontier AI systems' research ability has been doubling every three months since late 2025, and a wave of AI-generated mathematical breakthroughs that Clark reads as evidence the creativity component — the capability he previously identified as a missing prerequisite for autonomous model development — is now closing faster than he expected.

The update is Clark's most significant timeline revision since he first publicly assigned 60% odds to RSI arriving by end-2028 earlier this year, and it aligns him with xAI's Elon Musk, who has said he expects full automation of the Grok model improvement loop by end-2027. It sharpens the contrast with OpenAI, which has publicly acknowledged it does not yet know how to achieve aligned, full RSI safely while targeting an autonomous AI research intern by March 2028.

What Recursive Self-Improvement Requires

The term "recursive self-improvement" describes a specific technical threshold: the moment an AI system can autonomously handle the work required to build a more capable successor. This includes generating research hypotheses, designing training runs, running evaluations, and interpreting results — all without requiring human researchers to direct each step.

The threshold matters because improvement compounds rather than accumulates. A system that can improve its successor by 5% can run that same cycle on the next version, with each gain applied to an already-improved base. Clark has described the mathematics as unforgiving: a 99.9%-reliable alignment technique applied across 50 recursive generations retains about 95% reliability by generation 50 and falls to roughly 60% by generation 500. That degradation curve is why RSI is treated as a safety-critical milestone and not merely a productivity development.

The technical preconditions are partially measurable. METR, an AI evaluation organization, tracks the "task-horizon length" — the duration of tasks AI agents can complete at 50% reliability without human intervention. That figure has grown from roughly 30 seconds in 2023 to approximately 12 hours in mid-2026. Researchers estimating RSI prerequisites generally peg "weeks of autonomous task completion" as the capability level that would allow an AI to run a full training cycle without a human checkpoint, leaving a meaningful gap that the current trajectory is closing.

Read more: Hinton and Bengio paper finds automated AI R&D could compress a year of progress to five weeks

TasteVal: The Benchmark Behind Clark's Revision

The most specific data Clark cited in Import AI 476 comes from TasteVal, published by P-Zero Research on October 6, 2026. The benchmark attempts to measure something researchers have long treated as difficult to quantify: research taste, defined as the ability to design productive experiments given a fixed problem statement.

TasteVal separates research judgment from coding ability using a two-agent architecture. A "Researcher" model — the system being evaluated — proposes experiments and interprets results. A fixed "Coder" model implements those proposals and runs them on a single H100 GPU, capped at 40 GPU-hours per run. By isolating experimental design from implementation, TasteVal aims to measure strategic judgment rather than programming proficiency.

Eight novel tasks cover core frontier AI R&D activities: pretraining data curation, hyperparameter optimization, preference modeling, and robustness to adversarial prompts. The human baseline consists of 24 experts who recently worked at OpenAI, Google DeepMind, NVIDIA, and Microsoft Research, with at least two experts per task.

The core metric is a compute multiplier — the ratio of compute a human expert needs to reach a given score, divided by the compute the AI needs. Claude Opus 5.5 scored a compute multiplier of 2.30 (95% CI: 1.15–4.37), meaning it achieved equivalent experimental outcomes on roughly half the compute of the human experts and at approximately one-thirtieth the per-attempt cost.

More significant than the snapshot score is the rate of change. P-Zero Research reports that experimental research taste among frontier models has doubled every 3.0 months since December 2025 (95% CI: 1.7–5.0 months). Before that inflection, the doubling time was approximately 14 months. Substituting this trend into the AI Futures Model forecasting framework raises the estimated probability of a "taste-only singularity" from 51% to 88%, and moves the median projected ASI arrival date from mid-2030 to late 2028, according to the paper — though the authors explicitly label these as "naive" extrapolations rather than formal forecasts.

Caveats the Authors and Reviewers Flag

The TasteVal authors identify limitations in both directions. The benchmark may overstate progress because its tasks are "fast and cheap to verify," a profile that frontier labs consistently find easiest to optimize for. The human baseline may understate expert capability because it excluded the field's top researchers. The benchmark also does not measure research vision — the ability to identify which problems are worth pursuing — only experimental execution given a problem already defined.

One structural finding in the data limits the creativity narrative directly: of all model-proposed experiments in the TasteVal evaluation, none were classified as "Invented" — no fundamentally novel experimental structures appeared. Every proposal was a "Tuned" or "Composed" variation on existing techniques. The mathematical breakthroughs Clark cites as his second data point — OpenAI announced hundreds of AI-generated mathematical results in early October, following a report that a model solved one of mathematics' most complex problems — remain contested on the same axis. Independent mathematicians are divided on whether the proofs represent genuinely new creative contributions or final-step assembly of work human researchers had substantially prepared.

Clark's Previous Estimate and What Changed

Clark's earlier RSI estimate — from Import AI issue 455 — assigned 60% odds to AI systems autonomously training successors by end-2028, and 30% odds to that threshold arriving by end-2027. He described it as a "reluctant view" built from public benchmark data, resting primarily on SWE-Bench improving from approximately 2% to 93.9% and METR's task-horizon measure growing from 30 seconds to 12 hours.

Import AI 476 effectively raises the probability weight on 2027 without abandoning 2028 as a plausible outer bound. If frontier models are genuinely doubling research effectiveness every three months, a four-quarter doubling sequence between October 2026 and October 2027 would produce a 16x improvement over current capability levels — a trajectory Clark appears to find sufficient to shift his modal estimate one year forward.

The Governance Window This Revision Compresses

The shift to 2027 as a central estimate is not an isolated forecast. Musk's xAI targets the same year for full Grok automation. A July 2026 open letter from AI researchers asked governments to build capacity to slow automated AI development if needed. The Institute for Security and Technology launched a dedicated RSI governance initiative in August 2026 after concluding there was "no consensus on what signals should trigger closer monitoring, coordinated action, or development pauses."

The Institute for Progress published 23 specific policy recommendations in August 2026, but even its proposed 72-hour critical-incident reporting rule addresses only high-severity failures, not routine monthly capability updates. Cloud Security Alliance described a "governance velocity mismatch": AI capabilities are changing on monthly timescales while regulatory frameworks update over years. Anthropic's alignment science lead Evan Hubinger stated in September 2026 that the company does not yet have a plan to keep a future superintelligence under human control, according to reporting by The New Stack.

The most consequential implication of Clark's revised 2027 estimate is not the capability threshold itself but the time available to close that safety gap before it is crossed. If the TasteVal doubling rate holds and the mathematical creativity signal Clark is tracking continues, the effective window for deploying alignment mechanisms that can operate at RSI-level capability has compressed from several years to roughly one — leaving the governance and technical safety communities racing a timeline that, by design, may not announce itself with obvious milestones until it has already arrived.