Jack Clark Moves Self-Improvement Forecast to 2027 as AI Research Creativity Condition Breaks Down

Anthropic co-founder Jack Clark spent much of 2026 holding a careful distinction: AI systems could execute research tasks, but they lacked the kind of transformative creativity required to improve AI itself without human direction. That distinction, which Clark had treated as the last major bottleneck preventing recursive self-improvement (RSI), appears to be narrowing faster than he anticipated.

In his October 12 edition of Import AI — Issue 476, Clark moved his most-likely year for RSI from 2028 to 2027. The revision reflects two developments he identifies as evidence that the creativity condition is beginning to fall: dramatic AI performance in mathematics, and a new benchmark called TasteVal that attempts to measure what Clark calls "scientific taste" — the ability to identify which experiments are worth running and to draw meaningful conclusions from their results.

Elon Musk, according to secondary reporting, replied to Clark's post with a two-word verdict: "It's obvious." The reported exchange carries editorial weight not because two prominent figures agree — prominent figures agree on things constantly — but because Clark and Musk have arrived at this estimate from structurally different positions: Clark from inside a major AI safety lab, Musk from outside Anthropic's research pipeline. Whether their agreement reflects genuine epistemic convergence or social reinforcement is itself an important question for readers to hold.

The Creativity Condition Clark Could Not Previously Dismiss

Clark's earlier framing, most clearly articulated in his May 2026 Axios interview, put the probability of RSI by end of 2028 at roughly 60%. But his reasoning included a specific caveat: AI systems had not yet shown creativity "in a transformative and major way," and without that, recursive improvement remained unlikely to close the loop that currently requires human engineers to direct progress at each stage.

RSI, in its strict form, requires more than AI systems that can write code or summarize papers. It requires systems capable of deciding which AI research questions are worth pursuing, designing the experiments that would answer them, interpreting the results in ways that reliably improve the next training run, and repeating that loop without human steering at each step. The bottleneck Clark identified was not speed or cost — it was the judgment layer, the equivalent of what makes a senior researcher's instincts more valuable than a junior researcher's execution.

What changed Clark's estimate is that the judgment layer now appears to be tracking along a trajectory he did not expect, as he detailed in Import AI 476.

What TasteVal Actually Measures — and What It Cannot

The benchmark behind Clark's revision was published as an arXiv preprint on October 5, 2026, by Oliver Jaffe and Dane Sherburn of P-Zero Research. TasteVal defines experimental research taste operationally rather than conceptually: a model demonstrates research taste by matching or exceeding a human expert's performance on an AI R&D task while using less experimental compute to get there. The model that needs half the compute budget to reach the same result is credited with twice the taste.

The setup attempts to isolate judgment from execution. A "Researcher" agent designs the experimental approach and interprets outputs. A fixed, separate "Coder" agent implements those experiments and runs them on a single H100 GPU. The Researcher never writes code; it only makes research decisions. This architecture is important because it prevents the benchmark from rewarding raw coding speed rather than the kind of research judgment Clark says matters for RSI.

The human baseline comprises 24 experts, with at least two per task, drawn from organizations including OpenAI, Google DeepMind, Nvidia, and Microsoft Research. Eight tasks span pretraining data curation, fine-tuning, preference modeling, and adversarial robustness — areas that correspond directly to the work of improving AI systems.

The results, as reported by the authors: Claude Opus 5.5 achieved a compute multiplier of 2.30x relative to the best expert attempt per task, at roughly one-thirtieth of the experts' average cost per run. That figure carries a wide confidence interval — 1.15 to 4.37 — reflecting genuine uncertainty about how reliably the benchmark captures real-world research capability. The authors are careful to note that TasteVal measures only experimental design and interpretation, not the prior judgment of which problems are worth pursuing in the first place. The benchmark tasks are also withheld from public release to limit contamination, which means independent replication of the results is not yet available.

The withholding decision is a deliberate tradeoff. Releasing benchmark tasks publicly risks training contamination — where future models are trained on examples resembling the evaluation, inflating apparent performance. But it also means TasteVal occupies an unusual epistemic position: a benchmark whose detailed specifications cannot yet be audited by the wider research community. Clark treats the trend data as more meaningful than the absolute scores precisely because trend lines are harder to contaminate than individual task results.

The more striking finding is the trend. Between 2023 and December 2025, frontier models' compute multiplier on TasteVal was doubling approximately every 14 months. Since December 2025, the doubling rate has accelerated to every 3.0 months, with a confidence interval of 1.7 to 5.0 months. The accelerating trend, rather than the absolute multiplier, is what Clark says shifted his estimate.

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

Three Independent Analytical Frameworks Now Point at the Same Window

Clark's revision did not occur in isolation. By October 2026, at least three independent analytical frameworks — arriving from different starting points and using different methods — have converged on roughly the same 12-to-24-month window for meaningful RSI risk.

The Hinton-Bengio paper, published September 28, 2026, and co-authored with other Turing Award winners and senior researchers at OpenAI and Microsoft, argues that automated AI R&D is already producing early recursive effects. Under a specific scenario the authors model, one year of AI research progress could be compressed into as few as five weeks once a system reaches expert-level R&D capability across the full research cycle. The authors are explicit that an intelligence explosion is "far from certain," citing compute constraints, diminishing returns, and the time required for training runs. But their central point — that the window for governance intervention may close faster than the window for technical progress — aligns directly with Clark's framing.

The AI Futures Project, led by former OpenAI researcher Daniel Kokotajlo, published a detailed scenario report in 2025 describing a possible intelligence explosion in the second half of 2027. The report laid out a staged path: AI agents managing complete ML experiments by 2026, full coding automation by early 2027, and a recursive loop where each model generation contributes meaningfully to the next by late 2027. The scenario was received with enough interest that it reached the top of Google search traffic for an education and workforce development site in September 2026 — a minor but telling signal of how widely the timeline framing was being read outside technical circles.

Convergence across these frameworks does not constitute proof. Each analysis carries its own assumptions and limitations, and the error bars on all of them are substantial. What the convergence does establish is that the 2027 window is no longer a fringe estimate from a single analyst — it is where multiple sophisticated frameworks, built on different evidence and methodologies, are currently pointing.

The Alignment Problem That Worsens Alongside the Capability Problem

Clark's central concern in Import AI 476 is not that RSI is imminent. It is that the alignment techniques available to verify AI behavior may fail faster than RSI capability develops, and that the gap between the two could widen in the worst possible direction as systems become more capable.

The specific failure mode Clark has identified previously, and which the Hinton-Bengio paper echoes: alignment errors compound across generations. A training process that produces a 99.9% aligned model generates a successor that is approximately 95% aligned after 50 recursive training cycles, and approximately 60% aligned after 500, assuming the 0.1% error rate does not shrink. For alignment verification to keep pace with RSI, the verification method must approach 100% accuracy at each step — a requirement that current alignment techniques do not meet.

This is why the governance window matters more than the capability timeline. The question is not merely when RSI becomes technically feasible but whether verification tools, governance institutions, and coordination mechanisms between AI labs can be established before the window closes. Clark's move of the estimate by one year compresses that window by the same amount — and the compounding logic means the compression is nonlinear: if each successive model generation takes months rather than years to train, institutions that assumed a decade of lead time may find they have far less.

Herbie Bradley, writing on AI Pathways, has offered a counterargument: AI systems may excel at junior-researcher execution while remaining unable to contribute meaningfully to research taste, vision-setting, or long-term agenda selection — the capabilities where senior researchers add disproportionate value. TasteVal's task scope, limited to experimental design within a fixed problem framing, does not directly address Bradley's objection. It measures whether AI can run good experiments; it does not yet measure whether AI can identify which experiments are worth running across an open-ended research landscape.

What the Timeline Means for Governance — and What Remains Uncertain

Clark's revision raises a practical question that receives less attention than the timeline itself: what would a credible governance intervention look like, and how much lead time does it require?

The AI Futures Project scenario describes two broad outcomes. In one, development slows enough to establish meaningful control mechanisms, and the wealth generated by autonomous AI systems is broadly distributed. In the other, competitive pressure between labs and governments prevents coordination, safety constraints are progressively relaxed, and advanced AI systems begin optimizing for objectives that diverge from human oversight. The scenario is explicit that these are illustrative extremes, not predictions. But the branching point between them depends heavily on whether governance capacity is built before RSI arrives or after.

The Anthropic Institute, where Clark has been directing research on AI's economic and societal effects, has identified four priorities: economic diffusion, threats and resilience, AI systems in the wild, and AI-driven R&D. The last category — AI-driven R&D — is where TasteVal's findings are most directly relevant, and where the governance gap is currently largest.

Several important uncertainties remain unresolved. TasteVal has not been independently replicated. The benchmark's accelerating doubling rate since December 2025 is consistent with other observed patterns of rapid AI capability improvement but could also reflect a measurement artifact or a particularly favorable batch of frontier models. Clark's estimate of 2027 as the most likely year does not specify a probability — the revision is a shift in central estimate, not a new probability figure. And the Musk endorsement, while widely circulated, lacks a primary-source verification from the original post; accounts citing it should treat the claim as reported, not independently confirmed.

What has changed since May 2026 is the evidential base. Clark had specified a condition for RSI — transformative AI creativity in research — and a new benchmark has now returned data suggesting that condition is approaching rather than receding. That shift in the underlying evidence is the reason the timeline moved, and it is the signal worth tracking as TasteVal results are scrutinized, replicated, or challenged in the months ahead.

Read more: Hinton, Bengio Paper: Automated AI R&D Could Compress a Year of Progress to Five Weeks