Jack Clark Moves His RSI Forecast Up to 2027: What Changed and What TasteVal Shows

Anthropic co-founder Jack Clark has pulled his forecast for automated AI research forward by about a year. In his newsletter Import AI 476, published on October 12, 2026, he wrote: "I now think it's more likely we'll see full RSI become possible in 2027, whereas before I thought it was more likely in 2028." RSI means recursive self-improvement, the point at which AI systems can do the work of improving AI systems. This is a personal forecast from an influential insider, not a lab announcement and not a measurement, and this post explains what moved him, what the supporting benchmark actually shows, and where the caveats are.

If you follow the topic, this sits alongside our earlier coverage of the AI singularity debate, the superintelligence and Astra debate and the Ornith 1.5 self-improving open-weight model.

Quick reference

What exactly did Clark say?

The statement appears in a short "why this matters" paragraph under a benchmark item. Clark writes that in his earlier analysis of recursive self-improvement (Import AI 455, "AI Systems Are About To Start Building Themselves") he "repeatedly noted that creativity seemed like a key variable for how fast AI systems could develop the capacity to improve themselves." At the start of the year he had said AI systems "haven't yet displayed this [creativity] in a transformative and major way (though some of the results on accelerating math research are suggestive of this)."

Since then, he says, "we've seen both a tremendous advancement in the ability for AI systems to make creative-seeming math breakthroughs, and with benchmarks like this we see some signs they're becoming increasingly good at AI research as well." Then the forecast line. The primary text is on the Import AI site.

Note the careful wording: "become possible", "more likely", "2027 rather than 2028". It does not say it will happen in 2027, and it does not give a fresh percentage.

The May baseline: Import AI 455

To judge the change you need the starting point. In Import AI 455, Clark defined the target as a system that can train a more powerful successor with no human involved, and put the odds at roughly 60% by the end of 2028 and about 30% for 2027. The Decoder's summary lists the evidence he used then:

• SWE-Bench success rising from about 2% (Claude 2, late 2023) to 93.9%.
• METR time horizons growing from about 30 seconds for GPT-3.5 to roughly 12 hours.
• MLE-Bench top score rising from 16.9% to 64.4%.
• An internal CPU-training optimization test where mean speedup grew from 2.9x to 52x between Opus 4 and April 2026.

He also listed risks: alignment methods that "may break under recursive self-improvement", training environments that reward cheating, and models that can fake alignment. So the October line is an update to an existing, detailed view, not a new claim from nowhere.

What is TasteVal?

A benchmark bar filled to the top, representing how quickly AI research taste scores on TasteVal are climbing past a human baseline

TasteVal comes from P-Zero Research. According to Clark's summary, it tests how well an AI system can carry out 8 tasks representative of frontier AI development, "spanning pretraining data curation, pretraining and fine-tuning language models, preference modeling, and robustness to adversarial prompts."

The authors define research taste as "the ability to pick interesting problems to solve, design experiments, and interpret experimental results." But TasteVal measures only the middle part. Per the quote Clark gives, it measures "the experimental component of research taste; given a fixed research problem, we measure how well a model iteratively designs experiments and draws conclusions from their outcomes."

The human baseline is drawn from 24 experts, at least two per task, who recently worked at organizations including OpenAI, Google DeepMind, NVIDIA and Microsoft Research.

The numbers

A "compute multiplier" here means the model reaches the human expert level with that much less compute, or equivalently gets better results from the same compute. The interval is wide: the low end is 1.15x, barely above the baseline, and the high end is above 4x. That uncertainty is the main reason to treat the result as a signal, not a verdict.

The model named is Anthropic's, and Clark works at Anthropic. That is a fair disclosure point: the benchmark is by an outside group, but the person drawing the conclusion is at the lab whose model leads it. For readers who use the model, our Opus 5.5 launch coverage has the pricing and benchmarks.

Why creativity is the crux

Most RSI discussions focus on coding ability, because AI research is largely code. Coding is already strong. Clark's argument is that the limiting factor is something else: the ability to come up with ideas that are not in the training data. Examples from history are the transformer and mixture-of-experts architectures. If models can only polish known ideas, an automated research loop plateaus; if they can propose new ones, it compounds.

That is why two kinds of evidence moved him:

1. Math. Public reports of AI systems producing novel proofs and solving open problems. Our coverage of research agents finding a 1615 dodo hunt (including where they still fail) and Opus 5.5 agents proposing semiconductor candidates shows the kind of result in circulation.
2. Benchmarks of research behavior. TasteVal and similar tests that score how a model runs experiments.

Neither shows that a model can pick the right problem or hold a long research agenda. In May, researcher Herbie Bradley, as summarized by The Decoder, argued AI will absorb junior research work but not taste, vision-setting or long-term agendas. TasteVal's own definition covers all three parts of taste but measures only the experimental one.

What this means for what you build or pay

A benchmark ruler beside blocks of different heights, representing how builders should compare AI research benchmark claims

A forecast about 2027 does not change what you ship this week, but it has practical consequences.

• Expect agentic research tools to improve quickly. If experiment design is doubling every few months, tools that run training or evaluation loops will keep getting better. See how an open agent did on MLE tasks in our Sapient Praxist coverage.
• Treat benchmark cost claims carefully. "1/30 of the cost per attempt" compares model inference cost to expert time. It does not include the cost of building the harness or checking results.
• Safety investment follows capability. Clark's earlier risk list is the reason labs publish red-team work such as Anthropic's self-replicating IoT botnet research. If you deploy agents, plan for oversight now.
• Hiring and org design. Junior research tasks are the ones most exposed. For another insider's view of RSI and the data it needs, see Andrew Ho's departure from OpenAI and his RSI quote.

How to read a forecast like this

1. Separate possible from actual. "Possible in 2027" is a capability threshold. Whether anyone trains a successor without humans depends on compute, safety policy and incentives.
2. Check who is speaking. Clark is a co-founder of the lab building the leading model. That gives him unusual visibility and an obvious interest in the topic.
3. Look at the error bars. The TasteVal interval runs from barely above human to more than 4x.
4. Watch for a second source. At the time of writing, this is one newsletter issue and one benchmark paper we read through Clark's summary. We have not read the TasteVal paper itself, so quotes are as reported there.
5. Compare with counter-forecasts. Several independent writers have argued recursive self-improvement by 2028 is unlikely. A one-year move in one person's estimate does not settle that debate.

What else is in the same issue

Import AI 476 covers more than RSI. It also discusses the Humanity's Sixth Sense visual-reasoning benchmark, where the best model scores 53.6 against a human 93.1; Reflection's 501-billion-parameter Beam model with a planned open-weight release; and Mistral's Large 4, which we covered in our Mistral Large 4 open-weights preview post. We focus here on the forecast because it is the part with the widest implications.

What we are watching next

• Whether Clark publishes a new numeric probability for 2027.
• The TasteVal paper itself and independent replications on other models.
• Whether other lab leaders revise their timelines, as they did after earlier math results.
• Real-world signs: labs reporting how much of their own research code is written or planned by models.

Related reading

• The AI singularity, explained
• Has AI reached superintelligence? The Astra debate
• Ornith 1.5: a self-improving open-weight model
• Claude Opus 5.5 launch, benchmarks and pricing
• Sapient Praxist and MLE-Bench
• Mistral Large 4 open-weights preview

For teams deploying autonomous agents today, the AI-safety angle is practical: AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.

Details are accurate as of October 13, 2026. The forecast and benchmark figures are quoted from Import AI 476; we will update this post if the TasteVal paper or Clark's numbers change.