Anthropic Co-founder Jack Clark Moves His Prediction For Self-improving AI Forward To 2027
AIThis post was created with the assistance of artificial intelligence (AI).

TL;DR

Jack Clark has moved his estimate for when AI could independently train a stronger successor from 2028 to 2027, pointing to advances in models’ ability to plan and evaluate research experiments. A new benchmark reports strong results, but it does not test whether models can choose a research agenda or build successor systems, and its authors warn the trend may be overstated.

Anthropic co-founder Jack Clark has moved his forecast for when AI systems could independently train a stronger successor from 2028 to 2027, citing recent signs that models are getting better at research tasks. The prediction concerns recursive self-improvement, a scenario in which AI systems help develop more capable successors; the benchmark Clark points to does not show that this has already happened.

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

  • Business-only prices and quantity discounts
  • Tax-exempt purchasing
  • Multiple users, one account, clear invoices
As an affiliate, we earn on qualifying purchases.

The evidence at the center of the discussion is TasteVal, a benchmark developed by P-Zero Research to measure how efficiently AI models can plan experiments and interpret their results in AI model development. Its eight tasks set a research goal in advance. Models choose experiments and use the results to decide what to do next, but they are not asked to select the broader research question.

P-Zero Research reports that Anthropic’s Opus 5.5 achieved a compute multiplier of 2.3 against the benchmark’s human comparison: it reached expert-level scores using less than half the compute, by the benchmark’s measure. The reported statistical range was wide, from 1.15 to 4.37. The human baseline came from 24 experts with recent experience at organizations including OpenAI, Google DeepMind, Nvidia and Microsoft Research; the best human attempt on each task was used for comparison.

The test controlled for coding ability by using a fixed coding agent to run experiments on one Nvidia H100 GPU. Each attempt ended after 40 GPU-hours or 120 hours of elapsed time. P-Zero Research estimates that a model run cost about one-thirtieth as much as an expert attempt on average. These details describe performance on the benchmark, not the cost or reliability of running an AI research program in practice.

At a glance
updateWhen: Forecast updated after a benchmark repo…
The developmentAnthropic co-founder Jack Clark has brought forward to 2027 his estimate for when recursive AI self-improvement could become possible, citing progress in AI research capabilities.

What the Benchmark Says About AI Research

TasteVal focuses on a capability that has become central to arguments about whether AI can accelerate its own development: choosing and evaluating experiments. If models can produce useful results with less compute, AI labs could potentially test more ideas within a fixed research budget. That could increase the pace of model development, although the benchmark does not establish that outcome.

The result also bears on a distinction often lost in broad claims about self-improving AI. A system can help carry out experiments without deciding which long-term research direction deserves attention, judging whether a result is important, or autonomously building a stronger successor. TasteVal addresses only part of that chain. Clark’s earlier forecast made those harder capabilities relevant to the timeline; the new result offers evidence on experiment work, not proof that the full process is within reach.

Products Worth Considering

Kindle Paperwhite Signature Edition 32GB (2024 Model) - Metallic Black

Kindle Paperwhite Signature Edition 32GB (2024 Model) – Metallic Black

Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB-C MagSafe Charging Case

Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB-C MagSafe Charging Case

As an affiliate, we earn on qualifying purchases.

Clark’s Forecast and Its Limits

In a May essay, Clark estimated roughly a 60% chance by the end of 2028 that a system would independently train a stronger successor, and a 30% chance by 2027. He has previously described creativity as a key factor in how quickly AI might improve itself. The revised 2027 date is a forecast, not a report that such a system exists.

The timing matters because Clark co-founded Anthropic, whose Opus 5.5 leads the reported TasteVal results. That connection does not invalidate either the benchmark or his view, but readers should distinguish the lab’s reported measurements from an independent confirmation of the forecast. P-Zero Research also says the benchmark could overstate the rate of improvement: its tasks are comparatively quick and inexpensive to check, a quality that can make them easier for AI labs to optimize.

Other evidence points to a continuing human role. A Fudan University-led study of AI agents used in model development found that agents contributed up to 55% of methodology proposals, while people made more than 85% of final decisions. The study’s authors said it remains unclear how AI can generate diverse research directions and judge their value before results are available. Those agenda-setting abilities are not tested by TasteVal.

“He 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.”

— Jack Clark, in his May essay

What TasteVal Cannot Establish

TasteVal does not test agenda selection: each task supplies the research goal, leaving the model to choose experiments and interpret outcomes. It therefore cannot show whether a model can independently identify promising problems, sustain a coherent research plan, or train a stronger successor.

The reported trend also rests on a limited set of tasks and model generations. P-Zero Research says leading models’ compute multiplier has roughly doubled every three months since December 2025, but gives a wide estimated doubling-time range of 1.7 to 5.0 months. The authors say a single benchmark cannot establish whether that pace will persist. The source material does not specify the full underlying dataset or independent replication of the headline result.

It is also unclear how Clark’s revised probability estimates have changed. The reported shift is in the date he sees as possible; no new probability for 2027 is provided. The benchmark’s wide statistical range and its human comparison group, which did not include top-tier researchers, further limit what can be inferred about performance against the strongest research teams.

Evidence Needed Beyond Experiment Tasks

The next test is whether the reported gains hold across broader research problems and persist in independently evaluated benchmarks. More evidence would also be needed on whether models can generate and rank research directions, rather than only pursue goals selected for them, and whether they can use that work to produce a measurably stronger successor.

For now, Clark’s 2027 date should be read as a revised forecast grounded partly in benchmark progress, not a confirmed milestone. Further evaluations and real-world studies of AI-assisted model development may clarify whether models are taking on more than bounded experiment work—and whether the trend reported since late 2025 continues.

Key Questions

What does recursive self-improvement mean?

It refers to AI systems helping develop more capable successors. The forecast discussed here is about a system independently training a stronger next-generation system; the source does not report that this has happened.

What did Clark change in his forecast?

Clark moved his estimate of when recursive self-improvement could be possible from 2028 to 2027. His earlier May estimates assigned roughly a 30% chance to the capability by 2027 and 60% by the end of 2028. No updated probability is provided in the source material.

What does TasteVal measure?

TasteVal measures how efficiently models can choose experiments and interpret their results on fixed AI research tasks. It does not ask models to choose the research questions or long-term agenda.

Does the benchmark prove AI can improve itself?

No. It reports performance on a limited set of experiment-planning tasks. It does not show that a model can independently select a research direction or train a stronger successor, and its authors warn the measured improvement trend may be overstated.

Source: rss

You May Also Like

The Role Of AI In Innovation: Fields Medalist At OpenAI And ByteDance’s New STEM Scientist Program

A Fields Medalist has reportedly joined OpenAI, and ByteDance has launched a STEM Scientist Program, highlighting intensified competition for top research talent in AI.

GLM 5.2 And The Coming AI Margin Collapse

Analysis of GLM 5.2’s release and its implications for AI profitability and industry margins.

Quiet GPUs for Local AI: Acoustic and Thermal Roundup

A comprehensive roundup of the quietest and coolest GPUs for local AI in 2026, focusing on acoustic performance, thermal management, and practical recommendations.

How A 27-Year-Old Tsinghua Professor Entered AI Self-Evolution With A 35B Model

A headline reports a 35B-parameter model linked to a 27-year-old Tsinghua assistant professor, but provides no article details to verify the claim.