ResearchPapers, evidence & method
Dwarkesh Patel moves his forecast for AI matching top AI researchers to 2028, citing METR's estimate that Anthropic's progress has nearly doubled
Dwarkesh Patel now puts the median arrival of AI matching top human AI researchers at 2028, leaning heavily on METR's estimate that Anthropic's capability progress has nearly doubled since October 2025, mostly because of AI, while his math exhibit remains contested.
From the Dwarkesh Podcast podcast episode of November 13, 2026 and cited sources: Patel has cut his estimate from about a decade to a median of 2028 for AI matching the best human AI researchers, and 2030 for matching experts in every field. [1]
SOURCES[1] Dwarkesh Patel – The avalanche that hit math is coming for… dwarkesh.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Patel has cut his estimate from about a decade to a median of 2028 for AI matching the best human AI researchers, and 2030 for matching experts in every field. [1]
- Original source
- Dwarkesh Podcast podcast episode, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 13, 2026 podcast episode published
- Covered
- Nov 14 edition
- Evidence
- 14 sources across 14 sites; no transcript
- 1Dwarkesh Patel – The avalanche that hit math is coming for…dwarkesh.com
- 2METR – Anthropic AI R&D acceleration assessmentmetr.org
- 3Toby Ord – Mathematics is much more than prooftobyord.com
- 4OpenAI – Research acceleration: the view inside OpenAIopenai.com
- 5ITPro – OpenAI researchers' $7,000/day token useitpro.com
- 6CNN – OpenAI says it has solved a Millennium Problemedition.cnn.com
RisingHow it’s measured
YouTube video · 300.8K views · 5.5K likes
Reach blends public engagement (views, likes, comments, shares, stars) from the platforms where the story or its original source lives: YouTube, Hacker News, GitHub, Substack and similar. A video and the article written from its transcript share one score, taken from the video. Platforms with no data are left out rather than counted as zero, and the age of a story never lowers its score.

Patel has cut his estimate from about a decade to a median of 2028 for AI matching the best human AI researchers, and 2030 for matching experts in every field. His reasoning is that AI research, like math, has progress signals a system can check, such as held-out loss, internal benchmarks and inference efficiency, so it should be just as open to automation. [1]
After 11 weeks of high-access work inside Anthropic, METR estimated a 1.9x speedup in capability progress since October 2025 (95% CI 1.5–2.8x). It attributed about 80% of that speedup to AI and found no clear bottleneck likely to slow it within a year. The estimate rests on a regression fitted to Anthropic's internal capabilities index, staff interviews and pipeline data. Anthropic redacted much of that material, and METR could not query the models directly. [2]
Lab data points the same way, but the labs report it themselves. At OpenAI, the median researcher uses more than $600 a day of agent inference, and the 90th percentile uses more than $7,000. Anthropic says Claude leads over 26% of its R&D work, up from under 1% in February, with about 30,000 agents running at once. [4] [5] [10]
The math case Patel cites is still disputed. In OpenAI's Navier–Stokes result, the written proof claims more than the Lean-checked version. Experts also say OpenAI misdescribed the method and that the paper needs a human audit. Separately, Toby Ord argues that automating proofs is not the same as automating mathematics. [3] [6] [8] [11] [14]
In the reviewed sources, METR's estimate is the strongest evidence that AI is speeding up AI research at a frontier lab. If that pace holds, model release and evaluation cycles would shrink, and safety and governance work may struggle to keep up.
Sources are gathered. Writing the dossier now.
Executive brief
Dwarkesh Patel now argues that AIs as good as the best humans at AI research are "a couple of years" away, and possibly only months. His estimate came down from roughly a decade earlier in 2026 (Dwarkesh). The strongest evidence behind the shift is not his. METR spent 11 weeks with high access inside Anthropic and estimated that Anthropic's capability progress has roughly doubled since October 2025, with about 80% of that speedup likely due to AI. Anthropic, however, redacted most of the evidence METR found most informative (METR). His other main exhibit, OpenAI's Navier–Stokes proof, is still disputed.
What changed and event timeline
Ord separates proof from mathematics
Toby Ord argues that AI is automating proof but has shown little ability to ask good questions or open new fields of mathematics ().
OpenAI reports agent usage inside its research org
The median researcher uses more than $600 a day of agent inference. The 90th percentile uses more than $7,000. Agents log 3.1 agent-workdays per human workday (;).
OpenAI claims a Navier–Stokes solution
About 10,000 agents running an unreleased model produced a blow-up proof in 88 hours. Formalizing it in Lean took another 17 hours (;).
Researchers find the two versions of the proof don't match
A King's College London and Cambridge team found that the written proof states a stronger claim than the computer-checked Lean version ().
METR publishes its Anthropic assessment
METR estimates a 1.9x acceleration (95% CI 1.5–2.8x) and attributes about 80% of it to AI. It found no clear bottleneck likely to slow things within a year ().
Patel publishes the essay
It puts the median date for AIs matching human AI researchers at 2028, and AIs matching experts in every field by 2030 ().
Capabilities and access
- Navier–Stokes model: an unreleased OpenAI model described as significantly more capable than GPT-6 Astra. GPT-6 Astra handled the Lean formalization (OpenAI on X).
- Anthropic: about 30,000 agents run at once on its internal platform. Claude does not yet work fully autonomously (level AL5) (Anthropic).
- OpenAI: internal coding agents. The post does not name the models.
Read the full section
- Navier–Stokes model: an unreleased OpenAI model described as significantly more capable than GPT-6 Astra. GPT-6 Astra handled the Lean formalization (OpenAI on X). It is not publicly available.
- Anthropic: about 30,000 agents run at once on its internal platform. Claude does not yet work fully autonomously (level AL5) (Anthropic).
- OpenAI: internal coding agents. The post does not name the models.
Technical analysis for researchers and developers
- How METR measured acceleration: it fitted a piecewise linear regression with one breakpoint to AECI, Anthropic's version of the Epoch Capabilities Index.
- How Anthropic built its automation index: a task tree of 542 nodes drawn from employee work records.
- Patel's core argument: AI research has checkable progress signals: held-out loss, internal benchmarks and inference efficiency.
Read the full section
- How METR measured acceleration: it fitted a piecewise linear regression with one breakpoint to AECI, Anthropic's version of the Epoch Capabilities Index. It added about 15 staff interviews, Slack uplift polls (self-reported speedup of about 12x) and pipeline data. METR could not query the models, and a planned compute check was cancelled partway through (METR).
- How Anthropic built its automation index: a task tree of 542 nodes drawn from employee work records. Claude rated the automation level of each task, and humans agreed within one level 97% of the time (Anthropic).
- Patel's core argument: AI research has checkable progress signals: held-out loss, internal benchmarks and inference efficiency. That should make it as open to automation as math has been (Dwarkesh).
Claims and evidence
Read the full section
- Anthropic's pace doubled, mostly due to AI: this is METR's own analysis. It rests largely on data Anthropic redacted, and METR says a range of 1.3x to 3x would not surprise it (METR).
- Claude "leads" more than 26% of Anthropic's R&D work, up from under 1% in February: reported by Anthropic only, and a single snapshot (Anthropic).
- $7,000 a day at the 90th percentile: reported by OpenAI only (OpenAI).
- Millennium Prize problem "solved": OpenAI's claim. Outside experts dispute how it was described and say the written proof still needs independent checking (Scientific American).
Context and prior work
Patel traces a quick sequence in math: unreliable arithmetic in 2023, an IMO gold medal in 2025, then open Erdős problems (Dwarkesh). Ajeya Cotra read the METR results as implying roughly a 10x productivity gain for researchers. She warned that safety and governance work is not ready for that pace (Planned Obsolescence). Forecasting groups have moved their timelines in the same direction (AI Futures).
Limitations, safety and contested findings
- The proof's method was misdescribed. OpenAI said it followed the "Spanish program" of Córdoba and Martínez-Zoroa.
- Lean checks are not the whole answer. Mathematicians call the paper close to unreadable. Terence Tao says Lean proofs need a human audit (The Decoder).
- METR lists reasons it may be overestimating: hiring, an unusually slow baseline period, and gains concentrated on well-specified tasks (METR).
Read the full section
- The proof's method was misdescribed. OpenAI said it followed the "Spanish program" of Córdoba and Martínez-Zoroa. Gonzalo Cao-Labora and others say the approach is different (Scientific American).
- Lean checks are not the whole answer. Mathematicians call the paper close to unreadable. Terence Tao says Lean proofs need a human audit (The Decoder).
- METR lists reasons it may be overestimating: hiring, an unusually slow baseline period, and gains concentrated on well-specified tasks (METR).
- Anthropic's oversight figures: 6% of AI R&D compute goes to safety, and monitors block 0.002% of agent actions (Anthropic).
- The essay overstates in places. It calls Navier–Stokes settled, and frames METR's benchmark-based result as Anthropic's "overall rate of progress on well-scoped tasks."
Business and practitioner implications
- Agent spend is now research capital. A heavy user at $7,000 a day costs more in inference than many researchers cost in salary.
- If the 1.9x pace holds, vendor roadmaps and evaluation cycles will compress.
- Treat formal verification as necessary but not sufficient. Check that the Lean statement matches the claim being made.
Read the full section
- Agent spend is now research capital. A heavy user at $7,000 a day costs more in inference than many researchers cost in salary.
- Plan for faster model releases. If the 1.9x pace holds, vendor roadmaps and evaluation cycles will compress.
- Treat formal verification as necessary but not sufficient. Check that the Lean statement matches the claim being made.
- Press for third-party audits with less redaction. Both METR and Anthropic say outside verification and standard methods are needed.