Skip to main content
MindStudio
Pricing
BlogAbout
My Workspace
AI 2027superintelligence timelineDaniel Kokotajlo

Why Did AI Experts Move Up Their Superintelligence Timelines?

Dwarkesh Patel, AI 2027's authors, and Jack Clark all revised timelines forward. Here's the evidence that changed their minds.

Edited by Luis Chavez-Mattos, Director of Product RSS
Why Did AI Experts Move Up Their Superintelligence Timelines?

What changed in the superintelligence timeline debate?

In November, several of the most cautious voices in AI forecasting moved their predicted arrival of superintelligent AI research forward by months or years. Dwarkesh Patel, who argued a year earlier that AGI wasn’t imminent, now says AI as good as the best human AI researchers could be “a couple of months away.” The AI Futures Project, the team behind the AI 2027 scenario, published an update saying reality is running roughly five months ahead of their original forecast. Anthropic, in an amended SEC filing tied to its IPO, said it has crossed its own internal threshold for AI-accelerated research.

TL;DR

  • Anthropic’s SEC filing disclosed that the company believes it has crossed a self-defined safety threshold tied to AI accelerating its own research and development, based on expectations about future models rather than a single observed event.
  • Anthropic’s R&D automation index shows Claude went from handling under 1% of internal model-development tasks in February to 26% in the latest measurement, with more than 80% of code merged into Anthropic’s codebase reportedly written by Claude.
  • An independent report from METR, an organization that evaluates frontier model capabilities, found Anthropic’s internal capability scores have grown at roughly 1.9 times the prior rate since October of last year, with AI itself estimated to account for a large share of that acceleration.
  • AI safety researcher Ryan Greenblatt publicly said he was “seriously wrong” about the pace of AI-driven research speedup, revising his estimate from roughly 1.15 to 1.3 times to something closer to two times, with software-specific progress potentially running near 4.5 times faster.
  • Dwarkesh Patel and the AI 2027 team both cite the same underlying pattern: AI performance on math and coding has jumped from unreliable to superhuman in a few years, and they argue AI research itself shares the key trait, verifiability, that made that jump possible.
  • METR’s own task-length benchmark, which tracks how long a task an AI agent can complete independently, has reportedly stopped being updated because newer models have outpaced the test’s ability to measure them.
Let every team build.
Without shadow IT.
IT adminEvery app, approved
Campaign trackerMarketing✓
ExpensesFinance✓
OnboardingHR✓
InspectionsOperations✓
Shift schedulerNursing✓

What did Anthropic actually say in its IPO filing?

Anthropic maintains a “responsible scaling policy” with defined thresholds, essentially tripwires that trigger stronger safety measures once crossed. One threshold concerns AI accelerating Anthropic’s own research by around two times, substantially due to automation rather than just more compute or more hardware.

In a filing amended on November 10, Anthropic stated that its use of AI systems has already been accelerating its research and development, and that internal estimates point to further acceleration from future models. Because the policy applies when the company “observes or expects” that level of speedup, Anthropic said it has determined the threshold is met, based on expectations rather than a confirmed real-time measurement.

The filing also included a notable caveat: Anthropic said its evidence for this acceleration relies mostly on lagging indicators, meaning the company might not recognize a significant speedup until after it has already happened. That admission, that the company building the AI may not be able to see the acceleration in real time, is part of what drew attention from researchers and commentators tracking the filing.

Two days later, Anthropic’s founders published a letter ahead of the public offering stating that building AI safely may require the company, and ideally the rest of the industry, to deliberately slow the pace at which capabilities compound, particularly when that pace is driven by AI systems improving themselves. They wrote that there may be times when they choose to advance more slowly than they could, even at a competitive disadvantage.

How much of AI research is already being done by AI?

Anthropic tracks this directly with what it calls an R&D automation index, which rates internal model-development tasks on a scale from no AI involvement to full automation. By the most recent measurement, Claude was handling 26% of Anthropic’s model R&D work, up from under 1% roughly six months earlier.

A separate Anthropic report, “When AI Builds Itself,” coauthored by Jack Clark and Marina Favaro, said more than 80% of code merged into Anthropic’s codebase was written by Claude. In one internal safety research exercise, two human researchers working for about a week recovered roughly 23% of a performance gap on a test problem; Claude agents working on the same problem recovered about 97%. In head-to-head comparisons of research judgment, Claude Opus 4.5 reportedly picked a better next research step than a human researcher 51% of the time, and a newer preview model did so 64% of the time.

Extrapolating Anthropic’s current automation trend, one independent analysis estimated that 95% of the company’s model R&D could be fully automated by November 2027, the level Anthropic defines as fully autonomous recursive self-improvement. That projection assumes the current curve holds and is explicitly speculative since no task has yet been rated fully automated.

Why did the METR report change so many minds?

Developers have their tools.Enterprises have Remy.
Developers> add rate limiting to the API
● Reading src/server.ts
● Editing 3 files
+ import { limit } from 'rate'
- app.use(cors())
✓ Tests passed · ready for PR
Enterprises
Loan Approvals● Live
SOC 2Role-based accessSelf-hosted
Mortgage · $412,000Approved
Small business · $85,000Approved
Auto · $31,500Approved

METR, an organization that independently evaluates frontier AI capabilities, spent months analyzing internal Anthropic data and found that capability scores have been growing at approximately 1.9 times the earlier rate since October of last year, with a plausible range of 1.5 to 2.8 times. METR estimated that around 80% of that acceleration could be attributed to AI systems themselves rather than other factors like additional compute.

Ryan Greenblatt, a researcher who has previously tested models for METR, called the report “a big update towards super intelligence within 1 to 2 years” and said he had been “seriously wrong” about the expected pace. He had previously forecast only a 1.15 to 1.3 times speedup in 2026. Because some progress comes from simply using more compute rather than better methods, Greenblatt argued the software-specific acceleration must be even higher, and his own modeling put it at roughly 4.5 times. In the original AI 2027 scenario, that level of acceleration wasn’t expected until March 2027, meaning this estimate suggests reality is running months ahead of schedule. Greenblatt flagged an important caveat: much of METR’s underlying evidence was redacted, so the conclusions rely heavily on trusting the report’s authors.

Separately, OpenAI reported that its heaviest-using researchers went from spending about $1,000 a day on internal coding agents in early summer to over $7,000 a day by the end of August, a sevenfold jump in about three months, suggesting those agents were absorbing a growing share of real research work.

Is the AI 2027 scenario actually playing out?

AI 2027 is a forecasting scenario published by a team that includes Daniel Kokotajlo, a former OpenAI researcher. It projected a period of self-improving AI arriving around mid-2027, where AI systems handle most research progress and human contributions become marginal.

In their November update, the AI Futures Project said they are now roughly five months ahead of that original scenario, citing the METR report as the single biggest piece of new evidence. Kokotajlo had previously given a 50% chance of a superhuman AI researcher emerging before 2029. The team’s current view is that a superhuman AI researcher, defined roughly as the point where a company would rather lose all its human researchers than lose its AI systems, will likely arrive sometime in 2027 absent a deliberate slowdown. They now give a 35 to 65% chance that AI capable of automating all white-collar work arrives before 2028.

Jack Clark, Anthropic’s cofounder, wrote in his newsletter that he now considers full recursive self-improvement more likely in 2027 than his earlier estimate of 2028, citing recent progress in AI creativity, including novel math results, as the development he had been watching for.

Why does verifiability matter for AI research progress?

Dwarkesh Patel’s argument centers on a pattern seen in math and competitive programming: AI models went from unreliable at basic tasks to superhuman within a few years, specifically in domains where answers can be automatically checked. Models struggled with four-digit arithmetic in 2023, performed poorly on high school competition math in 2024, won a gold medal at the International Math Olympiad in 2025, and were reported to have solved a Millennium Prize problem in 88 hours this year, despite not being trained specifically on problems like it.

Self-host
Remy.
On-prem or in your private cloud. Remy and every app you build stay inside your walls.

The mechanism is verifiability. When a system can check whether an answer is correct, it can run through enormous numbers of practice attempts and improve iteratively. Patel’s claim is that AI research itself shares this property: whether a training technique lowers test loss, whether a model runs faster, or whether it uses less memory are all measurable outcomes, meaning they’re outcomes AI systems can be trained to optimize directly. If that holds, AI research could follow the same trajectory as math and coding, moving from assisting human researchers to outperforming them on a similar timeline.

Frequently Asked Questions

What does “recursive self-improvement” mean in this context?

It refers to AI systems meaningfully contributing to the design or training of future, more capable AI systems, creating a feedback loop. Anthropic’s policy defines a specific threshold for this: AI accelerating the company’s own research by roughly two times, substantially due to automation rather than added compute.

Did Anthropic say an AI model is now improving itself with no human involvement?

No. Anthropic’s filing describes AI significantly accelerating human-led research and development, with internal estimates pointing to further acceleration from future models. The company does not claim a model is independently building its successor without human oversight.

Why did Dwarkesh Patel change his view on AGI timelines?

He cites a few months of observing rapid AI progress on verifiable tasks like math and coding, plus ongoing debates with other researchers, as the reasons he moved from estimating “the better part of a decade” to possibly just months for AI reaching top human researcher level in AI research specifically.

What is the AI 2027 scenario?

It’s a detailed forecasting scenario written by a team including former OpenAI researcher Daniel Kokotajlo, projecting how AI capabilities and self-improving research might unfold, with a key turning point originally placed around mid-2027.

Is there solid proof behind these updated timelines, or is it speculation?

Some of it rests on measurable internal data, like Anthropic’s automation index and OpenAI’s agent usage figures. But key evidence, particularly in the METR report, was redacted before publication, and several of the most dramatic figures (like the 2027 automation projection) are extrapolations rather than confirmed outcomes.

Editorial standards

Presented by MindStudio

No spam. Unsubscribe anytime.