Our schedule is ahead of schedule!
On 11 November, Daniel Kokotajlo, a former OpenAI researcher, said, "I feel fear, sadness and anger", when he forwarded the team's new prediction to X.
He argued that the smart explosion would appear to have taken place in 2027: if the METR's judgement were true, AI-driven research and development would have accelerated about five months earlier than the initial AI 2027 launch.
When AI begins to help humans develop the next generation of AI, even the most radical predictors may underestimate the speed at which AI models progress.
In less than three months, the AI Futures Project, the forecasting team behind AI 2027, has scheduled the arrival of the super-AI researcher (SAR) for next year.
Previously, Daniel Kokotajlo, the most optimistic member of the team, gave only a 50% probability when referring to "super AI researchers before 2029".
On 10 November, the AI Futures Project released a new round of forecasts (Q3.5):
The signatory authors agree that super AI researchers are likely to emerge in 2027 without a significant and proactive slowdown in front-end AI research.
The most important basis for their ruling was a survey of Anthropic by the independent rating agency METR.
METR spent 11 weeks researching Anthropic, reviewing internal data, and interviewing employees, and eventually estimated a figure:
From October 2025, the benchmark for the Anthropic internal model has increased to the previous level.1.9 times。
According to this estimate, the one-year increase in Anthropic's model in this group of tests is equivalent to the growth of the past two years.
Running speed almost doubledHow much did AI do?On November 5, METR released the report.
To see how quickly the Anthropic model progresses, METR uses an internal indicator called AECI.
It combines the results of multiple tests into a single score, and most tests have a clear goal, and the results can be verified.
METR estimates that from October 2025, this total will increase at about the same rate as before.1.9 timesThe 95% confidence gap is between 1.5 and 2.8 times.
After October 2025, the annual increase in model integrated scores increased from about 15 points to about 29 points.
How to estimate the difficulty of the subject also affects the results.
METR reversed several assumptions and found that the rate of spin after October 2025 was about the same as before.1.1 to 2.9 times。
Several algorithms have different growth rates, but the trend is already clear: scores are gaining faster.
Why are the scores rising faster?
METR examines changes in training algorithms, recruitment, data and algorithms, and considers that changes in the rate of growth of training algorithms do not explain this acceleration.
Concrete evidence supporting this judgment was removed from the public version of the report.
As for how much AI contributes, METR believes that proportion also changes over time.
By the end of 2025, it is likely that factors other than AI will also play a role.
However, by October 2026, METR estimates that there will be an increase in the number of passengers on board.It's probably about 80% more AI than it was in the past.
Note that this "80 percent" is a subjective METR estimate. The report uses "We guess" (we guess) and makes it clear that there is a great deal of uncertainty in this judgment.
AI and other factors contribute to the model's acceleration, orange regions provide additional acceleration for AI, and gray is an undivided non-AI factor.
So what exactly did AI do?
One researcher said that in the past, it took a day or two to write code and run experiments.
With AI's help, he's willing to try a few more and see what works.
Other interviewees also mentioned that AI helps to write experimental code, monitor the running of experiments, and perform measurements that in the past required a lot of extra work to be done.
Some people even give ten ideas to an intelligent organism to help implement and test them.
From writing code to running experiments, AI is helping to develop the next generation of AI.
According to a new estimate from the AI Futures Project, this acceleration may be earlier than the initial introduction of AI 2027.
AI development is acceleratingIt's probably five months faster than AI 2027 predicted.The AI 2027 series, released in April 2025, demonstrates how AI can help develop the next generation of models.
As set at the time, the virtual company OpenBrain would have achieved the same amount of algorithmic progress, taking less than a week and a half of AI and less than a week of AI.
But algorithmic progress is only part of the model's progress.
An improvement that comes out of the experiment will have to be used in the next round of training to demonstrate the capabilities of the new model.
This time around, the AI Futures Project estimates that by early November 2026, the acceleration brought by AI alone could bring the overall progress of the model to an all-time high.1.7 times as much。
To see when this speed would appear in the old scenario, the team used the model to reproduce the overall trend of the AI 2027 model.
The results show that the same 1.7 times the speed,Around April 2027If the early November estimate is true, the model progresses 1.7 times faster than the previous scenario.About five months in advance.。
This result also exceeded the team's expectations less than three months ago.
Even on the basis of the most optimistic predictive model at the time, AI will push the overall rate of progress of the model to historical levels in November 2026.1.7 times or higherIt's not likely.4%。
Super AI researchersDeciding for Yourself What to StudyThe METR report also pushed ahead with the AI Futures Project's timetable for super-AI researchers.
According to their definition, super-AI researchers must first be able to sponsor programming for a cutting-edge AI project.
When it comes to choosing research directions, designing experiments, and interpreting the results of these things, it also needs the judgment of the top human research team.
When the code is written, which idea is the next to be tested? The experiment has failed, is it to continue to look for the cause, or elsewhere... these are decisions that are not already answered, and AI has to come up with their own ideas.
So when will AI cross this threshold?
In the same report, the three methods were used for more than a year.
One approach is to assume that AI has pushed the pace of model progress to 1.7 times its historical level.
According to the model, the predicted median is February 2027, which is about half the likelihood that super-AI researchers will appear before this point in time.
Another method is to push back the speed at which the model progresses in the test to estimate when it can be judged by the top research team.July 2027。
Another method of estimating the efficiency of anthropic employees' self-reported estimates has resulted inIn the middle of 2028。
What key evidence has Anthropic removed?At the beginning of the report, METR said that most of the evidence they considered most helpful in judging the core conclusions had been removed.
The report also lists a table of what internal data Anthropic has provided to METR and what results have been published publicly.
The runway trends of the internal model are public, with no estimates of algorithmic progress and key R&D inputs disclosed. As for AI's contribution to accelerating R&D, the public version lists only partial evidence.
By prior arrangement, Anthropic can delete non-public information, and METR can indicate which conclusions depend on these materials.
The reason given by Anthropic is that the disclosure of this confidential information could help other developers accelerate AI development.
After the deletion, what conclusions can the outside world verify?
According to the AI Futures Project,The speed of the run is about 1.9 times faster than before, with open curve support.
There is much less public evidence that "an additional acceleration could be 80% from AI" and 1.7 times that estimate.
The outside world is more dependent on METR's assessment of internal data to determine whether these two figures are reliable.
On 5 November, METR also released a review of the Anthropic August Risk Report.
METR and Anthropic both believe that automated R&D using the mid-July model is less likely to directly trigger disasters. Both parties are inclined to believe that the pace of R&D at the time was not up to the warning line set by Anthropic.
However, METR believes that Anthropic's public report underestimates the acceleration at the time, and that the actual acceleration may have already hit the warning line.
Next, let's see if AI researchers can write code faster and run experiments so that the next generation of models can be delivered sooner.
If an accelerated generation of accumulated AI, superintelligence (ASI) could arrive sooner than expected.
