AI 2027 authors revise their forecast: AI is five months ahead of their scenario

When AI writes code faster, a team's next task is to check whether that translates into delivering a working product sooner. Developing the models themselves introduces an additional effect: an improved model can help build the next one. This feedback loop prompted the AI 2027 authors to revise their forecast timelines. For developers and founders, it raises a practical question: which processes should they start redesigning now if their tools begin improving faster than expected?

What changed in the forecast

On November 10, AI Futures Project published an update: every member of the team now considers the emergence of a superhuman AI researcher, or SAR, in 2027 likely, in the absence of a substantial, deliberate slowdown in development. Previously, Daniel Kokotajlo had assigned a 50% probability to SAR emerging before 2029.

The authors define SAR as a system capable of fully automating programming within an AI project and choosing research directions at the level of the best human researcher. This is a substantially higher bar than an agent that independently completes individual tasks in a repository.

The team also puts the probability of AI outperforming the best experts at virtually every task before the start of 2028 at 35–65%. The initial summary said “in 2028,” but the primary source specifies an earlier cutoff. The authors advocate pausing or substantially limiting the pace of frontier development. This is their proposal for managing risk. AI Futures Project update.

What METR found

The revision was prompted by METR's November 5 report on research inside Anthropic. Its authors believe model capabilities on tasks with clearly defined conditions are now improving at roughly twice the historical rate. Uncertainty is considerable: the researchers would not be surprised by an acceleration of either 1.3 times or three times.

METR tentatively attributes most of the current acceleration to the use of AI. However, Anthropic has kept most of the evidence the assessors considered most informative confidential. Outside readers therefore cannot independently verify a substantial part of the reasoning. METR also cautions that internal metrics may not fully capture work under less clearly defined conditions. METR report.

Based on these estimates, AI Futures Project calculated an AI-driven acceleration in progress of roughly 1.7 times. It was on this measure that the team calculated a lead of about five months over its scenario. Applying that adjustment to every predicted event would be a mistake: the comparison concerns a specific metric, while the timing of subsequent milestones depends on additional assumptions.

How to apply this to your team

Imagine a lab where an agent prepares the code for an experiment in an hour instead of a working day. The experiment then waits for an available GPU, training takes several days, and a researcher still needs to interpret the result. Shortening the first stage helps, but the overall cycle will speed up only as much as the remaining stages allow. If the agent helps select better experiments, the effect could be considerably stronger.

The logic is similar in product development. A team can produce more code yet wait just as long for a release because of requirements validation, code review, or acceptance testing. A sensible next step is therefore to choose a recurring class of tasks and measure the time from assignment to an accepted result. Track corrections, execution costs, and hours of human review separately. This will reveal where automation actually frees up resources.

For a founder, this measurement helps inform hiring and budget decisions. If the bottleneck remains in checking the result, generating more code will not, by itself, remove it. Product managers should check whether their teams can define tasks and approve changes quickly enough; otherwise, improvements in models may fail to translate into improvements in the product.

The updated forecast offers a reason to revisit these processes more frequently. On its own, it does not yet establish that an autonomous research team will emerge in the stated year. The practical decision today is to gather your own data on agent effectiveness and retain the flexibility to switch tools as verifiable results emerge.