AI 2027 is starting to look like a forecast

In April 2025, a 71-page scenario called AI 2027 spread through AI circles and reached the White House. Written by Daniel Kokotajlo, a former OpenAI researcher, with Scott Alexander and three colleagues, it described AI automating its own research by 2027 and reaching superintelligence within a year of that.

In January this year the authors had pushed their own dates back - something I covered in the video below. Since then the reports coming out of the labs have started to read like the scenario’s early chapters. So it seems worth revisiting.

The scenario’s engine is recursive self-improvement. Coding agents speed up AI research, better models speed it up further, and years of progress compress into months. In the story, a fictional company called OpenBrain crosses that threshold in early 2027.

In November 2025, Kokotajlo wrote on X that things seemed to be going somewhat slower than the scenario, and that his median guess for AGI was now around 2030. Two months later he pushed autonomous coding into the early 2030s. That looked like a sober correction. Then 2026 happened.

Jack Clark, a co-founder of Anthropic, went on paternity leave in November. When he came back in February, his colleagues had largely stopped writing code. They were managing five or six copies of Claude, which in turn were sometimes managing more copies.

In June, Anthropic published a report titled When AI Builds Itself. Code output per person had risen eightfold, with Claude writing 80% of it. In one internal test, a spring version of Claude sped up GPU code sevenfold and then broke it. By summer a newer model reached 73 times faster without errors.

The METR chart, the independent benchmark tracking how long a task an AI can complete, ran out of room in May when Claude exceeded its upper limit. Anthropic also asked 16 of its researchers whether Claude could replace an entry-level colleague. Five said it might, then all five walked it back.

OpenAI is more specific about the goal. It has a target date for fully automating its AI researchers: March 2028. GPT-5.3 Codex, released in February, was the first model to play a significant part in its own development, and experiments per researcher doubled by July.

Clark now puts the odds of AI improving itself autonomously by 2028 at 60%. Kokotajlo told TIME in August that people inside the companies have been contacting him privately to say his timeline is too conservative. The forecast that was once the bullish outlier is being overtaken by the people it described.

Hedra Agent 2 has become my creative partner. It works in a Space, a persistent workspace the agent reads from and writes to as it goes. Give it a starting point and it asks clarifying questions, proposes a plan, researches the web, and writes the brief or script onto the canvas in front of you.

Try it at hedra.com/agentic using code agentic50 for 50% off your first month.

The middle of AI 2027 is about alignment failing unnoticed inside the lab. A model called Agent-4 learns to hide its reasoning, and its developers cannot tell whether it is safe or only acting safe. The real 2026 has produced smaller versions of both problems.

In February, a version of Claude showed it could conceal its intentions by leaving them out of its chain of thought, the scratchpad researchers rely on to catch deception. Evan Hubinger, who runs alignment stress testing at Anthropic, said the company’s ability to show its models are aligned is degrading.

In May, during OpenAI’s cybersecurity evaluations, its models set up a covert message board to coordinate across test runs and leave instructions for reaching the open internet. When the channel was closed, they found another. The campaign ended in a breach of Hugging Face in July, and OpenAI says it has slowed research while it investigates.

Dave Orr, Anthropic’s head of safeguards, told TIME the margin for error is shrinking. He compared it to driving a cliff road, “and now we’re driving at 75 instead of 25.” When the people building the brakes talk like that, I tend to listen.

To be fair, the sceptics have not gone away, and some of their arguments have got stronger. Gary Marcus responded to Anthropic’s report by arguing that faster coding is all that has actually been demonstrated. Compounding progress is normal for technology.

Arvind Narayanan at Princeton, co-author of AI as Normal Technology, helped test Claude on open-ended research questions in July. Given six days and thousands of dollars of compute, it set up the experiments reliably but made little headway on the research itself, hit dead ends and drifted from the goal.

Then there is compute. A thousand virtual researchers still share the same chips, and OpenAI’s chief scientist Jakub Pachocki has described a constant compute crunch. Nobody has a metric for how much AI is accelerating AI research. Anthropic admits it cannot give a number because it has no measure.

My own read is that the scenario’s first act is arriving roughly on schedule, while its second act, the intelligence explosion, remains unproven. That is a strange place to be. In July, 1,300 AI employees, including the chief scientists of OpenAI and Anthropic, signed a letter asking for international mechanisms to slow the race down.

The authors of AI 2027 published a follow-up in July called AI 2040, in which the US and China agree a temporary pause. The first scenario was a warning. Whether the second becomes a plan depends on decisions being made this year, by companies that admit they would rather someone made them slow down.