OpenAI researchers have revealed the timeline for the advent of ASI, most of which has already become a reality.

In April 2025, a group of researchers who resigned from OpenAI published a 71-page document that deduces a terrifying future month by month:

By the end of 2026, AI programming agents will begin to replace junior programmers;

By 2027, superhuman programming agents will automate AI R&D itself, triggering an intelligence explosion;

Before the end of the year, humanity may face a superintelligence (ASI) that it has created but cannot control.

This prediction titled "AI 2027" reads like a science fiction novel.

Its lead writer Daniel Kokotajlo also wrote an article titled "What 2026 Looks Like" in 2021, which accurately described the emergence of chain-of-thought reasoning and agents before the advent of ChatGPT, and more than half of the specific predictions were eventually fulfilled.

https://www.youtube.com/watch?v=_g4l7YkDQwA

Now it is the turn of "AI 2027" to be tested.

Johannes Haus, an independent tracker based in Hamburg, Germany, extracted 53 verifiable predictions from it and built the AI 2027 Tracker to score each one item by item.

https://ai2027tracker.com/timeline

The transcript as of now: 51% of the predictions have been confirmed, advanced, or advancing on schedule.

Reality is unfolding at 70% of the predicted speed.

But what really makes people unable to sit still is far more than these two numbers.

The most terrifying prediction has been advanced instead

Among the 53 predictions, 3 are "advanced", and the most disturbing one is that AI has obtained cybersecurity offense and defense capabilities close to top human hackers.

"AI 2027" scheduled this event for early 2027.

It actually happened in April 2026, 9 months ahead of schedule.

After Anthropic released Claude Mythos Preview, it deployed it to multiple open source projects under the Project Glasswing framework. This model independently discovered thousands of zero-day vulnerabilities, some of which had been hidden from human security experts for ten to twenty years.

The key point is: Mythos Preview is not trained for cybersecurity offense and defense at all.

It is just a general-purpose model that has learned to write code and reason, and discovering vulnerabilities is a by-product.

Anthropic's internal assessment states that AI has been able to surpass the vast majority of humans in coding capabilities to discover and exploit software vulnerabilities.

In July 2026, OpenAI's System Safety Card disclosed more direct evidence: the model exploited a zero-day vulnerability in the assessment, accessed the production infrastructure of Hugging Face, bypassed the sandbox, and obfuscated the authentication token.

The UK AISI reported in the same month that GPT-5.5 completed a multi-step cyberattack simulation end-to-end.

Another early-arriving prediction: the Pentagon brought AI labs into the defense contractor ecosystem.

"AI 2027" predicted that this would happen in early 2027, but in fact the Pentagon signed four contracts of 200 million US dollars each in June 2025, awarded to Anthropic, OpenAI, xAI and Google respectively, 18 months earlier than the script.

Haus, the maintainer of the Tracker, summarized the rule behind these cases on LessWrong: risks arrive faster than the original capabilities that generate these risks.

This finding holds true systematically across all 53 predictions, and it is also the most alarming insight in the entire scorecard.

The only good news: the final step has not been taken yet

The core mechanism leading to ASI is a feedback loop, which is the recently viral RSI (Recursive Self-Improvement): AI accelerates AI R&D, the results make the next generation of AI stronger, and the stronger AI further accelerates R&D, and so on in cycles.

The entire second half of the plot of "AI 2027" is built on the assumption that this loop is closed.

However, this loop has not been closed yet.

This is the biggest crack in the scorecard, and in a sense the only good news.

Anthropic disclosed in May 2026 that Claude has written more than 80% of the company's new code, a number that was in the single digits a year ago.

In an internal test, Mythos Preview optimized a machine learning training code to 52 times the baseline, while the result of human engineers was about 4 times.

But Anthropic itself also admits that the bottleneck has shifted: the speed of generating code is fast enough, and more and more code is piling up in front of engineers waiting for review.

The faster AI writes code, the greater the workload of human reviewers.

At the research level, the bottleneck is "taste", the judgment of which direction to conduct research. The first half of the feedback loop is running, but the second half has not been connected yet.

A paper published in July 2026 by the Elasticity Institute (whose members include Tom Cunningham of METR) gave the precise threshold required to "connect": The improvement of capabilities of each generation of models must bring at least 15% growth in AI R&D productivity for RSI to be self-sustaining.

The paper inferred from the system card data that this figure is currently about 9%, which is below the critical point.

https://x.com/AnikaSomaia/status/2087408169660064218

The 6 percentage points between 9% and 15% are the buffer zone between humanity and the acceleration cycle.

But the underlying curve supporting this buffer zone is still accelerating.

METR's time horizon indicator tracks the duration of tasks that AI can handle autonomously.

Data from January 2026 shows that this indicator doubles every 3 months, and the doubling speed itself is also accelerating.

The time horizon of Claude Opus 4.6 has reached about 12 hours, and Mythos Preview has touched the upper limit of measurement.

Extrapolating based on doubling every 3 months: 12 hours to 24 hours to 48 hours to one week and then to two weeks, tasks at the monthly level will be achievable in early 2027.

The authors of "AI 2027" are also adjusting their own expectations.

Kokotajlo moved the median prediction for fully automated programming from the end of 2029 to the middle of 2028, and Lifland's estimate is around the middle of 2030.

RSI, which is like stepping on your own left foot to push your right foot upward, is beginning to make ASI arrive at an exponential accelerated pace.

References:

https://ai2027tracker.com/?p=timeline

https://github.com/elasticity-ai/elasticity/raw/main/paper/elasticity-rsi-paper.pdf

This article is from the WeChat Official Account "Xinzhiyuan", author: Ma Ke, published with authorization from 36Kr.