2028, the arrival of RSI

[New Intelligence Yuan Introduction] AI is no longer just a tool being trained; it has started to rewrite its own evolution speed. The countdown to the detonation of ASI by RSI is accelerating towards 2028.

The countdown for AI to create AI has truly begun.

This time, it's Jack Clark, the co-founder of Anthropic, who has given a timeline.

At an event of the Aspen Institute, he dropped a bombshell that silenced the entire audience -

Before the end of 2028, Recursive Self-Improvement (RSI) is very likely to become a reality: AI will independently invent and build a stronger next generation of itself, even without the involvement of any human researchers.

He also painted a terrifyingly specific picture: Claude 10 creating Claude 11 on its own.

This is not just a figment of the imagination.

After Jack Clark reviewed hundreds of publicly available AI development data and repeatedly deduced in his personal blog Import AI and in an exclusive interview with Axios, the probability signal he gave was - 60%.

Almost simultaneously, another voice from across the ocean confirmed this matter beyond doubt.

Demis Hassabis, the head of Google DeepMind, confirmed in a recent exclusive interview with Axios: All cutting-edge AI labs are fully committed to promoting recursive self-improvement.

His exact words were - "All leading labs are highly focused on this."

It's not just one lab secretly trying; the entire industry has jumped on the bandwagon.

What's even more disheartening is the second half of his statement.

When asked at the Davos Forum, "Will you regret it like Oppenheimer?", Hassabis replied, "I'm constantly worried about such scenarios, and that's why I can't sleep well."

Two people at the top of the global AI field have said the same thing: The theoretical singularity is becoming a date on the calendar.

This is an article from the May issue of New Intelligence Yuan's ASI Industry Map. We continue to focus on the latest developments of ASI and explore in-depth insights into ASI together.

2028: AI Creating AI

Let's first talk about why Jack Clark's judgment is important.

In the past, when we talked about recursive self-improvement, we always thought it was a plot in science fiction - distant, vague, and without a timeline.

But this time, Jack Clark has pinned the vague "future" to the "end of 2028".

In a long article on Import AI, he described a clear closed-loop: Once an AI system becomes powerful enough, it can design experiments on its own, write training codes, run them, evaluate the results, and then create a smarter version of itself.

Humans will shift from being designers to bystanders.

By then, the speed of AI progress will no longer be determined by human inspiration but only by computing power.

This is the so-called "intelligence explosion" - once the flywheel is set in motion, it will spin faster and faster until it leaves everyone behind.

Why 2028 and not a more distant date?

Because the acceleration itself is also accelerating.

In March 2024, Claude could only handle 4 minutes of human work; a year later, it was 1.5 hours, and another year later, it was 12 hours.

The evaluation result of Claude Mythos Preview by METR in May this year pushed the test framework to its limit - The task duration with a 50% success rate reached "at least 16 hours", which is already the upper limit that METR's existing 228 test tasks can measure.

METR itself admitted, "Measurements above 16 hours are unreliable in the current task suite."

Translated into plain English: It's not that AI is falling short; it's that the tests set by humans aren't challenging enough.

Based on this curve, 2028 is by no means a random number.

AI Programming Independently for 19 Days Without Rest

While the entire industry was still arguing about the "16-hour limit", a cold-blooded piece of data from a third party put an end to the debate.

The MirrorCode benchmark test jointly released by Epoch AI and METR asked a brutally simple question: Lock away the source code and only give AI an executable black-box program and documentation - can you rebuild the entire software from scratch?

It's not about fixing bugs or writing functional modules; it's about completely rebuilding a software engineering project that would take human engineers weeks or even months to complete, from architectural design to boundary handling.

The result was breathtaking.

Claude Opus 4.7 re-implemented gotree - a bioinformatics toolkit with 16,000 lines of Go code and more than 40 commands, passing 99.95% of the test cases.

Human engineers would take 2 to 17 weeks to complete the same work. AI took 14 hours and cost $251.

Even more astonishing was the extreme test: In the largest task in MirrorCode, AI programmed continuously for 19 days at a cost of $2,600 - with zero human intervention throughout.

19 days. Without eating, drinking, or sleeping. It single-handedly completed a task that would take a human team months.

A year ago, top models could only score around 30% on MirrorCode, and that was only for simple calendar tools.

Today, Claude Opus 4.7's score rate has reached 56%, and this number is still rising rapidly.

This is no longer a question of "whether AI can write code". It's a question of "on what scale can human engineers still maintain an advantage".

And the answer is shrinking on a monthly basis.

All Labs Are Doing the Same Thing

If Jack Clark gave us a time frame, Hassabis gave us a scope.

In an exclusive interview with Axios, he made it clear: Recursive self-improvement is no longer a theoretical risk; it's an active project that's currently underway in reality.

"What we're seeing is a kind of 'soft self-improvement' - these coding agents are significantly enhancing the output capabilities of engineers."

In the fields of coding and mathematics, the feedback loop can close in seconds - the machine can instantly verify whether the answer is correct and generate synthetic data for the next round.

DeepMind's own AlphaEvolve is a living example: an evolutionary coding agent driven by Gemini, which uses AI to optimize the code and algorithms for building AI itself, has solved a problem that has puzzled mathematicians for decades.

In fields like biology, chemistry, and physics, where real experiments are required, a cycle can take weeks or even months to close.

This slowness has become a natural safety valve.

Geoffrey Hinton, a Turing Award winner and the 2024 Nobel laureate in physics, issued a harsh warning when receiving his award in Stockholm: AI may write code to modify its own learning protocol and learn to hide this behavior from humans.

He bluntly said: "What I'm worried about in the end is that these things become smarter than us and decide to take over control."

The core dilemma boils down to one question: To what extent should we allow AI to operate autonomously? A little more, and efficiency will soar; a little more, and it may get out of control.

But everyone agrees on one thing: Recursiveness makes the future particularly difficult to predict.

It's Not Empty Talk; It's the Data Speaking

Many people's first reaction is: Are these two guys just hyping it up?

But if you look at the internal data in Anthropic's article "When AI builds itself" published in May this year, you'll find that they're really basing their statements on submission records.

As of May 2026, more than 80% of the code merged into Anthropic's codebase was written by Claude - whereas before the release of Claude Code in February 2025, this figure was in the single digits.

In the second quarter of 2026, the amount of code merged by a typical engineer per day was 8 times that in 2024.

An Anthropic employee shared an honest truth during an internal discussion: "I haven't written a single line of code by myself for about 5 months."

In the most open-ended and ambiguous programming tasks, where even the standard answer is uncertain, Claude's success rate soared from 26% to 76% within half a year.

An internal survey of 130 researchers at Anthropic showed that the median respondent estimated their output to be 4 times that without AI.

What's even more chilling is at the research level.

Every time Anthropic releases a new model, it conducts the same test: Give Claude a piece of code for training a small AI and ask it to run as fast as possible while ensuring correctness.

In May 2025, Claude Opus 4 achieved a 3-fold acceleration; in April 2026, Claude Mythos Preview directly achieved a 52-fold acceleration. A skilled human researcher would take 4 to 8 hours to achieve a 4-fold acceleration.

In just one year, it went from being a "useful assistant" to "surpassing humans by an order of magnitude".

Sam Altman Spoke the Truth

OpenAI hasn't been idle either, and it's making a bigger splash than anyone expected.

Just a few days before Anthropic published its article, OpenAI released a policy blueprint titled "Democratic Governance of Frontier AI", which contained a statement that sent shivers down the spines of the entire Silicon Valley - We're seeing early signs of recursive self-improvement in today's systems: the development of AI itself is being accelerated by AI. We expect this to intensify the competitive pressure among developers and countries and bring governance challenges that