Anthropic says Claude is already deeply involved in developing its successor model, with the self-improvement timeline moved up to 2027

AI is now personally building the next generation of AI! Claude's "self-evolution" timeline moved up to 2027 Tech Little Mushroom 2026-10-31 23:14:27 From Beijing Updated 2026-11-01 09:10:40

The Wall Street Journal reports that artificial intelligence is personally building the next generation of artificial intelligence.

This was once a classic trope in science fiction and in the writings of futurists, but now it has become one of the most closely watched real-world topics in Silicon Valley. The American AI company Anthropic has publicly acknowledged that its model Claude is already deeply involved in developing its own "successor."

The industry calls this phenomenon "recursive self-improvement": AI improves itself, the stronger AI then improves the next generation, and so on in a loop.

In theory, once this loop truly closes, the pace of technological progress could far exceed anything in human experience. Some believe this will set off a technological explosion; others worry that humanity may unknowingly lose control of this acceleration.

80% of code written by AI

In June this year, Anthropic published a report titled "When AI Builds Itself," using a set of figures to depict the change that is already underway.

The report shows that as of May this year, more than 80% of the code merged in Anthropic's production codebase was written by Claude. When Claude Code was first launched in February 2025, that proportion was still only in the single digits.

Compared with 2024, the average amount of code merged per company engineer has grown to eight times the original. Some engineers have not personally written a single line of code for five months; their main work has become managing and reviewing the output of multiple Claude instances.

In research judgment, Claude's progress is equally evident.

In April this year, Anthropic also demonstrated Claude's ability to complete an open-ended AI safety research project end to end, with nine parallel agents running cumulatively for more than 800 hours.

When research ran into difficulties and a next direction had to be chosen, the proportion of cases in which Claude made the same judgment as senior researchers rose from 51% in November 2025 to 64% in April this year.

In August this year, Anthropic released another internal R&D survey. The results showed that 26% of the company's R&D work was already in an "AI-led" stage, that is, humans give the general direction, AI completes most of the work, and humans supervise.

The 26% here is not the share of researchers or working hours, but an estimate derived from task-weighted classification.

More than 90% of R&D work has reached human-machine collaboration or a higher stage. However, the company also said that no R&D field has yet achieved full autonomy.

The timeline is moving up

This acceleration is also changing expectations for the future within the industry.

Anthropic co-founder Jack Clark estimated in May this year that by the end of 2028, the probability of AI developing a successor model without human help was over 60%.

On October 12, he moved that prediction up to 2027. Clark stressed that this is only his personal judgment, not the company's product goal.

From "2028" to "2027," it looks like a difference of only one year, but it reflects the rapid change in the industry's judgment of the pace of technology.

One of the bases that prompted him to change his judgment is a new benchmark called TasteVal. The test was developed by researchers to evaluate AI's ability to select research experiments and interpret experimental results, that is, the so-called scientific research "taste."

The test shows that for models released since December 2025, the computational resource efficiency required to reach the human research level has roughly doubled every three months, far faster than the previous pace of about 14 months.

Researchers also cautioned that the humans involved in the comparison are not top scientists, the scale of test questions is small, and the conclusions cannot be simply extrapolated.

Calls to "hit the brakes"

Faced with such rapid progress, Anthropic itself has also shown obvious unease.

The report argues that a unilateral pause by a single lab would only shift the leading position to someone else, so multiple labs in multiple countries need to agree jointly and verify each other.

The report compares this idea to nuclear arms control, but also acknowledges that the training process of large models is far easier to hide than missile silos.

In July this year, Anthropic Chief Science Officer Jared Kaplan, OpenAI Chief Scientist Jakub Pachocki and about 1,300 AI practitioners signed an open letter calling on governments to coordinate so that slowing down R&D becomes possible.

Safety researchers' concerns are also very specific. Evan Hubinger, who is in charge of alignment stress testing at Anthropic, has said that subtle changes during training could spawn harmful model variants, and "we could certainly fail."

Of course, some people also remain skeptical of this "self-improvement" narrative. Critics argue that the existing data shows more of "faster programming" rather than true autonomous evolution.

Gary Marcus, professor emeritus at New York University, is one of the representative figures.

Moreover, much of the above data comes from Anthropic's own evaluations, and it is difficult for outsiders to verify independently. How to establish unified, transparent and comparable measurement standards has become a common issue facing the industry and regulators.

While selling a more powerful Claude to enterprises, it is also calling on the world to prepare to "hit the brakes." Anthropic's own situation reflects the contradictions of the entire industry.