Anthropic engineers are no longer writing code: Claude is training the next generation of Claude, and the CEO says "it's uncertain how much time is left."

Anthropic’s internal engineers have almost completely stopped writing code by hand, leaving it all to Claude; the deeper issue is that Claude is assisting in training the next generation of Claude, creating a rapidly tightening feedback loop of AI self-evolution.
This article is derived from Dario Amodei’s interview on the Dwarkesh Podcast, edited and translated by PA Media.
(Previous context: A decade-long feud: If OpenAI had never been hypocritical, there would be no strong Anthropic.)
(Background note: 360周鴻禕: Tokens can never be like mobile data “unlimited”, AI will only become more expensive the more it is used.)

Article Directory

Toggle

• What are engineers doing if they aren’t writing code?
• Dario explicitly states, “I’m not sure how much time is left.”
• The logic of the feedback loop and overlooked issues.

What is the daily routine of Anthropic’s engineers now? According to various sources compiled by PA Media, the answer is quite straightforward: open Claude, watch it finish writing the code, and then “take a glance” to confirm there are no issues. A report from Fortune provided quantitative figures—AI writing code accounts for between 70% to 90% of the entire company at Anthropic, with the percentage of top engineers stating it is directly at 100%.

What are engineers doing if they aren’t writing code?

This question deserves a serious answer, as the answer is more unsettling than it appears on the surface.

Boris Cherny, the founder of Claude Code, predicts that the title of software engineer will “disappear” starting in 2026. He did not use any euphemisms, directly adding, “It’s going to be painful for a lot of people.” If even the internal engineers at Anthropic are not writing code, then what exactly are they “doing”?

The answer is: they are training the next generation of AI.

And the tool for training the next generation of AI is Claude. In other words, Claude is participating in shaping the next version of Claude using its own generated code. This loop sounds like a science fiction plot, but it is now part of Anthropic’s daily operations.

Dario explicitly states, “I’m not sure how much time is left.”

CEO Dario Amodei publicly admitted that he is no longer writing code himself. What is more noteworthy is what he said next: “This loop is closing rapidly, and I’m not sure how much time we have left.”

This statement can be interpreted in two ways. The optimistic version is: AI is accelerating evolution, and humanity is entering the next level of productivity. The pessimistic version is: those present, including Dario himself, are uncertain where this acceleration will stop, or whether it will stop at all.

It is worth mentioning that Anthropic concurrently predicts that AI models will have the capability to replace software engineers within 6 to 12 months. This time window is from mid-2026 to early 2027. That’s not far from now.

The logic of the feedback loop and overlooked issues

The feedback loop of “AI training AI” is not technically new. Synthetic data, RLHF, and model distillation all have similar self-referential structures. However, Anthropic’s case has a subtle difference: it is not about using AI to generate training data but allowing AI to directly participate in engineering decisions—determining which features to implement, how to implement them, and how to fix bugs.

In this process, the role of human engineers has devolved from “decision-makers” to “reviewers,” and the standard for review is still human evaluation. The problem is, when the output speed and complexity of AI far exceed human review capabilities, how much substantive meaning remains in this “taking a glance” gatekeeping?

Currently, no one has provided an answer. Dario says he is uncertain how much time is left, Boris says it will be painful, and Anthropic’s figures suggest 70% to 90%. These statements are precise but do not explain “what comes next.”