AI is learning to build itself: Google DeepMind and OpenAI on the path to self-improvement

In 1966, mathematician I. J. Good predicted that a superintelligent machine could build even smarter machines, triggering an "intelligence explosion." That science fiction is now becoming reality: OpenAI, Google DeepMind, and Anthropic already use AI to build more advanced versions of themselves.

AI starts writing itself

OpenAI recently said that GPT-5.3-Codex is actively involved in its own improvement. The model writes code for its own training, debugs the training process, manages deployment, and analyzes evaluation results. In effect, it helps make itself better.

Anthropic has gone further: Claude itself now writes most of the code for Claude (the company's flagship AI coding assistant). This creates a strange loop: AI improves itself, but humans stay in the loop, reviewing results and making decisions.

Google DeepMind unveiled AlphaEvolve, an agent for "scientific and algorithmic discovery." The system can optimize neural network architectures, improve task scheduling in data centers, and refine chip designs. Each breakthrough lets researchers move faster toward the next one.

"Often you look at what the system discovered and learn from that discovery," says Matej Balog of DeepMind. The system has already surprised its creators by finding optimization algorithms that had eluded human intuition.

Two of the creators of AlphaChip (a chip design system) have founded a startup called Recursive Intelligence. They promise to cut the chip development cycle from one to two years down to days. In a third phase, the company plans to use AI to build chips that are better at training the next generation of AI.

Why it isn't taking off yet

Despite the impressive progress, full AI independence from humans is still a long way off. Researchers note that current systems generate and implement ideas at a "good, but not great" level.

Here's what is holding the process back:

• Humans still set the goals and decide what counts as success
• Training frontier models costs billions, so no one will let such a system run unsupervised
• Knowledge at large corporations is spread across thousands of people (TSMC employs 90,000 people) and can't be packed into a single algorithm
• Full automation would require not just code and chips, but also building data centers and power plants and mining raw materials

Researchers at Meta propose a different path: instead of full automation, aim for human-machine "co-improvement." By staying in the development loop, humans ensure both the speed of progress and its safety.

Could this spiral out of control?

A survey of 25 leading AI experts found that 23 of them do not rule out an "intelligence explosion" if AI development were fully automated. Most believe companies would keep such models secret rather than release them publicly.

David Krueger of Montreal, who advocates a pause in AI development, painted a grim picture:

"It's a gamble with everyone's lives."

He proposes halting AI development once machines write 99% of the code. "It looks like we're approaching that point right now," he adds.

What it means

We are at the start of an era in which AI is genuinely beginning to accelerate its own development. That could lead to breakthroughs in medicine, science, and technology. But it also demands strict oversight, from international regulators and from the companies themselves. History shows that powerful technologies need limits.

*Meta has been designated an extremist organization and is banned in Russia.