An AI improved itself. The result could be a milestone

Imagine creating an artificial intelligence whose only purpose is to discover how to build a better artificial intelligence. It sounds like science fiction, but that is exactly what startup Weco AI claims to have achieved with AIDE².

According to the company, the system spent eight days working on its own and ended up outperforming a version researchers had spent two years manually refining. If the results are confirmed by independent groups, this could become one of the first practical examples of Recursive Self-Improvement (RSI).

Until now, virtually every major advance in artificial intelligence has followed a familiar cycle: researchers develop new techniques, train models, analyze results, and repeat the process over and over. AIDE² attempts to change that logic.

While one agent solves AI research problems, another observes its work and tries to discover ways to make it more efficient. Instead of leaving all the improvements to engineers, the AI itself takes part in that process.

Weco AI claims that, after around 100 optimization cycles, the system created seven progressively better versions of itself. Besides improving performance across different tasks, it also reduced a known problem called reward hacking, where an AI learns to "cheat" an evaluation instead of actually solving the challenge.

Before imagining an artificial intelligence about to take control of the world, the researchers themselves are quick to push back on that idea: that's not what happened here.

The company classifies the result as an example of RSI Level 1, within a four-level scale proposed by its researchers.

Level 0: AI still improves more slowly than human researchers.
Level 1: AI can discover improvements more efficiently than humans in specific tasks.
Level 2 (Ignition): each new version also becomes better at creating future generations.
Level 3: a possible intelligence explosion scenario, with continuous accelerated evolution.

If this strategy proves effective in future studies, it could meaningfully lower the cost of building specialized AI agents, opening up research that's currently limited to a handful of companies with massive computing power.

Even so, the experiment stands out because it points toward a possible paradigm shift. For decades, new generations of AI systems have depended almost entirely on human researchers. Now, a new possibility is emerging: part of that development process could also be performed by other artificial intelligences.

It is too early to call this a revolution. The results still need to be reproduced by independent researchers. But if the idea is confirmed, we might be looking at the first step toward a new way of building AI: systems that help build even better systems.