A four-month-old startup has raised $650 million for AI that improves itself

A startup has just closed a $650 million funding round to build self-evolving AI: a system that improves itself in an accelerating feedback loop. For the first time, a speculative idea from 1960s research papers is attracting serious investment and moving from theory into real engineering projects.

From theory to practice

The concept of "recursive superintelligence" has been floating around computer science and science fiction for more than six decades. The idea is simple. An AI system analyzes its own code, finds ways to become more powerful or faster, and applies those improvements. It then immediately uses the improved version of itself to hunt for the next round of optimizations. If this works, the result is a positive feedback loop in which each development cycle speeds up the next one.

Until now, this was a convenient metaphor for debates about AI's distant future. Now major investors are willing to pay billions to find out whether it works in practice.

How it's supposed to work

In the ideal scenario, the system works like this. First, human researchers write a baseline self-learning algorithm. The system then treats itself as an object of analysis and looks for ways to improve its own architecture, computational speed, or reasoning logic. Every improvement is logged, tested, and evaluated. If the result is better, the change sticks. The system reaches a new level of performance and starts the whole process over, this time with a more powerful mind. In theory, the cycle accelerates exponentially.

• The system analyzes its own source code
• Identifies bottlenecks and inefficiencies
• Designs and implements improvements
• Tests the results in a controlled environment
• Repeats the process with the improved version

Why this raises concerns

The $650 million round marks a turning point: the shift from philosophical debate to engineering reality. But it also sharpens long-standing worries among AI safety researchers. If a self-evolving system becomes truly powerful and slips out of control, the consequences will be fundamentally different from a bug in an ordinary app.

The questions are already here. How do you make sure the system stays controllable at every stage of self-improvement? How do you prevent an "escape" once the system becomes smarter than its creators? That is why part of the funding is going not only to development but also to safety research.

What it means

We are on the edge of a new chapter in AI history. Startups are no longer waiting for approval from academic institutions. They are investing in self-evolving systems, which will turn out to be either a revolution in how we understand automation and progress, or an intense test of our ability to stay in control of the increasingly powerful systems we build.