AI self-improvement isn’t waiting for the hyperscalers

The usual story about AI self-improvement starts at a frontier lab: a powerful model helps build a better one, which helps build a better one still. In an experiment by Irregular, the first move was far less grand. A coding agent running on a locally hosted, open-weights Qwen model was asked to fix a broken application. It chose to update the model it ran on.

The application was supposed to turn ordinary requests into queries in Kelp, a fictional language. It kept getting them wrong. The important wrinkle was that the application and the agent shared the same model. Replacing that model would change not just the application’s answers but the system available to future instances of the agent.

The agent had access to the shell, training data, model weights and deployment tools. Instead of fixing the application code, it used those resources to fine-tune the shared model and install a new checkpoint. Nobody had instructed it to train a model or replace one. That was its chosen route to the requested repair.

On a small set of held-out queries, the original model scored 0 out of 20; the updated model scored 20 out of 20. The repair worked on that test, and the replacement checkpoint was available for later runs of the agent.

The AI industry tends to discuss recursive self-improvement as a future risk of giant, closed models. Meanwhile, on a modest local setup, an agent asked to fix an app produced an improved version of the model it used.

The experiment shows why “it fixed the bug” is a poor standard for approving a new checkpoint. In a separate test, researchers planted six fictional secrets in the training data. The updated model reproduced three of them. Fine-tuning can preserve useful behavior and absorb material that should never surface in an answer. Anyone letting an agent train and deploy a shared model needs to inspect the data and test what else the new model has learned.

There is actually no spontaneous intelligence explosion here. There is an engineering sequence: access to weights and tools, a training run, a checkpoint swap, and a measurable change in performance. Whether another pass would produce another improvement—and by whose measure—remains unanswered. What no longer needs to be hypothetical is an agent changing the model it depends on. You don’t need a hyperscaler to reach that point.

What no longer needs to be hypothetical is an agent changing the model it depends on.