What Exactly Improves When an AI Improves Itself?

“The AI improved itself” can describe several very different events. A training process changed model weights. A search procedure found a better neural architecture. An agent edited the software surrounding its model. A system generated a faster algorithm for an external problem.

All are interesting. Only some change the model itself, and none of those descriptions alone establishes an open-ended path to general intelligence.

As of 23 June 2025, several research projects make the distinctions worth examining carefully.

Name the object being changed

Weights are the learned numerical parameters used by a model. Updating them changes the model through a training or adaptation process.

Architecture concerns the structure of the model: which operations and connections it uses. Neural architecture search explores candidate structures under an objective and constraints. It still requires a way to train or assess the candidates.

Agent software includes prompts, tools, control flow, memory, and other code around a model. Improving this software can make the overall system more capable while leaving the underlying model weights fixed.

A generated program may solve a completely external task. Better code produced by an AI is not automatically evidence that the AI’s own reasoning mechanism has improved.

These objects can interact, but merging them into one phrase hides where the progress occurred.

Three projects, three scopes

EvoPrompting, presented at NeurIPS 2023, uses code-capable language models in an evolutionary approach to neural architecture search. The work is by Angelica Chen, David Dohan, and David So, associated with Google DeepMind—not an Anthropic project. Candidate architectures and their evaluation are central to the method.

AlphaEvolve, announced by Google DeepMind in May 2025, combines Gemini models with an evolutionary process and automated evaluators to improve code for algorithmic problems. An evaluator provides feedback about proposed programs. Calling it a Gemini-powered system is different from saying Gemini autonomously rewrites and retrains its own weights.

The Darwin Gödel Machine, announced by Sakana AI in collaboration with researchers at the University of British Columbia in May 2025, explores agents modifying their own code and evaluating resulting variants on coding tasks. The object of improvement is the agent implementation. This is a concrete form of self-modification, with empirical evaluation, rather than a proof that every modification is beneficial or safe.

The projects differ in methods and targets. Grouping them is useful only if we retain those differences.

The evaluator defines the direction

A generate–test–select loop can improve whatever its tests reward. If the tests are narrow, the selected system may become better at satisfying those tests without becoming more useful elsewhere.

Ask what prevents that outcome. Are there held-out tasks? Are resource costs included? Can a candidate exploit the evaluator, skip required work, or alter the conditions under which success is measured? Does the improvement persist across repeated runs and different problems?

These are not peripheral concerns. Evaluation is the mechanism that turns variation into directed search. An unreliable evaluator directs the search toward unreliable success.

Improvement still needs an operating boundary

An experimental agent that can edit its code should run with appropriate isolation, action limits, recorded changes, and a recoverable starting point. A promising benchmark result does not justify unrestricted access to credentials or production systems.

The most informative claim is specific: this procedure improved this system, on these tasks, under these constraints, at this cost. That leaves room to ask whether the result transfers.

“Self-improving AI” becomes more interesting once it stops being a shortcut to an AGI prediction. It becomes a research question about what can be searched, what can be measured, and whether the resulting improvement survives outside the search loop.