What's this about?
Google states that both releases are powered by the same foundational intelligence, and "further accelerated by long-running agentic loops designed to recursively evaluate and refine the underlying models."
What is being described there?
Using agents to improve the model that runs the agents. The loop evaluates model outputs, identifies weaknesses, generates or selects training material addressing them, and feeds that back — running long enough to compound. It is recursive self-improvement in the narrow, practical sense: not a model rewriting its own weights, but a development process where the model is a substantial part of its own supervision. Every major lab now does some version of this, and this is an unusually plain statement of it in a product announcement.
Why is it worth noticing?
- It lands in a week when Zhipu announced Full Self-Training as a product direction and Anthropic published what happens when the supervision signal is exploitable.
- Three Flash releases in six weeks is the kind of cadence such a loop would produce.
- "Recursively evaluate and refine" is doing a lot of work in one sentence and is not elaborated anywhere in the post.