Autonomous software now manages portions of the experimentation and analysis phases within Google's model development cycle, allowing for the independent formulation of hypotheses. Chief Koray Kavukcuoglu stated that the company trusts these agents to run parts of the training process under human supervision, a shift in operations that began within the last six months. This change is intended to move the company closer to the goal of models that can train themselves.

The firm is preparing to release its flagship Gemini 4 model well before the end of the year, moving from post-training to an early public version to facilitate rapid iteration. This strategy is designed to catch up with OpenAI and Anthropic and is informing the hardware roadmap for the next two or three generations of TPUs. By giving chip engineers visibility into model needs, Google intends to compete more aggressively with Nvidia as it begins selling its TPU chips directly.