Overview
- Google CEO Sundar Pichai said Friday that researchers now use AI agents alongside humans in parts of model development, creating hybrid loops rather than fully autonomous systems.
- Pichai reported that AI now deeply influences nearly all Google coding work and credited tools such as Gemini 4 Argon with a roughly one-third drop in code rollbacks and clearer developer productivity gains.
- Pichai warned that the “end state” of recursive self‑improvement, where models redesign stronger successors without human input, remains distant and will likely evolve through 2027.
- Other signals differ: Google cofounder Sergey Brin described recent Flash models and Gemini 4 as showing AI helping to build itself, and a Gladstone AI report said many interviewed researchers thought RSI could appear within three to 12 months though that forecast is contested.
- The shift matters for work and governance because agent-assisted research speeds iteration, concentrates compute and tooling in a few labs, and could accelerate capabilities if oversight, access to GPUs and safety practices do not keep pace.