Particle.news

Pichai Says Google Is Far From Fully Self‑Improving AI

Google is shifting cloud compute to frontier model work, a choice that could alter how quickly AI capabilities and safety concerns evolve.

Overview

  • Sundar Pichai told podcast audiences on Friday that Google is still "pretty far from the end state of RSI," meaning the company does not view fully autonomous recursive self‑improvement as achieved.
  • Pichai said AI agents are already assisting researchers in parts of model development, creating semi‑automated loops where humans and models iterate on successor designs.
  • Google reports AI now deeply influences almost all of its code work and that use of tools like Gemini 4 Argon has cut code rollbacks by about one‑third and raised developer productivity.
  • Pichai warned the industry faces a shortage of frontier compute and said Google is deliberately reallocating cloud resources and forgoing short‑term revenue to reserve capacity for model research.
  • Outside analysts and a Gladstone AI report say many researchers expect RSI could begin within three to 12 months, a contested forecast that heightens questions about oversight, safety testing, and competitive advantage.