Pichai称AI距离自主递归自我改进(RSI)仍远，Google靠人机协作开发下一代Gemini

Alphabet Inc. (NASDAQ:GOOGL) CEO Sundar Pichai says artificial intelligence is still far from independently building more powerful versions of itself, even as Google increasingly relies on AI to develop its next models.
“We’re pretty far from … the end state of RSI,” Pichai told Alex Kantrowitz on the Big Technology Podcast Tuesday.
Fully autonomous RSI, or recursive self-improvement, would allow AI to develop increasingly capable successors without human researchers directing the process.
If achieved, it could dramatically accelerate scientific breakthroughs, but some researchers fear increasingly powerful AI systems could advance faster than humans can understand or control them.
Days before Pichai’s remarks, Google co-founder Sergey Brin said the company was “starting to get onto that RSI treadmill,” with researchers and AI models collaborating on new Gemini versions.
In a separate interview with Alex Heath, published Wednesday, Pichai explained what that collaboration involves.
“There’s a lot of humans in the loop on these things,” he said. AI models already suggest training methods and evaluate results, and Pichai expects further acceleration from self-improvement over the next 12 months.
Google’s Own Research Reveals the Gap
Google DeepMind’s Oct. 13 safety report shows how much work AI already performs, and how much remains dependent on human researchers.
AI writes more than 75% of DeepMind’s production code. But only 11.9% of code-change submissions were fully delegated to AI in September, up from 7.5% in April.
Despite that automation, Google’s researchers found “there is currently no significant end-to-end acceleration of AI R&D, AI-driven or otherwise.”
They tracked model progress across 92 benchmarks and found that an apparent increase in the pace of development was not statistically significant.
Google’s assessment found AI most useful for clearly defined tasks, while human researchers still lead decisions about which scientific ideas to investigate, how to design experiments and how to resolve unexpected training problems.
AI is already producing improvements. Google’s AlphaEvolve system, for example, discovered an optimization that reduced Gemini training time by approximately 1%.
The Race to Automate AI Research
OpenAI said in September it had achieved its automated research-intern milestone and is targeting a more capable automated researcher by March 2028, still under human supervision.
Anthropic reports that Claude writes more than 80% of code merged into its codebase, but says people remain important in choosing worthwhile research problems.
On Polymarket, traders give Anthropic a 49% chance of having the highest-ranked AI model at year-end, compared with 43% for Google. The $2.77 million market measures model rankings, not autonomous self-improvement.
Pichai has already acknowledged that Google is sacrificing Cloud revenue to prioritize AI research, underscoring the commercial stakes in developing models capable of performing more of that work themselves.
Image: Shutterstock
© 2026 Benzinga.com. Benzinga does not provide investment advice. All rights reserved.
To add Benzinga News as your preferred source on Google, click here.