Brin says Google is entering an AI self-improvement cycle

Brin said Google was starting to enter recursive self-improvement, or RSI, which refers to AI helping build better AI systems.[1][2] He made the comments in a roundtable about the creation of Google’s Gemini 4 model, Argon, with Google’s Josh Woodward, Tulsee Doshi and Melvin Johnson.[2][1] Google shared the roundtable video on a Wednesday.[2] Google released Argon publicly during the week of TNW’s report, which was published on October 17, 2026.[1]

Brin pointed to Google’s Flash model releases before Gemini 4 as evidence that AI was helping develop successor models.[2] He also said Google would be a couple of models ahead internally by the time viewers watched the video.[1] The idea behind that progression is a flywheel that could speed up the pace of building new models.[2]

Google DeepMind’s Argon safety report said AI agents author a majority of production code and bug fixes, with humans reviewing changes before submission.[1] More than 75% of Google DeepMind’s production code is now AI-written, according to the report.[1] Yet about 88% of pull requests still require repeated human steering or code written by a person.[1] The share of fully delegated pull requests rose from 7.5% in April to 11.9% in September.[1] The report also said humans remain key to choosing research directions, designing experiments and diagnosing training problems.[1]

Argon solved around half of the challenges in Google’s internal test of AI research tasks.[1] Google tracks Gemini’s improvement with an internal index based on 92 benchmarks.[1] The index’s monthly increase rose from 0.091 points before April 1, 2026, to 0.118 points after that date, but the report said the difference was not statistically significant.[1] Google suspected increased investment and training compute, rather than AI-driven acceleration, might explain the change.[1] The report rated current risk from AI-driven research and development as low and surveyed 12 senior leaders and researchers.[1]