Seventy-five percent of new code written at Google was being generated by AI, the company said in April. Sundar Pichai now says that number is meaningless. “That number in our world today is like meaningless in the sense that every part of code we now touch is influenced by AI deeply,” he said on the Big Technology Podcast this week, as Hugh Langley reported for Business Insider. The retraction matters more than the figure: if Google’s own 75% metric no longer measures anything, then the more interesting line Pichai drew in the same interview deserves scrutiny by the same standard.
That line is recursive self-improvement, or RSI: AI systems participating in the construction of their successors. On whether Google has reached it, Pichai was explicit. “We are pretty far from the end state of RSI,” he said. “The right way to describe it is, you have RSI loops along with AI researchers in various aspects of model development. I would say you almost have agents helping AI researchers, in some automated way, work together to build these models.” He expects that continuum to keep progressing into next year.
His phrasing is softer than that of Google’s other public voice on this. Earlier this month, cofounder Sergey Brin said the Flash model series and Gemini 4 showed AI helping to build itself, and that Google was “starting to get onto that RSI treadmill.” “And now our researchers and the model work hand in hand to make the next version of the model,” Brin said. Brin has been taking an increasingly active role in the Gemini effort, using his position as cofounder to shift resources toward projects he judges most effective — including, specifically, more focus on RSI, Business Insider has previously reported.
The forecast everyone is quoting
An outside estimate for the transition arrived this month from the firm Gladstone AI. Most researchers interviewed for its report, from companies including Anthropic and DeepMind, believed RSI would start within the next three to 12 months, and possibly sooner absent government intervention. Gladstone shared the report with TIME before publication; the firm said it worked with researchers at Anthropic, OpenAI and Google DeepMind, all quoted anonymously.
A three-to-12-month window and Pichai’s “pretty far” are not necessarily in conflict, because neither the Gladstone researchers nor Google has published a measurable threshold for when RSI has started. Google already runs systems in which agents write model code and run experiments beside human researchers; whether that counts as RSI or its precursor is a matter of definition, not measurement. The claim that would settle it needs a denominator: the share of a new model’s total development work performed without a human in the loop. No lab has published that number.
What is measurable
On the coding side, Pichai did offer one hard number. A code rollback — software reverted to a previous version after a bug or system fault — now happens one-third less often at Google, he said, and he attributed it to AI: “And so I think people are writing code more productively, and better code with it.” The baseline rollback rate and the measurement period are not public, so the figure cannot be converted into a per-release failure rate from outside. With Gemini 4 Argon — the frontier model Google put on general sale this month at $19.99 a month — Pichai said users would take on harder problems.
Pichai is a user of the stack as well as its spokesman: on the Sources podcast this week, he said he uses AI to track Google’s model development and internal discussions. The number that would answer the RSI question empirically has not yet been quoted by anyone — what fraction of the work that produced Gemini 4 was done by a model, and whether the people building its successor report a higher one.