Google's co-founder says the company is stepping onto the path of recursive self-improvement. Google's own safety report, released the same week, says it cannot find the speedup. Both claims come from the same building, which makes for an unusually candid kind of disagreement.
What Brin actually said
At a roundtable about Gemini 4 Argon, the flagship model Google opened to the public this week, Sergey Brin said the company is beginning to move toward recursive self-improvement. In his telling, Gemini models already help produce the next version of themselves, and the engineers and the model work hand in hand.
His evidence was the Flash models that shipped before Argon. He added that by the time the video of the conversation went online, Google would be a couple of models further along. Google posted the video last Wednesday, and a business publication picked it up. Brin, for the record, has worked on Gemini since returning to Google.
It is a bold claim, delivered with the relaxed confidence of someone who does not have to fill in the safety paperwork. That job belongs to DeepMind.
What the safety report says
Alongside the launch, Google DeepMind published a safety and security report on Argon. Its central finding is almost the opposite of a victory lap. It records no meaningful end-to-end acceleration of AI research, and it rates the risk of AI speeding up AI research as low for now.
The assessment rests partly on a survey of 12 senior leaders and specialists. That is a small sample, though it is presumably the group best placed to know what is happening inside the lab.
The numbers behind the caution
The report is not claiming that AI does little. It is claiming that doing a lot of typing is not the same as speeding up research.
| Metric | Figure |
|---|---|
| Share of DeepMind working code written by AI | More than 75% |
| Pull requests fully delegated to an agent, April | 7.5% |
| Pull requests fully delegated to an agent, September | 11.9% |
| Changes that still need human edits | About 88% |
| Tasks in DeepMind's internal test that Argon solves | About half |
Agents already write most of the code and fix a large share of bugs. A human still reviews every change before release, and humans remain the ones who choose research directions, design experiments and work out why a training run went wrong.
So the machine does most of the keystrokes while people keep the judgment. That is a real change in how the work gets done. It is also a long way from a lab running itself.
The improvement index, and why it settles nothing
DeepMind tracks an internal index of Gemini improvement, built from 92 tests. Since 1 April 2026 it has risen faster than before, at 0.118 points a month against 0.091 earlier. The report says the difference is not statistically significant.
Argon itself sits above the trend line. The company's explanation is more investment and more compute, not AI accelerating AI. That is a respectable account, and also a convenient one, since it is hard for an outsider to distinguish from the alternative.
The flywheel view
Vihan Jain, who leads the RSI and AI for science effort at Google DeepMind, wrote on X that the model is built into its own evolution. His example was telling. When the team responsible for post-training stops asking for raw chips and starts asking for a quota on Argon, the RSI flywheel shifts into a new gear.
That is a more concrete picture than Brin's, and a more modest one. It describes a change in how internal teams request resources, not a measured jump in research speed. Whether that shift matters depends on what the quota buys, and the report's answer so far is that it has not yet shown up in the output.
Outside estimates and the political clock
A report published this week by Gladstone AI surveyed specialists at Anthropic, OpenAI and Google DeepMind. They expect full recursive self-improvement to begin in 3 to 12 months. They also say it could come sooner if government intervention does not slow the process.
Policy is already moving. Last month, US Representative Ro Khanna announced a bill to ban self-improving AI. OpenAI, for its part, asked the United States to lead the development of global standards in the area.
How to read the gap
There are three ways to square Brin with his own company's report.
- Different definitions. Brin may be describing models contributing to their successors in a loose sense, while the report measures acceleration of the whole research pipeline.
- Different timing. The report looks at data up to September. Brin says the company is already a couple of models past Argon.
- Different audiences. A roundtable rewards ambition, and a safety report rewards defensible numbers.
The source material does not say which explanation is right, and it would be unwise to pick one on its behalf. What it does show is that the lab's most senior voice and its most formal document are describing the same system in very different registers.
What to watch
The measurable signals are fairly plain. If the share of agent-only pull requests keeps climbing from 11.9%, if the 88% needing human edits shrinks, and if the improvement index pulls away from its old trend by a margin that is statistically significant, the report's low risk rating will need revisiting. Until then, Google has told the public both that the flywheel is turning and that nobody can yet measure it spinning faster.