If you read one AI headline today about whether models are about to improve themselves, it is this one: a Google DeepMind researcher says the company sees no measurable acceleration of AI research driven by AI, even though AI already writes most of its code. The claim comes from a The Decoder report on October 14, 2026, which cites Haydn Belfield on X and a safety report. We could not open the underlying document, so this post lays out what is claimed, what we could and could not confirm, and how to check for yourself.
Claimed vs verified at a glance
| Claim (as reported by The Decoder) | Status on explainx.ai |
|---|---|
| AI writes more than 75% of production code at Google | Reported; not found in Google pages we checked |
| About 88% of code changes still need human oversight or review | Reported; unconfirmed |
| Share of code changes fully delegated to AI rose from 7.5% to 11.9% | Reported; unconfirmed |
| 1,389 internal Argon agent runs: success without human input falls as tasks get complex | Reported; unconfirmed |
| Gemini 4 Argon solves about half of the 74 RE-Bench tasks vs 27% for Gemini 3.7 Flash | Reported; RE-Bench is a real public benchmark family, the DeepMind numbers are unconfirmed |
| A General Capability Index (123 models, 92 benchmarks) shows about 0.1 points per month, no acceleration | Reported; no public description found |
| DeepMind rates the risk from AI-driven research automation as low | Reported; unconfirmed |
| 20+ researchers including Hinton, Bengio and Pachocki warned of an intelligence explosion | Reported by The Decoder as a separate paper |
What the report says
Per The Decoder, Belfield, a Google DeepMind researcher, argues that self-improving AI remains far off even as significant automation takes hold. The numbers it attributes to him paint a specific picture: AI writes over 75% of production code, yet roughly 88% of changes still need human oversight, and every change is reviewed by a human before submission. The share of changes fully delegated to AI is described as rising from 7.5% to 11.9%.
The report also says that the more complex and time-consuming a task, the less likely an agent is to finish it alone. Across 1,389 internal runs of Argon agents, success without human input dropped sharply as complexity rose. A survey of twelve senior DeepMind researchers reportedly matched this: automation is highest for well-defined, verifiable tasks and lowest for open-ended scientific ideation, experiment design, diagnosing subtle training instabilities and long-term planning.
On DeepMind's RE-Bench-style evaluation of 74 realistic ML engineering tasks, Argon reportedly solves about half, up from 27% for Gemini 3.7 Flash. That is a big jump, but far from the near-perfect performance full automation would need. Finally, the General Capability Index reportedly shows Gemini capabilities growing steadily at about 0.1 points a month, with Argon sitting slightly above trend, which the team attributes to more compute and organizational investment rather than AI-driven acceleration.
What we could and could not verify
We checked what we could reach. Google's own Gemini 4 Argon safety post from October 13 covers misuse defenses, a misalignment monitor and agent-control work, but we found no mention of the 75%, 88% or 11.9% figures, the GCI, or research automation in it. The Decoder points instead to a separate safety report and to Belfield's X post; we could not locate either through our tools. Searches for the numbers turned up nothing independent.
That does not make the claim false. A researcher's post or a long report section can easily sit outside the public blog post. It means you should treat every percentage above as a secondhand report, which is exactly how we handle other unconfirmed numbers, for example in our claimed-vs-verified piece on Amodei and Meta's compute request. The Decoder introduces Belfield as a Google engineer and attributes the findings to him and to a safety report; we did not independently confirm his role.
Why this matters even unverified
The question of whether AI will speed up its own development is the crux of most timeline arguments. If AI research is bottlenecked by human judgment on novel ideas and debugging, the curve stays steady. If agents start doing that work, the curve bends. The reported DeepMind data point is a measurement, not an opinion, which is rare in this debate: a steady 0.1 points a month and no statistically significant acceleration, with the company saying it is building measurement infrastructure to catch the first sign of change.
It also fits other recent evidence. OpenAI's own account of 3.1 agent-workdays per human researcher shows heavy use of coding agents inside a lab, while Sakana's MASS work on a 27B model that trains itself through agent teams shows the research direction is real. None of these shows a runaway loop. They show a lot of delegation and a human still in the loop.
The timeline: how the self-improvement debate got here
The argument has moved from philosophy to instrumentation in about a year. Early talk of an intelligence explosion rested on thought experiments: if a model can do AI research, it can build a better model, which does better research. The practical version arrived with coding agents. Labs started publishing how many agent-hours their researchers use, and open projects such as Sakana's MASS showed models generating their own training signal. Meanwhile Google fast-tracked Gemini 4 after skipping an intermediate Pro release, and Argon launched this month with a safety post that stresses monitoring over capability claims.
Against that background, a lab reporting a flat trend line is notable precisely because it is the less exciting result. Labs have commercial reasons to emphasize rapid progress, and a measured "no acceleration so far" cuts against that, which is a mild point in its favor, though still not a substitute for seeing the data. Whether other labs publish comparable indices, with methods, is the thing to watch next.
How the claim sits with the intelligence-explosion warnings
The Decoder notes that more than 20 researchers, including Geoffrey Hinton, Yoshua Bengio and OpenAI's Jakub Pachocki, recently argued in a paper that AI could automate all AI research and development within a few years, compressing years of progress into months. Both things can be true: no acceleration today, and a risk that arrives once agents cross a capability threshold on open-ended research. The strongest reading of the Belfield claim is narrow: as of the data collected, human creativity is still the bottleneck, and the gap between 11.9% fully delegated changes and everything is large.
The honest unknowns are the ones a skeptic would ask. How was "production code" counted, by lines, by commits or by merged changes? Does 75% include autocomplete and boilerplate? Is the GCI weighted toward benchmarks that saturate smoothly, which would hide a sudden jump on a few hard tasks? Without the primary document none of these can be answered.
What builders should take from it
- Plan for review, not autonomy. Even in a lab with the strongest tooling, the reported figure is that most changes keep a human in the loop. If you run coding agents, a mandatory review step is not overhead, it is the norm. Our Argon safety report coverage shows the control stack Google pairs with that.
- Expect task length to dominate success. The reported drop-off with complexity matches what practitioners see: break work into verifiable pieces. See how Gemini 4 Argon compares with Opus 5.5, Grok 4.7 and GPT-6 Astra on agentic benchmarks.
- Track your own delegation rate. Count the share of changes your team merges with no human edits. It is the cleanest version of the metric DeepMind reportedly uses, and it will tell you more than any vendor figure.
- Secure the loop. Agents that write and run code need runtime guardrails; AgentBeam, from the explainx.ai team, stops AI agents before they take dangerous actions.
How to check this yourself
- Search for Haydn Belfield's thread on X and see whether it links a report with the 1,389-run and twelve-researcher survey details.
- Open the Gemini 4 Argon Safety and Security Report and model card linked from Google's announcement and look for sections on AI R&D or automation risk.
- Compare the RE-Bench numbers with the public benchmark definition before treating 27% versus about 50% as a like-for-like jump.
- Wait for independent replication: a lab-internal capability index is only as credible as its published method.
We will add an update here if the primary document surfaces. Until then, the right summary is: DeepMind is reported to see steady, not accelerating, progress, with humans still gating most code, and that report has not yet been independently confirmed.
Related reading
- Google DeepMind fast-tracks Gemini 4 after skipping Gemini 3.5 Pro
- Gemini 4 Argon launch, benchmarks and pricing
- Gemini 4 Argon safety report: probes, monitor and 2% cheating
- OpenAI's research acceleration post
- Sakana MASS: a model that trains itself through agent teams
- Argon vs Opus 5.5 vs Grok 4.7 vs GPT-6 Astra
Figures are as reported on October 14, 2026 by The Decoder and have not been confirmed against a primary Google document.