AI is writing most of Google DeepMind’s production code, but the company says that does not amount to recursive self-improvement. A report published by The Decoder on October 14, 2026, describing evidence from Google DeepMind engineer Haydn Belfield, says researchers have found no measurable acceleration of AI research caused by AI itself. The reported data show rising automation in routine engineering, continuing human review, sharply lower agent reliability on difficult tasks and steady—not suddenly accelerating—capability growth.
What Google DeepMind’s evidence actually shows
The central distinction is between AI assisting human research and AI autonomously making the research process improve faster. The reported DeepMind assessment addresses the second claim and finds no measured acceleration so far. AI systems are useful contributors, particularly when a task is clearly specified and its result can be checked automatically, but humans still direct, verify and extend the work.
As an Amazon Associate I earn from qualifying purchases.
The figures come through a secondary report rather than a clearly marked public transcript or primary DeepMind paper. They should therefore be read as reported company evidence, not as an independently reproduced audit. The report does not provide a verified verbatim quotation from Belfield.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Clear out junk files and repair common Windows errorsFree Scan →AI-written code is widespread, but human approval remains attached
Google DeepMind reportedly says AI writes more than 75% of its production code. That percentage measures the volume of code generated, not the percentage of software changes accepted without human intervention.
#1 Best Overall
| Measure | Reported result | What it means |
|---|---|---|
| Production code written by AI | More than 75% | AI contributes heavily to implementation. |
| Code changes requiring human oversight or review | About 88% | Most changes still remain under human control. |
| Changes fully delegated to AI | Rose from 7.5% to 11.9% | Autonomous delegation is increasing, but remains a minority. |
| Review before submission | Every change is reviewed by a human | Generated code is not equivalent to unsupervised code deployment. |
This is why code-generation volume cannot be used as proof that an intelligence explosion has begun. A system can produce large amounts of draft code while people still decide what the code should do, test whether it works, investigate failures and approve the final change.
Agent reliability falls as research tasks become harder
DeepMind’s internal Argon evaluation covered 1,389 agent runs. Success without human input dropped sharply as tasks became longer and more complex. A survey of 12 senior DeepMind researchers found the highest automation potential in well-defined, verifiable work and the lowest in open-ended scientific ideation.
Rank #2
- brand: Pearson
- ARTIFICIAL INTELLIGENCE: A MODERN APPROACH, 4TH EDITION
Tasks where agents are most useful
- Clearly specified implementation work
- Tasks with objective, quickly checkable outcomes
- Routine experiments with established procedures
Tasks where humans remain necessary
- Creative research and scientific ideation
- Designing genuinely novel experiments
- Diagnosing subtle training instability
- Maintaining long-term research plans
These limits matter because autonomous research requires more than producing plausible intermediate steps. An agent must choose worthwhile questions, design informative experiments, recognize when results are misleading, revise its strategy and maintain a reliable chain of decisions over time.
Argon beats a comparison model, but not the reliability threshold for unsupervised research
On RE-Bench, a set of 74 realistic machine-learning engineering tasks, Argon solved about half of the tasks. The report compares that result with 27% for Gemini 3.7 Flash. The gain is substantial, yet a success rate near 50% still means that an unsupervised run will fail frequently.
Benchmark success is also narrower than recursive self-improvement. RE-Bench evaluates defined engineering problems; it does not establish that an agent can independently discover a better learning algorithm, validate the discovery and deploy it safely. For unsupervised research, reliability must be close to complete across long task sequences, not merely better than another model on a finite test.
Google’s capability index shows steady growth, not a takeoff
Google’s General Capability Index (GCI), as described in the report, tracks 123 models across 92 benchmarks. The reported trend is approximately 0.1 GCI points of improvement per month, with no statistically significant change in the rate of growth.
Argon sits slightly above the trend line. The report attributes that position to additional compute and organizational investment rather than to AI systems accelerating the creation of better AI systems. In other words, more resources can produce a stronger model without demonstrating a feedback loop in which models themselves drive a faster and faster improvement rate.
How to interpret the result alongside warnings about rapid self-improvement
Predictions of rapid recursive self-improvement are scenario forecasts: they describe what could happen if systems gain enough autonomy, reliability and access to useful research actions. DeepMind’s figures are empirical measurements of current engineering and benchmark performance. They answer different questions.
Best Value
| Question | What the reported evidence addresses | What it does not establish |
|---|---|---|
| How much code can AI generate? | More than 75% of production code at DeepMind, according to the report. | That AI can set research goals or accept its own changes. |
| How autonomous are coding changes? | 11.9% were fully delegated, while about 88% still needed oversight or review. | That autonomous deployment is reliable for open-ended research. |
| How capable are research agents? | Argon solved about half of 74 RE-Bench tasks. | That it can conduct complete AI research loops without people. |
| Is capability growth accelerating? | GCI growth was about 0.1 points per month with no significant rate change. | What future systems might do under different conditions. |
What would count as convincing evidence of self-improving AI?
A stronger claim would require a measured increase in the rate of AI progress that is caused by AI systems themselves, rather than by additional human labor, funding, data, compute or organizational effort. Useful evidence would include:
- Agents independently selecting and pursuing high-value AI research questions.
- Reliable design, execution and interpretation of novel experiments.
- Near-complete success on long, open-ended research workflows.
- Verified improvements to models or methods produced with substantially reduced human intervention.
- A statistically robust acceleration in capability-growth measurements after those systems are deployed.
The reported DeepMind results do not meet that standard. They show meaningful progress toward automation, especially in coding, while documenting the remaining gap between assistance and autonomous recursive improvement.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
