iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
AI is writing most of Google DeepMind’s production code, but the company reports no measurable acceleration of AI research caused by AI itself. A report published by The Decoder on October 14, 2026 describes substantial automation in routine, testable work alongside persistent human review, falling reliability on harder tasks and a capability index that is improving steadily rather than suddenly. That is evidence of rapid progress—not of an intelligence explosion or fully recursive self-improvement.
What Google DeepMind’s evidence actually shows
The reported assessment comes from Google DeepMind engineer Haydn Belfield, as described by The Decoder. DeepMind has seen no measured acceleration in AI research that can be attributed to AI improving AI research itself. The distinction matters: an AI system can make engineers more productive without independently discovering, validating and deploying better generations of AI at an accelerating rate.
The figures are reported second-hand rather than in a clearly identified primary DeepMind publication. They should therefore be read as a report of Google’s internal evidence, not as independently reproduced measurements. No verbatim Belfield quotation is established in the available material.
Free tools Windows power users keep installed
One-click scans. No signup required.
AI writes most production code, but humans still control the changes
Google DeepMind reportedly says AI produces more than 75% of its production code. That number measures generated code, not code that an agent can safely ship without supervision. About 88% of code changes still require human oversight or review, and every change is reviewed by a person before submission.
#1 Best Overall
The portion of changes fully delegated to AI increased from 7.5% to 11.9%. That is a meaningful shift toward automation, but it also means the large majority of changes are not fully delegated. Code generation is especially favorable to automation because requirements can often be stated precisely and outputs can be checked with tests, type systems, static analysis and review.
| Measure | Reported result | How to interpret it |
|---|---|---|
| Production code written by AI | More than 75% | High generation volume in an engineering environment |
| Code changes needing human oversight or review | About 88% | Most reported changes still retain human control |
| Fully delegated changes | 7.5% rising to 11.9% | Delegation is growing but remains a minority |
Why coding volume is not recursive self-improvement
Recursive self-improvement would require more than producing code. An AI system would need to identify valuable research directions, design novel experiments, diagnose failures, implement changes, evaluate them reliably and repeat the cycle with little or no human intervention. A large amount of machine-written code can support that loop, but it does not demonstrate that the loop is operating autonomously or accelerating.
Human review can catch security flaws, incorrect assumptions, regressions and misleading benchmark gains. It also supplies decisions that are difficult to reduce to a pass-or-fail test: which problem to pursue, what evidence is convincing and whether an apparent improvement will generalize.
Rank #2
Agent reliability falls as research becomes harder
Across 1,389 internal Argon agent runs, success without human input declined sharply as task complexity increased. A survey of 12 senior DeepMind researchers found the highest automation potential in well-defined, verifiable tasks and the lowest in open-ended scientific ideation.
Tasks where agents perform best
- Work with a precise specification.
- Problems whose outputs can be checked objectively.
- Engineering tasks with repeatable tests or measurable targets.
Tasks that still require researchers
- Creative research and choosing promising questions.
- Designing genuinely novel experiments.
- Diagnosing subtle training instability.
- Long-term planning across uncertain results.
This pattern is central to the self-improvement question. An agent that succeeds on bounded implementation work is not equivalent to an agent that can run an entire research program, decide what to try next and recognize when its own conclusions are wrong.
RE-Bench shows progress, not dependable autonomy
On RE-Bench, described as 74 realistic machine-learning engineering tasks, Argon solved about half of the tasks. Gemini 3.7 Flash solved 27%. Argon’s result is a substantial improvement over the comparison model, but a roughly 50% success rate is far below the near-perfect reliability needed for unsupervised research automation. In a research loop, one serious failure can invalidate an experiment, waste substantial compute or send later decisions in the wrong direction.
| System | RE-Bench result | Scope |
|---|---|---|
| Argon | About 50% | 74 machine-learning engineering tasks |
| Gemini 3.7 Flash | 27% | Same 74-task comparison |
Benchmark gains should therefore be separated from deployment claims. A higher score indicates stronger performance on the measured tasks; it does not establish that an agent can safely operate without review across open-ended AI research.
Google’s capability index shows steady growth
Google’s General Capability Index (GCI) covers 123 models and 92 benchmarks. The reported trend is approximately 0.1 GCI points of growth per month, with no statistically significant change in the rate of that growth. In other words, capabilities are improving, but the measured slope has not suddenly steepened.
Argon sits slightly above the trend line. The report attributes that position to additional compute and organizational investment rather than to AI-driven acceleration. That distinction weakens the claim that systems are independently creating a faster improvement cycle.
How to read the result alongside intelligence-explosion warnings
Warnings about rapid recursive self-improvement are scenario forecasts; the DeepMind evidence described here is an empirical measurement of current systems. They address different questions and should not be treated as direct contradictions.
| Question | What the reported DeepMind data addresses | What it does not establish |
|---|---|---|
| What is being measured? | Current agent runs, coding delegation and benchmark performance | A guaranteed future trajectory |
| Which work is easiest to automate? | Routine, well-defined and verifiable tasks | Open-ended scientific discovery |
| What does code volume mean? | AI contributes heavily to implementation | AI independently accepts and deploys better AI systems |
| What does benchmark success mean? | Argon outperforms the cited comparison model on RE-Bench | Reliability sufficient for unsupervised research |
| What does the GCI trend mean? | Capability growth is steady at about 0.1 points per month | A statistically significant acceleration |
Answers to the practical questions
Is AI close to improving itself?
Based on the reported DeepMind measurements, no. AI is improving engineering productivity and can contribute to parts of the development process, but there is no measured AI-driven acceleration of AI research.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Scan for outdated or missing drivers - takes under a minute3Clear out junk files and repair common Windows errorsCan AI do its own AI research yet?
It can complete portions of machine-learning engineering work, especially when tasks are specified and verifiable. The reported success rates and researcher survey do not support fully autonomous, open-ended AI research.
Best Value
Does AI-written code mean recursive self-improvement has started?
No. Generated code is an input to software development. Recursive self-improvement would require a reliable, largely autonomous research-and-validation loop, which these figures do not demonstrate.
How much human review do coding agents still need?
Google DeepMind’s reported figures say about 88% of code changes still need human oversight or review, and every change is reviewed by a human before submission.
Bottom line
Google DeepMind’s reported data describes powerful but bounded automation. AI can now generate most production code in the measured environment, yet humans remain attached to most changes, agent reliability drops on complex work and the GCI shows no statistically significant acceleration. The evidence supports “rapid capability growth with human supervision,” not “AI is already improving itself in a runaway loop.”
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.
