Artificial Intelligence
Google says self-improving AI remains far off
Haydn Belfield said Google DeepMind had not seen measurable AI-driven acceleration in research, even as automation expanded in coding and internal benchmarks showed limits on harder tasks.
- Haydn Belfield said Google DeepMind had not seen measurable acceleration in AI research caused by AI systems.
- He said AI wrote more than 75% of production code, while about 88% of code changes still needed human oversight or review.
- The share of code changes fully delegated to AI rose from 7.5% to 11.9%.
- Across 1,389 internal Argon agent runs, success without human input fell as task complexity increased.
- More than 20 researchers, including Geoffrey Hinton, Yoshua Bengio and Jakub Pachocki, warned in a paper that AI could automate all AI research and development within a few years.
Haydn Belfield, a Google engineer, said Google DeepMind had not seen any measurable rise in AI research speed caused by AI systems. He said AI already wrote more than 75% of production code, but about 88% of code changes still needed human oversight or review.
Belfield said every change was reviewed by a human before submission, and the share of code changes fully handed to AI increased from 7.5% to 11.9%. He said that as tasks became more complex and time-consuming, AI was less likely to complete them without help.
Across 1,389 internal Argon agent runs, success without human input fell sharply as complexity rose. A survey of twelve senior DeepMind researchers found the highest automation on well-defined, checkable tasks and the lowest on open-ended scientific ideation.
On RE-Bench, which contained 74 machine-learning engineering tasks, Argon solved about half, compared with 27% for Gemini 3.7 Flash. Google's General Capability Index, which covered 123 models and 92 benchmarks, showed Gemini capabilities rising by about 0.1 GCI points a month with no clear change in pace.
The team said Gemini 4 Argon sat a little above that trend because of more compute and organisational investment, and it rated the risk from AI-driven research automation as low. More than 20 researchers, including Geoffrey Hinton, Yoshua Bengio and OpenAI's Jakub Pachocki, had recently argued in a paper that AI could automate all AI research and development within a few years.
Named in this story
People
- Haydn Belfield
- the Google engineer who presented the findings
- Geoffrey Hinton
- one of the researchers who warned of a possible intelligence explosion
- Yoshua Bengio
- one of the researchers who warned of a possible intelligence explosion
- Jakub Pachocki
- an OpenAI researcher named among the warning paper’s signatories
Companies
- employer of Haydn Belfield and source of the reported findings
- OpenAI
- the company identified as Jakub Pachocki’s employer
Organisations
- Google DeepMind
- the research lab whose internal data and benchmarks were cited
Products and systems
- Argon
- the internal agent whose runs were analysed
- RE-Bench
- the benchmark used to test machine-learning engineering tasks
- Gemini 3.7 Flash
- the model used as a comparison on RE-Bench
- General Capability Index
- the internal index tracking model capability over time
- Gemini 4 Argon
- the model said to sit slightly above the capability trend
How the source tells it
Measured and data-heavy, with a reassuring register built from internal metrics, trend lines and benchmark scores.
- reassurance steady-growth metrics and a low-risk assessment are used to dampen claims of near-term self-improvement