AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Google says self-improving AI isn't anywhere close

Collected Oct 14, 2026

Haydn Belfield, an engineer at Google DeepMind, has published an assessment arguing that self-improving AI remains far off, and that the company has seen no measurable acceleration of AI research driven by AI itself. The claim rests on internal tooling the team built to look for that acceleration directly rather than infer it from capability headlines.

Automation inside Google's engineering pipeline is already substantial. According to Belfield, AI writes more than 75 percent of production code. Human involvement has not disappeared alongside it: roughly 88 percent of code changes still require human oversight or review, and every change is reviewed by a person before submission. The share of code changes fully delegated to AI, with no human in the loop at that stage, rose from 7.5 percent to 11.9 percent.

The pattern gets sharper as work gets harder. Across 1,389 internal runs of an agent called Argon, the success rate without human input fell sharply as task complexity increased. Belfield identifies the areas where people remain essential: creative research, designing novel experiments, diagnosing subtle training instabilities, and long-term planning. A survey of twelve senior DeepMind researchers matched that division of labor, with automation highest on well-defined, verifiable tasks and lowest on open-ended scientific ideation.

On RE-Bench, an in-house benchmark containing 74 realistic machine learning engineering tasks, Argon solves about half of them. That is a large improvement over Gemini 3.7 Flash at 27 percent, but still well short of the near-perfect performance that full automation would demand. This suggests the gap is not a matter of a few more model generations on easy tasks; the remaining half of the benchmark is where the difficulty concentrates.

To track progress systematically, the team built a General Capability Index covering 123 models and 92 benchmarks. Gemini capabilities are growing at a steady pace of roughly 0.1 GCI points per month, with no significant change in that speed. Gemini 4 Argon sits slightly above the trend line, which the team attributes to more compute and organizational investment rather than AI-driven acceleration. The team currently rates the risk from AI-driven research automation as low, but is building measurement infrastructure to watch for signs of acceleration going forward.

Belfield's findings line up with some recent research, though the field is not unified. More than 20 AI researchers, among them Geoffrey Hinton, Yoshua Bengio and OpenAI's Jakub Pachocki, recently warned in a research paper of a possible "intelligence explosion," arguing that AI systems could automate all AI research and development within a few years and compress years of progress into months.

The disagreement is partly about measurement. A capability index like GCI tracks benchmark performance over time, which is a lagging signal: it can detect a speedup once tasks have already been solved, but a genuine feedback loop could first show up as faster internal iteration before it appears on any public or private eval. Argon's success rate on 74 engineering tasks and the 11.9 percent delegation figure are point-in-time measurements of a moving system, and both could shift without the headline GCI number changing. In practice, the practical distinction for engineering teams is between assistance that speeds up work a human still directs and autonomy that removes the human from the loop. Google's numbers describe the former at scale, with the latter confined to a small slice of changes.

Why it matters: teams planning around an imminent self-improvement takeoff now have a concrete counter-datapoint from inside one of the labs best positioned to observe it, while those planning around steady capability growth get support for that assumption. Developers should read the 75 percent code figure and the 11.9 percent full-delegation figure as describing different things: widespread assistance and narrow autonomy, with review still the default gate. If the GCI trend line is the right instrument, expect capability gains to keep arriving at roughly the same monthly rate rather than compounding, though that reading is contested by researchers who predict a much sharper curve.

Read at The Decoder

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Google engineer Haydn Belfield says there's no measurable acceleration of AI research driven by AI at Deepmind. AI now writes more than 75 percent of production code, but 88 percent of changes still need human oversight. The company's in-house General Capability Index shows steady growth with no sudden speedup. Self-improving AI isn't anywhere close, Belfield says. The article Google says self-improving AI isn't anywhere close appeared first on The Decoder .