Google DeepMind finds no acceleration in AI research progress

Google DeepMind released its initially detailed internal evaluation of AI-driven device finding out research study on October 14, 2026. The audit discovered no quantifiable velocity in AI development that can be credited to AI itself.

news What the audit really determined

The evaluation appears inside the Gemini 4 Argon Safety & & Security Report. Scientist Haydn Belfield led the work.

The group tracked frontier Gemini abilities with a metric it calls the General Capability Index, or GCI. Throughout 123 designs and 92 criteria, that rating has actually climbed up by around 0.1 GCI points monthly.

The report states comprehensive analytical analysis reveals no substantial structural speedup after April 2026. For public designs, the month-to-month modification before April 2026 was 0.091 points. After that date, it was 0.118 points. The report discovered no considerable distinction in between the 2.

Gemini 4 Argon itself landed partially above the anticipated pattern line. DeepMind’s report recommends that bump primarily originates from more calculate and organizational financial investment, not from the design accelerating its own advancement.

news AI composes the code, human beings still sign off

AI tools add to editing 75% of production code at DeepMind. Some 88% of code modifications still require human evaluation.

Completely handed over modifications, where AI deals with the work end to end, increased from 7.5% to 11.9%.

Studies of 12 senior DeepMind scientists assist describe the pattern. Automation works best on plainly specified jobs where the response can be confirmed. It has a hard time terribly with open-ended clinical ideation and complex preparation.

news The benchmark scorecard

DeepMind likewise ran its representatives through RE-Bench, an internal standard of 74 practical maker finding out engineering jobs. The Argon representative resolved approximately half of them. Gemini 3.7 Flash handled 27%.

A different evaluation of 1,389 Argon representative runs discovered that success rates fell significantly on intricate jobs when no human was included.

news What this suggests for the AI market

The share of completely handed over code modifications is one pattern to view, given that its relocation from 7.5% to 11.9% is the closest thing to an automation pattern line in the report. A continual break above approximately 0.1 GCI points each month, without a coordinating dive in calculate, would be the signal the self-improvement camp has actually been waiting on. In the meantime, DeepMind’s own journal reveals a constant climb and a great deal of people still checking out the code.

Disclosure: This short article was modified by Diego Almada Lopez. To learn more on how we produce and evaluate material, see our Editorial Policy.