rsi.watch

Google DeepMind · 2026-10-13

Gemini 4 Argon Safety & Security Report: no significant end-to-end acceleration of AI R&D

“there is currently no significant end-to-end acceleration of AI R&D, AI-driven or otherwise”
Who
Google DeepMind
Date
2026-10-13 Date of the Google blog post that released the report. The report itself says October 2026.
Evidence
Very low certainty Starts at Low: the developer’s own numbers.
  • ↓ indirect measure. Acceleration is read off an internal capability index (GCI v1). The report says the index does not capture speed-ups in parts of R&D while other parts are the bottleneck. “the rate of frontier capability improvement does not capture acceleration of the many subprocesses that make up AI R&D as long as overall progress remains bottlenecked by other factors”
  • ↓ imprecise. After the breakpoint the interval runs from 0.053 to 0.171, which includes rates close to twice the earlier 0.091. The intervals come from 200 bootstrap refits. “The estimated slope after 1 April 2026 is 0.118 [0.053, 0.171] GCI points per month”
how we rate evidence
Replication
None recorded
Checked
2026-10-16, against the source page
Links
Source

Also on the page

  • The test: a regression with a breakpoint on 1 April 2026. The slope was 0.091 GCI points per month before and 0.118 after; "we do not observe a statistically significant difference". With internal models included: 0.097 before, 0.107 after.
  • The report calls this "a small (but not statistically significant) uptick". Gemini 4 Argon itself sits above trend, which DeepMind suspects "is due to increased organizational investment and an increase in training compute".
  • "Most production code at GDM is now >75% AI-written". Experimental code is not included. Fully delegated pull requests rose from 7.5% in April to 11.9% in September 2026.
  • On its internal ML R&D evaluation (GRB), "Gemini 4 Argon solves only around half of the challenges".
  • How much is automated comes from a survey of 12 senior leaders and researchers, and from 1389 internal trajectories graded by Gemini Flash 3.8.

← Ledger Report an error