DeepMind: no measurable AI R&D acceleration yet
Confirmed
Importance: major (4 of 5)The takeaway
The Gemini 4 Argon Safety & Security Report (Oct 13, 2026) contains Google DeepMind’s first holistic assessment of whether AI is speeding up its own AI research.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Major (4 of 5)
- Last verified
- 16 October 2026
Your AI and this story
- GPT-6 Astra167 days after its cutoff
- Claude Opus 5.5106 days after its cutoff
- Gemini 4 Argon197 days after its cutoff
- Grok 4.7136 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 106 days before it.
Key facts
- Finding 1: ‘there is currently no significant end-to-end acceleration of AI R&D, AI-driven or otherwise’
- Measure: GCI v1, an internal index like Epoch’s ECI, fitted on 123 models × 92 benchmarks with 200 bootstrap refits; frontier Gemini models gain ~0.1 GCI points per month
- Hinge regression with a breakpoint on April 1, 2026: 0.091 [0.071, 0.108] points/month before vs 0.118 [0.053, 0.171] after (public models); 0.097 vs 0.107 including internal models; not statistically significant
- Gemini 4 Argon sits above trend, which DeepMind attributes to ‘increased organizational investment and an increase in training compute’, not AI-driven acceleration
- Code: >75% of GDM production code is AI-written; fully delegated pull requests rose from 7.5% (April) to 11.9% (September 2026); ~88% of PRs still need iterative human steering or human code
- Survey of 12 senior leaders/researchers: automation highest on well-specified, verifiable execution tasks; lowest on research ideation and long-horizon ‘research taste’
- 1,389 sampled internal Argon trajectories (labelled by Gemini Flash 3.8): the larger the task, the less often Argon finishes it alone
- No ‘automation overhang’: Argon solves only about half of GRB, the internal ML R&D benchmark; full automation would need a near-perfect score. ML R&D acceleration and automation CCLs: not reached
Show 1 more
- Contrast from Google’s leadership: in an Argon roundtable video Google shared on Oct 14, Sergey Brin said models now help build their successors and Google is ‘starting to get onto that RSI treadmill’: ‘our researchers and the model work hand in hand to make the next version of the model’ (Business Insider, Oct 15). DeepMind’s Vihan Jain posted on X that post-training teams now ask for ‘Argon quota’ rather than chips
What happened
Section 3.4.2 of the 217-page Gemini 4 Argon Safety & Security Report, “Holistic Assessment of ML R&D Risks”, tests whether AI is speeding up AI research at Google DeepMind. It combines four kinds of evidence:
- a capability index fitted to 123 models on 92 benchmarks;
- telemetry on how much code is written by AI;
- a sample of 1,389 internal Argon agent trajectories;
- a survey of 12 senior researchers.
The pace of frontier Gemini capability gains has been steady at about 0.1 index points a month. A regression with a break on April 1, 2026 shows a small rise that is not statistically significant. Most production code is now AI-written, and the share of fully delegated pull requests rose from 7.5% to 11.9% between April and September. About 88% of pull requests still need human steering. Argon solves about half of DeepMind’s internal research-engineering benchmark, far from the near-perfect score that full automation would need. DeepMind concludes “the risk from AI-driven AI R&D is currently low”.
On Oct 14 Haydn Belfield, a DeepMind researcher who worked on the assessment, posted a nine-part thread calling it “our first holistic empirical assessment of ML R&D automation and acceleration inside Google DeepMind”. The thread had about 3.4k views on Oct 16, per fxtwitter. Decrypt’s Oct 15 article contrasted the finding with Anthropic leaders’ claims that recursive self-improvement is already speeding up.
Why it matters
This is the first published lab measurement of end-to-end AI R&D acceleration rather than code share or anecdotes. It shows that “AI writes most of our code” and “AI is speeding up AI progress” are different claims. The result is a counterpoint to the 2027 timelines for recursive self-improvement given by Jack Clark and others, and to Sergey Brin’s “RSI treadmill” remark in the same week. The two can both be true, because Brin describes the process and the audit measures its effect on the rate of progress. It has caveats. The index and the benchmark are internal and unpublished. A post-April window of about six months leaves wide confidence intervals: the upper bound of 0.171 points a month is almost twice the earlier rate. Argon itself is above trend.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 1 primary, 2 press, 1 reaction
Primary
- Google DeepMind: Gemini 4 Argon Safety & Security Report (PDF), section 3.4.2storage.googleapis.com, official
Press
- Business Insider (via MSN): Sergey Brin says Google’s AI models are reaching a critical moment of self-improvementbusinessinsider.com, press
- Decrypt: Google Says AI Isn’t Building Itself Yet—Even Though It Writes 75% of the Codedecrypt.co, press
Reactions
- Haydn Belfield on X: 9-part thread on the ML R&D assessment (Oct 14)x.com, discussion
Changes
- Filed (Belfield thread via fxtwitter)