Geminiexchange
DevelopingResearch

Google DeepMind Says AI Research Is Not Accelerating Despite AI-Written Code

Google DeepMind's Gemini 4 Argon safety report finds no significant speedup in AI research, even though the lab says AI now writes more than 75% of its production code.

Why it matters

The finding runs against the runaway-development scenario that the report says has spooked AI executives into calling for a slowdown, and it comes from the lab behind Gemini 4 Argon itself. The report pairs that conclusion with a warning that existing guardrails would be 'severely inadequate' in the runaway case, a line aimed at the policy question of how AI systems should be tested and regulated. Nothing in Decrypt's account reaches crypto markets, tokens, exchanges or regulation, so its direct significance for this desk is limited.

Key facts

Google DeepMind's safety assessment for Gemini 4 Argon found no statistically significant acceleration in AI research after April 1, 2026.

Reported by Decrypt

DeepMind's capability index has risen a steady 0.1 points a month.

0.1 index points per monthReported by Decrypt

More than 75% of production code at Google DeepMind is now AI-written.

75 %Reported by Decrypt

About 88% of code changes at the lab still require a human to steer the AI or write the code themselves.

88 %Reported by Decrypt

Gemini 4 Argon solved roughly half of the internal research engineering tasks it finished.

Reported by Decrypt

DeepMind treats a 90% score on those tasks as a red flag for full automation.

90 %Reported by Decrypt

DeepMind researcher Haydn Belfield wrote on X that there is 'currently no significant end-to-end acceleration of ML R&D, AI-driven or otherwise' and concluded the risk from AI-driven ML R&D is currently low.

Reported by Decrypt

The report describes recursive self-improvement as reaching 'a situation where existing guardrails are severely inadequate'.

Reported by Decrypt

What happened

Decrypt reported that Google DeepMind's safety assessment for Gemini 4 Argon found no statistically significant acceleration in AI research after April 1, 2026, with the lab's capability index climbing a steady 0.1 points a month. The same report says more than 75% of the lab's production code is now AI-written, yet about 88% of code changes still need a human to steer the AI or write the code themselves, and Gemini 4 Argon solved roughly half of the internal research engineering tasks it finished, well below the 90% score DeepMind treats as a red flag for full automation. DeepMind researcher Haydn Belfield wrote on X that there is 'currently no significant end-to-end acceleration of ML R&D, AI-driven or otherwise' and that the lab therefore puts the risk from AI-driven ML R&D at low. The report frames the thing it looked for as recursive self-improvement, the chain in which AI builds a smarter version of itself, and warns that this would reach 'a situation where existing guardrails are severely inadequate'. Decrypt is the only outlet in this evidence set, so every figure and the low-risk conclusion rest on that one report, and no second outlet has weighed in.

How the story developed

  1. First report

    First report by Decrypt

    Google Says AI Isn't Building Itself Yet—Even Though It Writes 75% of the Code

    Decrypt

  2. Official statement

    Google DeepMind safety report for Gemini 4 Argon concludes there is no significant end-to-end acceleration of ML R&D

    The lab's assessment of whether AI is speeding up AI research found a capability index rising about 0.1 points a month after April 1, 2026, alongside a Gemini 4 Argon solve rate of roughly half of the internal research engineering tasks it finished.

    Decrypt

  3. Update

    Decrypt reports the DeepMind assessment

    Single-outlet report carrying the capability index figures, the AI-written code share, the human-steering share and Haydn Belfield's X thread.

    Decrypt

Affected entities

Original sources

PublisherReportRolePublished
Decrypt
Crypto media
Google Says AI Isn't Building Itself Yet—Even Though It Writes 75% of the CodeOriginal