Google's own audit cannot find the AI flywheel its founder says has started

Google put out two documents this week that disagree, and only one of them describes what Alphabet's▼−2.37% — Alphabet, down 2.37 percent today spending is buying. On October 14, 2026, Google DeepMind published the first internal audit of whether AI is accelerating AI research↗, tucked into the safety report for its new frontier model Gemini 4 Argon, and found no speedup the machines could claim credit for ¹².

A day later, Google co-founder Sergey Brin told a roundtable that his researchers and their models "work hand in hand" and that Google is "starting to get onto that RSI treadmill↗" ³⁴. Recursive self-improvement, or RSI, is the loop in which AI helps build the AI that replaces it, with each generation supposed to arrive faster than the last ³.

The audit found a straight line, not a flywheel

The assessment, led by DeepMind researcher Haydn Belfield, tracks frontier capability on a General Capability Index: one number per model, blended from scores across 123 models and 92 benchmarks, something like an IQ score for AI ¹². Since late 2024 the score has climbed about 0.1 points a month, a remarkably steady line ². For its public models, the monthly gain before April 2026 was 0.091 points. After April it was 0.118, and the report found no statistically significant difference between the two ¹.

Argon itself sits marginally above the trend, and the team traces that to increased compute and organizational investment rather than self-acceleration ¹². As Belfield put it, "There is currently no significant end-to-end acceleration of ML R&D, AI-driven or otherwise↗" ². One caveat cuts both ways: the numbers describe only DeepMind's own lab, not OpenAI'sPVT$1.18T — OpenAI, private, latest valuation $1.18T or Anthropic'sPVT$350B — Anthropic, private, latest valuation $350B ².

AI writes the code, humans still steer almost every change

The flywheel would have to engage in the middle of the chain, and that is where it stalls. AI tools now write more than three quarters of DeepMind's production code, yet about 88% of code changes still need a human steering the AI or writing code before anything ships ¹². The share AI completes end to end moved only from 7.5% to 11.9% between April and September ².

On RE-Bench, an internal suite of 74 machine-learning engineering tasks, the Argon agent finished about half ¹. Gemini 3.7 Flash, the previous generation, managed 27%. Both sit far below the 90% score DeepMind treats as a red flag for full automation ¹². Across 1,389 reviewed agent runs, the bigger the task, the less likely the model finished it alone ².

A survey of 12 senior DeepMind researchers found automation highest on well-specified, verifiable execution and lowest on open-ended ideation and long-horizon research taste ¹². What stays scarce is verification, review, and the judgment about which idea deserves compute.

The spending assumes the founder's version

Brin says Google is already a couple of models ahead of Argon internally and expects the pace only to accelerate ³⁴. Vihan Jain, a Gemini RSI lead at DeepMind, posted that in post-training, the stage where a finished base model is refined on feedback and data, demand had shifted from asking for raw chips to asking for "Argon quota" ³.

The thesis moves real money. Chip stocks fell almost 6% a month ago when Anthropic's chief executive, Dario Amodei, argued that recursive self-improvement was already accelerating ². A report published on Wednesday by Gladstone AI↗, a firm that advises the U.S. government on AI policy, found most researchers it interviewed expect RSI within three to twelve months ³⁵⁶. Alphabet spent about $45 billion on capital projects in the second quarter↗, against a measured record in which every point of capability still comes from chips and people ⁷.

Read the ledger before pricing the flywheel

None of this is a case against AI. The index climbs steadily, the agents are real, and Argon, unveiled on September 30, is genuinely finishing long-horizon engineering work across Google ²⁸. It is a case for separating the measurement from the story: on DeepMind's own numbers, every point of capability so far traces back to compute and headcount rather than to machines compounding their own progress.

Two lines in DeepMind's ledger will settle the question: the monthly slope on the capability index, and the fully delegated share of code changes. A slope that steepens without new compute, and changes that stop waiting for review, would be the flywheel arriving. Until then, the value goes to whoever builds the review and trust layer between what AI writes and what an organization ships.