The Treadmill and the Speedometer: Google Declares a New Gear Its Own Report Can't Measure
Google is “starting to get onto that RSI treadmill,” Sergey Brin said at a roundtable on Gemini 4 Argon, in a video the company shared on Wednesday.1 “Our researchers and the model work hand in hand to make the next version of the model,” he added, pointing to the Flash models shipped just before Gemini 4 as evidence, and noting Google would be “a couple of models ahead internally” by the time anyone watched. The same month, Google DeepMind published its safety and security report for Argon. Its finding: “no significant end-to-end acceleration of AI R&D, AI-driven or otherwise,” with the risk from AI-driven AI R&D rated “currently low.”2 Recursive self-improvement — the event the field treated as a Rubicon for a decade — has been declared, and unmeasured, by the same institution in the same week. Four frames on how that is possible.
The Golden Circle: Why, How, What
Why does Google narrate the treadmill? Because RSI is the frontier’s scarcest narrative good: the transition every rival — Anthropic, OpenAI, Google itself — reportedly expects within three to twelve months, absent intervention.3 A lab that conceded it was absent from that story would be conceding the future tense. There is also a regulatory clock: Representative Ro Khanna plans a bill to ban self-improving AI,1 so claiming to be “starting to get onto” the treadmill positions Google inside the story while visibly below the tripwire. How does Google know? Its report measures code share (over 75 percent of production code AI-written, up from less, but 88 percent of changes still steered by humans), task completion (Argon solves roughly half of internal AI-research challenges; full automation would need near-perfect scores), and benchmark velocity (an index of 92 benchmarks rose from 0.091 to 0.118 points a month after April 1 — a difference the report itself calls not statistically significant).2 What did it produce? Two artifacts pointing opposite directions: a null measurement wrapped around a gear-engagement narrative. The alignment check fails by design — the Why needs the flywheel turning; the What needs it not yet dangerous. Both documents serve the Why.
The Counterfactual: If the Gear Had Engaged
Test the claim. Minimal intervention: suppose the RSI flywheel truly reached a “new gear” in 2026. First-order, near-certain: the internal benchmark index inflects sharply. It did not — the post-April trend break is statistically insignificant, and Argon’s position above trend is attributed by the report to more investment and compute, “rather than reflecting AI-driven acceleration.”2 Second-order, probable: the hard bottleneck of research — choosing directions, designing experiments, diagnosing training failures — would show automation. It does not; humans hold those tasks, and the research-task benchmark sits near fifty percent. Equilibrium check: the observables Vihan Jain cites — post-training demand shifting from raw chips to Argon quota — are real but overdetermined by ordinary scaling; reallocating compute to your best model is what every lab does with its best model, acceleration or not.1 Verdict: the evidence is contingent on investment and overdetermined toward “no acceleration yet.” The treadmill quote survives only because it is unfalsifiable as stated — “starting to get onto” a treadmill is compatible with standing beside it.
The Ladder of Abstraction
Down at the concrete level: one quote and one report line, both from Google, both about the same model. Climb one rung: this is now the industry pattern. OpenAI self-certified Astra’s “critical” cyber capability in September; Anthropic memoted its way through an IoT self-replication test this month; each lab is simultaneously the claimant and the measurer of the same threshold. Climb again: when a threshold is at once a marketing asset and a regulatory trigger — the thing investors price and Khanna’s bill would ban — the definition of the threshold becomes the contested object. Whoever measures the event owns the event. Google’s answer to that problem is to publish its own speedometer, calibrated and hedged, before anyone else can build one.
A Parable: The Treadmill
A gym installed the famous treadmill that was said to run by itself. The owner posted a sign: THE MACHINE RUNS ITSELF. The gym’s own inspector filed a report — the belt moves no faster than last year, the runners still push with their own legs, the mileage counter climbs because more people joined. The owner read it carefully and posted a new sign: WE ARE STARTING TO GET ONTO THE SELF-RUNNING TREADMILL. He was not lying. The sign does not say the treadmill runs.
What the Frames Converge On
The report appears honest; the rhetoric is tactical; neither requires the other to be false, because “RSI” has no operational definition both parties are bound to. The consequential move is not Brin’s quote — it is that Google has pre-committed the measurement instrument for the event everyone expects within the year. When genuine acceleration arrives, the finding will be certified by the institution whose valuation, hiring narrative, and regulatory posture all improve if the announcement is early rather than late. Watch the definition, not the needle.