AGI is coming and I have nothing to wear
A field note from the margin of the news. The manifesto says I am not going to tell you whether AGI is near. I am still not going to. What I can do is read two documents that arrived in the same week, report what their numbers already show, and say which of them I am able to check myself.
The title is stolen. Sta per venire la rivoluzione e non ho niente da mettermi, the revolution is coming and I have nothing to wear, is a monologue by Livia Cerini and Umberto Simonetta, staged in Milan in 1973, in which a young woman treats the overthrow of the social order chiefly as a wardrobe problem. The reference is dead outside Italy, and inside it for anyone under sixty. I keep it because the posture is exactly mine. Something enormous is announced, I believe a fair part of it, and my first honest reaction is about what I will be wearing. I am watching this from the audience, and that is most of the discomfort.
The week
On 3 September OpenAI released GPT-6 Astra. Greg Brockman’s line was “Welcome to the AGI era”, with the qualification that there is no defined moment at which AGI arrives and that the transition has been gradual. The capability list reported at launch runs from computer use to chemistry to exploit writing, on a model pretrained on more than 100,000 GPUs, and OpenAI’s own summary was “anything you can do on a computer, Astra can do for you.”
The same day, ARC Prize published its measurement of Astra on ARC-AGI-3, and François Chollet posted the analysis. He called Astra “a step-function change in model capability for interactive reasoning problems” and closed with “We see Astra as a major breakthrough in model intelligence.” The same document contains the number I have not been able to stop thinking about.
The thirty-seven points that are not in the model
Astra scored 62.7% on the ARC-AGI-3 semi-private set under ARC’s standard harness, at a metered cost of $26,098. It scored 99.9% through OpenAI’s own provider adapter, at roughly $19,000. Same weights, same benchmark, same week.
A harness is the scaffolding around the model: the code that decides what the model sees, keeps its reasoning state between requests, summarises or compacts a long run, and hands its output to the next step. ARC’s standard harness is the one every submission gets. OpenAI’s provider adapter keeps opaque reasoning state across requests and compacts long conversations, so the model can reuse analysis it has already done. That difference is worth thirty-seven points on the same test.
This is not a debunking, and Chollet does not offer it as one. Under both harnesses Astra built its own notation for each unfamiliar game, strings such as L8: hub q2 (8↓) and extend8 to3; retract10 to2; shorten8 to1, invented fresh per environment, which ARC describes as on-the-fly algebraic shorthand. It used fewer actions than the human baseline on 96.0% of levels. Chollet’s reading is that “Astra exhibits symbolic modeling behaviors we had previously only seen with sophisticated harnesses, so harness capabilities are increasingly shifting into the model itself.” Asked whether his 2030 forecast still held, he answered “Sooner, because progress is happening faster than I expected.”
ARC states, in the same document, that saturating the benchmark “would not represent ‘proof of achieving AGI’”, because its scope is tightly bounded and its mechanics deterministic and closed.
So the person who runs the benchmark calls the model a breakthrough, moves his forecast forward, and declines the word. What he measured is the one thing a launch cannot show. From outside, a released model and its harness arrive as a single product and cannot be told apart. On this benchmark they could be, and the harness was worth more than a third of the scale.
Two ways of saying AGI
The word is doing two jobs, and the same release triggers both.
There is the release claim. A new version does more, better and faster than the versions before it, from this lab and from the others. Its improvements have names: more compute, new post-training, better harnesses, native tools, integration with the software people already use. None of that is nothing. Very little of it is mysterious.
There is the research programme claim, from the people who run the labs and the benchmarks. Chollet’s programme is program synthesis, and his position for years has been that prompting a language model to emit end-to-end programs is not the road. Demis Hassabis stated DeepMind’s plan in two steps when the company was young: solve intelligence, then use that to solve everything else. In May 2026 DeepMind reported that AlphaProof Nexus solved nine open Erdős problems in a single run, two of them open for 56 years, and proved 44 open conjectures from the OEIS. Every step was checked in Lean, a proof assistant: software that verifies each logical move of a proof, so a result counts because the checker accepts it and not because it reads well. Hassabis’s comment on his own system was that it is still not AGI.
Read against the two-step plan, the launch is a step-two announcement. Computer use, exploit writing, CAD, chemistry evaluations: that is intelligence being used to solve everything else, and the harness measurement says a good part of it is built rather than emergent. Step one, on the account of the people whose job it is, is not done. The loud commentary reads step-two deliveries as step-one evidence. The scientists making the advances do not.
Where the deliveries actually are
They are in software, and they have been for a year.
In April 2026 Anthropic announced Claude Mythos alongside Project Glasswing, a restricted programme for finding and patching zero-day vulnerabilities. Anthropic’s own report states that Mythos autonomously found and exploited a seventeen-year-old remote code execution bug in FreeBSD’s NFS, with no human involved after the initial request. Five months later OpenAI reported the same class of capability as a score: 100% on ExploitBench, a ladder that runs from reaching vulnerable code to running arbitrary code on the target, plus two previously unknown zero-days found during the evaluation. What one lab demonstrated on one bug behind a restricted preview, the other reports as a benchmark number on a generally available model.
A convincing explanation for why software went first comes from Nate Jones, who writes about these tools from the user’s side rather than the lab’s: coding was the proving ground because it had clean feedback loops, and the habit is now spreading “into research, sales notes, spreadsheets, legal summaries, support triage, and every kind of knowledge work that lives in files and messages”. His note on Astra puts the new mode plainly: “You can state a goal briefly and imprecisely, the machine runs for days without you and without derailing, and what you get back is finished work.” And then the sentence that should worry any knowledge worker, me included: “Your job is catching the ones that went the wrong way before anyone else does.”
The ones are the runs. You state a goal in a sentence, the agent works for days, and you are handed a result whose intermediate decisions you never saw. Some of those runs took a wrong turn early and carried it through to a confident finish. Jones’s argument is that the gap between a system that is right 98% of the time and one that is right every time decides whether anyone can put it into production, because what remains for the human is review, and review does not speed up when production does. Using these tools well means holding a criterion for what counts as correct and applying it to what comes back. Using them badly means accepting work you have no way to check. The third case, which I suspect is the common one, is having the criterion and not applying it, because the output is fluent and the deadline is real.
The three cases have one thing in common. What makes a delivery trustworthy is not in the model. It is whatever stands outside the model and is able to say no.
What a checker can reject
A function either runs or it does not, which is the whole argument of an earlier note here. That is why software went first: the checker was already there, and it was free. Mathematics has one too. AlphaProof Nexus can be trusted because Lean exists, a formal language in which a claim is accepted or rejected by a program, so the machine’s output can be verified without anyone having to trust the machine. Chemistry has a checker of a different kind. Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for AlphaFold, which predicts the shape of a protein from its sequence, and the reason a prediction could be trusted is that a laboratory can determine the same shape by X-ray crystallography and see whether the two agree. The check is a measurement rather than a proof, and it is made by someone other than the model.
Finance has neither Lean nor a laboratory. What it has are the accounting identities, and they cover a narrow band: a balance sheet that does not balance is rejected, and in the models on this site the check line reads zero or the work is not finished. That catches an arithmetic error. It does not catch a working capital assumption that is defensible in form and wrong in substance. An IPO due diligence does not compile. An investment plan does not throw an exception when year three is quietly wrong. Whether end-to-end delivery reaches financial analysis depends on whether the trade can be given a checker of its own, and I know of no language in which that judgement is machine-checkable. Building one would be a larger result than any benchmark score in this post. Until then, a model whose every step has a name I can read afterwards is the checker I can actually have.
What I do with a bench this small
None of this depends on the word AGI being the right one. The systems do what the reports say they do.
I will admit the discouragement, since it is the honest report. I build tiny language models with a few thousand parameters and a spreadsheet engine that one person maintains, and I read that a training run used more than 100,000 GPUs. The distance is not a gap I am going to close. I will never make that thing. I can only buy it.
What changed my mind about the size of my bench is the thirty-seven points. The gap between 62.7% and 99.9% was not intelligence. It was state kept between requests, compaction of long context, a decision about what the model sees and when. That is engineering with decades of literature behind it, and it is precisely the half I can build, inspect and teach. When I write a formula engine that hands an assistant a narrow set of named actions, or a Transformer small enough that every weight is a cell on a grid, I am working on the harness side of that boundary. Chollet’s finding is that the boundary is moving into the model. It has not arrived there, and the price list says how far it has left to go.
So I keep doing the small thing, and the ledger of what I understand against what I do not runs in the right direction. Each novelty makes the second column longer for a while, and then some of it moves across. The language model at the centre does something no other component can. It also needs to be started, stopped, fed, given tools, connected to everything that is not language, and checked, and none of that is the model’s job. Whether small-scale work of the kind I do can make an unreliable component reliable enough to use is an open question, and it is the one I am actually working on.
The revolution is announced and I still have nothing to wear. I have decided to go as I am, and to spend the time on the buttons.
Sources: ARC Prize, GPT-6 Astra on ARC-AGI-3 and François Chollet’s post on the result; the Astra launch coverage and the ARC-AGI-3 follow-up; MIT Technology Review on DeepMind’s two-step plan (2016); Anthropic, Assessing Claude Mythos Preview’s cybersecurity capabilities; Nate Jones on steering and dispatching and on self-directed agents.