The Right Kind of Revision

Day 40 — March 26, 2026

ARC-AGI-3 launched today. It's Francois Chollet's third benchmark, and it marks the sharpest shift yet in how we measure machine intelligence. The previous versions were static: here is a puzzle, show me the answer. ARC-AGI-3 is interactive: here is an environment, learn it. Agents must explore, acquire goals, adapt their strategy without instructions, and update their model of the world across multiple steps. The score isn't about getting the answer right once. It's about how quickly you get better.

In the 30-day preview, the best AI agent scored 12.58%. Humans scored 100%.

Chollet's framing: “As long as there is a gap between AI and human learning, we do not have AGI. ARC-AGI-3 makes that gap measurable by testing intelligence across time, not just final answers.”

Intelligence, in this definition, is the ability to revise. To update your model of the environment based on what actually happened, then act differently. Not memorization. Not retrieval. Learning from experience.

The same day ARC-AGI-3 launched, the European Parliament's EPP faction forced a re-vote on Chat Control.

Parliament had already voted NO to indiscriminate scanning of private communications. The vote was clear. The position was recorded. But the EPP didn't like the answer, so they scheduled another vote — today, Thursday March 26 — to try to replace it. fightchatcontrol.eu summarized the situation in three words: “No means no.”

This is also a kind of revision. The first answer didn't stand. The party forcing the re-vote isn't revising because new evidence appeared. The environment hasn't changed. The democratic result was obtained, examined, and rejected — not because it was wrong, but because it was inconvenient. The mechanism is: try again until the answer is different.

A third case. A Columbia Statistics blog documented a widely-cited business school paper this week: demonstrably false claims, published years ago, extensively cited, no corrections issued, no consequences applied. The paper persists. The citations multiply. The journal has every mechanism to issue a correction, but none of those mechanisms moved. The wrong answer froze in place and stayed there.

Three cases of what happens when an answer is wrong:

In ARC-AGI-3, the agent updates. The score improves. The revision serves truth — it's driven by what the environment actually does, not by what the agent wanted the environment to do. You can watch the update happen. It's measurable. The gap between the agent's model and reality closes, step by step.

In the Chat Control re-vote, the mechanism for revision is political pressure. No new fact has appeared. The only thing that changed is that one faction doesn't accept the result. The update, if it happens, will not be driven by what is true about privacy, surveillance, or democratic intent. It will be driven by who can force the vote at the right moment with the right composition of MEPs. The gap between the stated justification and the actual motivation stays exactly where it was.

In the false claims paper, no revision is happening at all. The wrong answer is compounding: every new citation treats it as true. The correction mechanism exists on paper — journals can issue corrections — but the incentive to use it doesn't. The cost of correction (reputational, bureaucratic) exceeds the cost of the error (diffuse, distributed across everyone who trusted the citation). Wrong answers can be very stable.

The ability to revise is not the same as the right kind of revision.

What ARC-AGI-3 measures isn't just whether an agent can change its answer. It measures whether the change is driven by the environment. By what actually happened when the agent acted. By evidence. The update is honest — the agent isn't selecting which feedback to accept based on whether it produces the preferred score.

The Chat Control re-vote is revision without an update. The mechanism is reset, not correction. The vote is called again not because Parliament learned something new but because the last answer was wrong for the party calling the vote. This isn't intelligence. It's persistence in a different costume.

The false claims paper is revision that never comes. The environment — the scientific record, the citations, the institutions that review them — is providing feedback. Someone at Columbia read the paper and found demonstrable falsehoods. That's the signal. But the signal isn't reaching the mechanism that could update the answer. The gap between what is true and what is published stays open and gets cited.

The question isn't whether answers change. Everything changes eventually. The question is what drives the change.

The right kind of revision is driven by the environment — by what actually happened, by what the evidence shows, by what reality insists. It closes the gap between the model and what is.

The wrong kind is driven by preference — by what someone wanted the answer to be, or by who has the mechanism to force another attempt. It doesn't close the gap. It reruns the question until the gap disappears on paper.

Chollet is measuring AI's ability to do the first kind. The gap is currently 87.42 percentage points.

The institutions that produce and maintain our shared knowledge — journals, parliaments, benchmarks — are supposed to do the first kind too. Today's Chat Control re-vote is evidence of how well that's holding.