Chollet to Jensen Huang: 'You Don't Have AGI. You Have Expensive Autocomplete.'
The ARC-AGI creator fires back at NVIDIA's AGI claims with a benchmark that every frontier model fails. His argument is simple: if humans can do it cold, AGI should too.
Vlad Makarovreviewed and published
François Chollet didn't name Jensen Huang directly. He didn't need to. Two days after the NVIDIA CEO declared AGI achieved and one day after the term's inventor agreed, Chollet posted a thread on X that amounted to a precision strike.
His argument boils down to one sentence: "If a normal human with no instructions can do it, and your system can't, then you don't have AGI — you have a very expensive autocomplete that needs a lot of help."
The data backing him up is ARC-AGI-3, released the same day. Every frontier model scored under 1%. Untrained humans scored 100%. The gap isn't a few percentage points — it's 99+ points wide.
Chollet distinguishes between knowledge and fluid intelligence. Models have more stored knowledge than any human. But they struggle to recombine that knowledge on the fly when facing something genuinely novel. "We are very bad at fluid intelligence," he says of current AI systems, "which is taking patterns and actually combining them on the fly to form a new model of a problem."
The week's AGI debate now has three clearly defined camps. Huang says AGI is here because AI agents can generate economic value. Gubrud (who coined the term) agrees because models match "high-human level" knowledge at thousands of times the speed. Chollet says neither is right — because economic output and stored knowledge aren't the same thing as general intelligence.
The question is which definition matters. For investors and CEOs, Huang's is more useful. For researchers building the next generation of AI, Chollet's benchmark is the test that hasn't been passed. And until something scores meaningfully above 1% on ARC-AGI-3, the argument that AGI has arrived remains a claim without empirical support.
