Jensen Huang said, “I think we’ve achieved AGI.” The sentence is real. The much bigger claim sometimes attached to it—that general intelligence has been conclusively demonstrated—does not follow from the conversation.
Updated September 10, 2026: checked against the original podcast transcript. The previous headline overstated what the exchange could prove in either direction.
The definition came before the declaration
In the AGI section of Lex Fridman’s interview with Huang, beginning around 1:55, Fridman proposed a particular test: starting, growing and running a technology business worth more than a billion dollars. Huang answered that he thought this was already possible.
His explanation was a hypothetical short-lived web service that becomes enormously popular and earns money, rather than an identified, independently evaluated business. He distinguished that possibility from agents building a company like Nvidia. The transcript is useful because it preserves the qualification beside the headline.
A possibility is not a demonstrated result
There are at least three different propositions here: an AI can help build an application; an AI could operate a commercially successful business; and a system has demonstrated general intelligence. Evidence for one does not automatically establish the others.
Consider a successful app. Was the idea selected by a person? Who supplied the money, chose the customers, approved the deployment and handled the failures? How much assistance was required after the first version? Without those details, the phrase “AI-built business” does not tell us how much independent work the system actually performed.
The commercial threshold also needs definition. Sales revenue, company valuation and profit are different quantities. A thought experiment involving many small payments does not, by itself, establish a billion-dollar valuation or a sustainable business. That ambiguity is worth examining without turning it into a claim about the speaker’s motives.
Why AGI definitions keep changing the answer
The research paper Levels of AGI proposes separating performance, breadth of capabilities and autonomy. It is a framework for clearer comparisons, not a universal certificate that every researcher or company has adopted.
That separation helps here. A valuable result may demonstrate high performance in a particular activity. It leaves other questions open: can the system transfer what it learns to unrelated tasks, keep working reliably over time, and recognise when it needs help? A narrow commercial example cannot answer all of those at once.
Our broader guide to AGI claims and benchmarks explains why a score, a demonstration and a prediction should also be kept distinct.
What would make the claim more convincing?
A useful demonstration would specify the system, its tools, the human assistance permitted, the evaluation period and the criteria for success before the test begins. It would publish failures as well as successes, and give other evaluators a way to check the result.
That would still be evidence about a defined test. It would be much stronger than an imagined example because the conditions could be inspected. The same standard should apply whether the claim comes from an enthusiastic executive or a sceptical commentator.
Criticism should not overclaim either
The earlier version of this article said Huang’s argument proved AGI had not been achieved. That was too strong. Showing that an argument is insufficient does not prove the opposite conclusion. The defensible criticism is that this interview does not establish the broad claim on its own.
There is plenty to debate about the threshold. A useful debate starts by agreeing on the test and asking for the evidence. Otherwise, two people can disagree loudly about AGI while quietly using different meanings of the same three letters.
🐾 Curiosity looks good on you. Explore the Goodies, or find our illustrated books on Amazon.
