NVIDIA Jensen Huang’s explanation of AGI is telling

Artificial General Intelligence, or AGI, has spent the last year or so as the AI ​​industry’s favorite term. With the industry’s leading companies burning capital at historic rates, racking up energy costs and growing investor expectations struggling to meet quarterly, the promise of near-machine intelligence has become a useful thing to have in your back pocket.

Whether we are close to that milestone depends entirely on how you define it. That flexibility, it turns out, does a lot of work.

Take, for example, Jensen Huang, CEO of NVIDIA – a company currently valued at around $4 trillion, built largely on the GPU hardware powering the AI ​​boom – who recently sat down with podcaster Lex Fridman for a wide-ranging conversation that included data centers, geopolitics, and the question of whether AGI has already arrived. Huang thinks it has. The rationale for that claim, however, is questionable.

As Fridman points out, Huang has said that the timeline for AGI depends on what you define. At the 2023 New York Times DealBook conference, Huang described AGI as software that can pass tests that measure general human intelligence at a fairly competitive level. He expected AI to clear that bar within five years.

For his part, Fridman gave Huang an open definition to work with: A true AGI, in Fridman’s formulation, would look like an AI that could start, grow, and run a technology company worth more than a billion dollars. He asked if that could be achieved in the next five to 20 years, given the recent proliferation of agent AI tools like OpenClaw.

Huang didn’t need it for five to 20 years. “I think so now. I think we’ve got AGI,” Fridman replied.

That, however, is based on a narrow interpretation of what Fridman is asking. The way Huang sees it, AI doesn’t need to build anything permanent. It does not require managing people, navigating the board, or supporting the business. It just needs to hit a billion dollars once.

“You said billion,” Huang told Fridman, “and you didn’t say forever.”

The line in both cases is not a consistent theory of machine intelligence. It’s a consistent pattern of defining the threshold in any way that makes “yes, we are” a very easy answer. His illustration of what that might look like.

After his initial answer, Huang lays out his thoughts, describing a scenario where AI creates a simple web service — some program that goes into the bloodstream, is used by a few billion people at 50 cents a pop, and then quietly folds. He then points to the dot-com era as an example, arguing that most of those websites were less sophisticated than what an AI agent can produce today.

Huang was also blunt about the idea’s ceiling. “The 100,000 chances of those agents building NVIDIA,” he said flatly, “is zero percent.” That’s no small caveat. It’s a whole football game.

What Huang is really describing — a viral app that makes money briefly and then dies — is a far cry from the revolutionary, economy-fixing AGI that dominates the public conversation. So, by his own admission, the kind of integrated institutional intelligence needed to build something like NVIDIA is nowhere in the picture yet.