To The Moon

Jensen Huang put one of the biggest phrases in AI back into circulation this week. On Lex Fridman’s podcast, after Fridman described AGI, artificial general intelligence, as a system able to start, grow, and run a company worth more than $1 billion, Huang replied, “I think it’s now. I think we’ve achieved AGI.” That line travelled fast because it sounded final. But the fuller exchange was narrower than the headline suggested.

The qualifier came immediately after the claim. Huang did not point to a system that could build and sustain a company like Nvidia. He gave a smaller example instead: an AI system creating a simple web service or app, going viral for a short period, charging a small fee, and then fading. He later added that the odds of 100,000 of those agents building a company like Nvidia were effectively zero. Read plainly, that is not a clean declaration that full AGI has arrived. It is closer to saying that current systems are showing flashes of broader capability.

That is the right way to read the moment. AI is getting strong enough for top industry figures to speak in bigger terms. But broad performance in bursts is not the same as stable performance across many kinds of work. A paper makes that point clearly. Its argument is simple: strong results on known tasks do not prove general intelligence, because real intelligence shows up when a system can learn and adapt under new conditions rather than repeat patterns it has already absorbed.

That is why reliability sits at the centre of this debate. The issue is not whether today’s systems can produce moments that look surprisingly broad. They can. The issue is whether those moments are dependable, sustainable, and consistent. For business leaders, that is the difference between a system that looks impressive in a demo and one that can be trusted in real operations, where inconsistency creates delay, rework, and risk. Huang’s own example pointed to a short burst of success, not durable execution.

The next constraint in the argument is not only intelligence. It is the lack of a body. Even if AI keeps improving at writing, reasoning, analysis, and planning, it still cannot directly move through the world, pick things up, navigate a factory floor, or carry out physical tasks on its own unless that intelligence is placed into a machine. That is why the discussion moves so quickly from AGI to ASI, artificial superintelligence, and then to robots.

That line of thinking is no longer science fiction. A 2023 paper showed that a model trained on internet-scale language and image data could improve robot control, handle unfamiliar objects better, follow commands it had not seen during robot training, and perform simple reasoning while carrying out physical tasks. That does not prove human-level robots are here. But it does show that the gap between digital intelligence and physical action is being worked on directly.

The same point appears in a 2025 filing. In a 2025 Tesla proxy filing and related remarks, Elon Musk said he would not feel comfortable overseeing what he called a “robot army” unless he had at least a “strong influence” over it. The language is dramatic, but the underlying signal is plain. Once advanced intelligence is discussed as something that can sit inside a machine and act in the real world, the conversation stops being only about software. It becomes a discussion about deployment and what capable machines can actually do once they leave the lab.

So the clean reading of this moment is straightforward. We do not have clear proof that full AGI has arrived in a reliable and repeatable form. But we may be seeing the first credible glimpse of something broader than narrow AI systems built for one task at a time. Leaders should take the signal seriously without treating the headline as settled fact. The sequence is becoming clearer: first broader reasoning, then dependable performance, then the question of what happens when that intelligence gets a body.