Chapter 5: Jensen Huang declares AGI has arrived

Ah, interesting. Are you a humanist or a speciesist? The same conversation that Elon had with Larry Page.

All right, I'm going to move us on. Uh, our next series of stories here focus on the accelerating conversation around AGI that we're having and that the world is having. So Jensen Huang posted this week that OpenAI trained GPT-6 Astra on more than 100,000 Nvidia Gracewell — Blackwell GPUs, uh, and posted three magic words: AGI has arrived. Congratulations to OpenAI.

So Jensen is calling it uh in Q3 of 2026. Uh, remember on September 1st, we reported Sam Altman expects AGI internally by the end of 2026. And of course, Alex, you've been saying that we've had AGI for what, four years now?

Three years now. And no — since summer of 2020 at the latest.

Okay. All right. And Salim, you've been asking the question, what the heck is AGI anyway?

Why are we even talking about this, right? It's a semantic argument.

Meanwhile, there's economic capability running rampant. If an AI can perform 70 or 80 or 90% of economically valuable cognitive tasks, whether you call it AGI or not becomes completely irrelevant.

I'll just remind people that at last count there were 14 different definitions of AGI and we have no idea what the hell we're talking about, because it's not artificial, it's not really general, and it's not really intelligence. Apart from that, everything is fine. I mean, for God's sakes, I think the more important question is, what scarcities are we now making abundant? Let's just focus on that.

Emad, what do you make of Jensen's proclamation?

I mean, again, for his — his one is a very functionalist one, and it's actually good intelligence, shall we say. I mean, it's clear that Astra is kind of at that level. As to ASI and [unclear], well, maybe we're getting there the next day or two. Um, it is a question of kind of these things. I think you also notice there's 100,000 chips that was trained on. That's a billion dollar training run roughly over 2 months. And the next one is 400,000 chips with those of Vera Rubins. So it's an order of magnitude more compute will be used for the next lot.

Interesting.

If it needs to be used at all, which we'll get to in a bit.

And Dave, what do you make of it? Is this uh — is this just marketing on Jensen's behalf or does he really believe it?

Uh, no, of course he believes it and I think it's real. There's going to be some capability leap. We don't quite know — the chinchilla laws, you know, don't necessarily hold up at larger and larger scales — but some capability leap that in my opinion is only a good thing. Uh, as long as it's contained and kept inside the big labs. Um, but I think that uh, you know, the more GPUs you throw at these all the way up until today, the more GPUs you throw at these training runs, the more spectacular the resulting model has been. Why would that end? I don't think it will end. I think it just gets bigger, smarter, better, more helpful, solves diseases, cures problems. You just have to really be thoughtful about how to contain it. But, but you know, this is Jensen's wheelhouse, too. I think inference is moving off of Nvidia inevitably, but training is not. And so to the extent that we keep getting improvements, it keeps driving Nvidia stock up and up and up.

Yeah. There are two other stories to hit on this AGI theme. Um, the first story: OpenAI's internal data says that AI research agents are now completing 3.1 days of research work for every one day done by a human researcher. Earlier in this year — you know, in this year, like in the last 5 months — agents were doing less than one day of work as compared to humans. So OpenAI is saying that we've now crossed the line and current systems are exceeding AI research interns.

The second story I'll just put forward, uh, and then we'll talk about it, is a statement by OpenAI's engineering lead for Codex. This is Thibault Sottiaux, who posted the following, and let me read from uh from his comments. He said, Astra was probably our biggest competitive advantage while it wasn't generally available. Right? While they had it internally, it was our biggest competitive advantage. Since we've had it, our productivity jumped so much that we've shifted some of our plans 6 months ahead and we'll ship them at DevDay instead of mid next year.

So, I mean, Dave, you know, 6 months of roadmap being pulled forward by one model.

I mean, this is without a doubt the most important moment in human history. And I truly believe that they're not lying. You know, a lot of naysayers will say, look, they're trying to promote their capabilities in advance of an IPO, blah, blah. It's not true. If you talk to the actual researchers working in the labs, many of whom are friends of ours, there's no doubt that those numbers are exactly right and accelerating, though. Just a few months ago, it wasn't true. Now, with the Fable 5.1 and Astra models, it's absolutely 3 to 1. Sure. But it's on the tipping point of being 300, 3,000, 3 million to one imminently. This is the moment in time that's most important in societal history.

Yeah. Um, Alex, your thought on this. I mean, we've talked about the fact that the frontier labs have the best models they're retaining for their own use. Here you hear uh, you know, the head of Codex, you know, making that statement. And how many models ahead are they beyond uh, beyond Astra?

They're only a few months ahead, I think. But look, recursive self-improvement is here. Please, please, please let me say something about Navier-Stokes, because that is —

Okay. Okay, we can — we can jump —

I'm out of order, Peter. We have to — we can't —

You're so excited. So — so everybody, you know, on text this morning, I'm getting like red alerts from Alex. Yeah, go for it, Alex.