Meta’s AI Relativity Theory

To hear Mark Zuckerberg tell it, it’s all relative. Seriously, basically everything. At least when it comes to Meta and AI. That was the takeaway from his brief chat with Jessica Lessin on The Information’s inaugural TITV broadcast.

After kicking off with his clearly prepared talking points, Lessin asked him about the Llama 4 situation — i.e. the model that launched a thousand hackquihires — and thus began Zuck’s theories of relativity. You see, it's not that Llama was a mess, it’s that the rest of the field was “accelerating” and “moving faster”, relatively speaking.

And they had to act because “superintelligence” is going to be here relatively soon. While others debate if it’s going to be here in three, five, or seven years, Zuck is more bullish:

"I just think that we should bet and act as if it’s going to be ready in the next two to three years. I believe there’s a shot at that. And if that is what you believe, you are going to invest hundreds of billions, for hours, whatever it takes to build out the strongest team."

Whatever it takes. That, of course, leads to the question about the massive pay packages Meta has been dishing out. The “Godfather” deals and those being accepted for mere seven or eight figures. Zuck was sure to reiterate that what has been reported on such deals “aren’t accurate by themselves” — which reads a lot like a non-denial denial on the grounds of some semantics; aka, the standard Meta playbook — but then he immediately goes on to seemingly confirm at least the directional accuracy of such reports.

"But it is a very hot market. I mean, as you know, and there’s a small number of researchers, which are the best, who are in demand by all of the different labs. So I think that it certainly is quite competitive."

Anyway, back to relativity:

"You know what I’d say. I think in perspective, if you’re going to be spending hundreds of billions of dollars on compute and building out multiple gigawatt of clusters, then, you know, it really does make sense to compete super hard and do whatever it takes to get that, you know, 50 or 70 or whatever it is, top researchers to build your team."

Yes, that’s right, if you’re spending hundreds of billions of dollars on data centers, what’s a few hundred million on people? I mean, he’s not wrong. But that people market also happens to be relative! And compared to what others had been getting paid before the Meta offers started rolling in... well, it certainly reset the market.

At the same time, maybe that’s the right way to look at and act upon the situation! What’s the point of spending those hundreds of billions if you can’t leverage all the compute to the greatest extent possible? And that takes the best people building the best models to leverage that infrastructure. And that takes money. A lot of money relatively speaking for people. Because Meta currently doesn’t employ most of those people. And they’re coming from behind here.

"I think that the physics of this is you don’t need a massive team to do this. You actually kind of want the smallest group of people who can fit the whole thing in their head. So there’s just an absolute premium for the best and most talented people."

Yes, relative to what you might need for other types of jobs and projects, Zuck thinks you don’t need a big team to make this all work. And that makes the comp packages for a relatively small amount of people all the more doable.

"So it kind of makes sense when you kind of think about it that, from that perspective, the amount that is being spent to recruit the people is actually still quite small compared to the overall investment and all when you talk about super intelligence."

Again, he’s not wrong. But he’s also floating right past the relevant and relative market before he started doing these deals...

"I think personal superintelligence is a different vision than what the others are going for. It is kind of like when the internet was first getting to scale, people were like, all right, is the internet going to be for productivity?"

“Superintelligence” too, as it turns out, may be relative! It’s not just a one-size-fits all thing in Zuck’s mind. There will be many facets of it. And the aspect Meta is going to focus on is “personal superintelligence”. Yes, for those keeping score at home, “AGI” has more or less been rebranded as “superintelligence” (though yes, many people argue they’re different, the fact that no one has an actual definition for AGI means you simply cannot argue this either way) and now Zuck is trying to brand subsections of the rebrand, well before we’ve achieved any of it, naturally.

Now on the to the translation round:

"But I think the products that we’re going to build are going to be, are going to be pretty different, and I think are just going to touch billions of people’s lives.

We’re kind of the premier company at developing products and scaling them to billions of people"

Translation: relative to our peers, we’re uniquely positioned to get such technology in the hands of far more people than anyone else. Again, not wrong, but that was also true with Llama and Meta AI to date. In fact, Zuck keeps touting the scale that Meta's AI products are currently at. It apparently did matter — at least not enough to avoid having to reboot the entire team and effort to start anew.

"And we’re, of course, leading in the next generation of computing platforms with the glasses that are doing exceptionally well.

And I continue to think that that’s going to be the best form factor for AI, because, you know, they can see what you see and hear what you hear, and you can talk to them throughout the day."

Translation: relative to everyone else, we’re also best-positioned with regard to devices to actually use the AI. That may end up true, but it’s such early days of all this that it’s impossible to know that right now. To be fair, Zuck does couch it all a bit with “think”.

"And it’s, it’s almost gonna be like, look, if you need vision correction and you don’t have glasses, just optical glasses, you’re sort of at this cognitive disadvantage.

I think in the future, if you don’t have AI glasses, you’re basically going to be at a cognitive disadvantage too."

Translation: without our glasses, you’re going to be a Neanderthal relative to other human beings.

As someone who already wears glasses, I'm naturally fine with this world. But will the entire world be fine needing to wear glasses all the time? After all, only a literal idiot would not in this future hypothetical world.

"And one of the benefits of reinforcement learning is it gives you a venue to, you know, potentially convert very large amounts of capital into a better and better service, and potentially a better service than other less well-funded or less bold competitors will be able to do so.

I view that as a competitive advantage. If we can get this to work well, and that’s why we are basically all in on this. We’re building, you know, we’re building multiple, multigigawatt data centers, and we can basically do this all funded from the cash flow of the company. We have a very strong business model that supports this. We have a very strong infrastructure team that is doing novel work to build up data centers."

Translation: Relative to our peers, we’re printing cash. Every day. Every quarter. Every year. We can pour this back into everything we’re doing here. And because we have an experienced infrastructure team, we can do this even more efficiently compared to pretty much everyone else.

This is obviously more of a direct shot at OpenAI — and to a lesser extent, Anthropic and the other AI startups. Because, of course, Google is in the same, and arguably better position here with regard to cash (certainly in that their revenue base is far more diversified than Meta’s) and infrastructure. Ditto for Microsoft. And Amazon.

"You know, I wanted them to not just take four years to build these concrete buildings.

So we pioneered this new method where we’re basically building up these weatherproof tents and building out the networks and the GPU clusters inside them in order to build them faster. They are hurricane-proof tents."

Translation: Relative to building actual data centers, we’re going to move far faster. With data… tents!

I mean, okay! The question is how fast they’ll be able to move relative to how fast Elon Musk spun up “Colossus” — are Titan names also relative? Meta/Facebook was the first “move fast and break things” company — are they still willing to do that to the extent that it extends into the real world with these data center builds?

"We have people working around the clock on Prometheus and Hyperion, which are our first two Titan clusters that are going to be greater than a gigawatt.

Hyperion is going to scale up to five gigawatts over the coming years. And I shared this image of it, but the size of the site covers a significant portion of the footprint of Manhattan in terms of space. That thing is just massive."

A literal relative example! These data center footprints will be nearly the size of Manhattan! (I’m going to assume the data centers are not actually going to be 2,000-feet tall as the comparison image implies, but I digress...)

"You have a large amount of the energy—of people’s energy—in the area where it is getting built, going towards constructing it. It’s great for the community."

That is certainly a relative statement. It’s great for the community compared to no money being spent perhaps, but what about relative to building, I don’t know, a park? Or a zoo? Or something else that’s not a massive power-sucking data center?

"I’m very excited that I think we’re going to have the largest compute fleet of any company, and focusing on that on being powered by a small and talent-dense team, I think we’re gonna have by far the most compute per researcher to do leading edge work. And I think that it’s just gonna be a really exciting, big few years coming."

“Compute per researcher” definitely sounds like a compelling pitch relative to their peers. For now at least. But the whole "open source AI" thing was also supposed to be Meta's killer pitch. It didn't really land, it seems...

"I’m obviously spending a lot of time on recruiting and a lot has been written about money and all that. Like I said before, a lot of the numbers specifically have been inaccurate, but I think it discounts the other key reasons why people are super excited to come work on Meta Superintelligence Labs.

And one of the biggest is that you can just have more leverage as a researcher. You have more compute right? I mean, basically historically, when I was recruiting people to different parts of the company, you know, people are like, OK, what’s my scope going to be?

And, you know, here, people say, I want the fewest number of people reporting to me and the most GPUs. And so having basically the most compute per researcher is definitely a strategic advantage, not just for doing the work, but for attracting the best people.

And I actually think that has been underreported as one of the main reasons why people are excited to come do this. The other is just having a clean slate is fun, right? I mean, a lot of times people refer to building up the infrastructure for these things."

The "clean slate" point is fair enough. Though it's a classic attempt to turn a weakness – needing to start over – into a strength. But the problem with the "leverage" point is that it will obviously also just turn into a race. Zuck would say – and is saying, indirectly, at least – that OpenAI, for all their fundraising prowess, can't match what Meta can spend here. But again, Google certainly can! So can Amazon! Microsoft may be pulling back a bit from that particular race, but if they start clearly losing to those rivals, they have the capital as well, clearly. Oh yes, and Elon. xAI may not have all that capital, but Elon's conglomerate of companies will figure out a way to outspend Zuck here if he needs to, one imagines...

And that leaves... paying more in comp. Which remains the biggest question here. With Meta clearly resetting this market, will the others step up to match these offers? OpenAI is said to be in some cases. It seems like Anthropic may be doing so as well in others. Google? Again, if Meta can afford it, Google certainly can. But will their comp committees make the same determination about relative spend here? Is one desperate actor resetting the market enough?

Remember, that actor also has something the other big, public players don't: founder control. So maybe it's less about whether or not they'll call Zuck's raise, but if they can afford to – not literally, which they can, but in the eyes of the market? Or the flipside might also be true: the market may determine they can't afford not to...

One more thing: the way Zuck ends the chat is fun...

"Someone gave me this analogy that it’s like, it’s like building a house. The first one you build for your enemy, the second one you build for your friend, the third one you build for yourself. And you know, I think the ability to start from a clean slate, you know, where you’re not just at another company working on some other infrastructure, and you get the opportunity to build a new lab from scratch. It’s going to be, I think it’s just really appealing to a lot of people. Anyhow, yeah, I’m out there very focused on this."

Quite the analogy. Essentially: “All your work today has been relative shit.” A newfangled "Do you want to sell sugar water for the rest of your life, or do you want to come with me and change the world?" Some might say that's a hard sale for Meta to make. But coming from Zuck directly... with millions and millions of dollars attached... It's all relative.