Zuck Is The Most Important Person This Decade
Meet the next Rockefeller.
TIME showcased its Top 100 Leaders in AI, and their cover was appropriately ripped to shreds. Jensen is on there, sure, and Lisa Su, and Sundar. But where's Sam Altman? Satya? Elon? Greg Brockman? Dario? Andrej? And what about Ilya — what did Ilya see?
There's lots of important people missing from the cover, and plenty more that are confusing. I get that ScarJo was the voice of "Her", and that she sued OpenA for their 4o voice model for using her likeness, and I get that this is emblematic of a larger set of questions around copyright law. But that's a footnote in the AI revolution, not a headline. Gosh, it's almost like TIME is using her very photogenic mug for the same reason OpenAI nabbed her voice. And MKBHD? Love his videos, but giving a fair and negative review of some early AI products feels highly tenuous.
To be fair, the criticism is really for the cover. Most of the AI heavyweights I mentioned are on the list, except for Elon (no surprise there) and Greg Brockman. And they threw in some strong lesser known figures like Dwarkesh Patel — kudos for that. On the other hand, Gina Raimondo, Scott Weiner, and Becky Pringle are all on there too... and these are not paragons of deep thinking on AI. They are the regulatory weak links in a strong link problem, looking to stamp their signature on slowing down progress in an accelerating field. So maybe they deserve to be there, just not for good reasons.
But the one I really want to talk about is Zuck. He's on the list but I think he's special. In honor of o1, here's the chain of reasoning.
1. AI is Transformational.
It's all the rage right now to say this but all the value of AI is still in the future. LLMs are taking some of the world by storm, often in surprising ways, but they are hugely capital intensive and have yet to return any of that investment beyond Nvidia investors. They probably will build value over the next couple of decades, assuming there's not a barrier to progress like the total amount of training data, but after a few years of incredible LLM progress, the insane transformations that people are casually throwing around are still not guaranteed.
Most of the commentary on AI points to superintelligence accelerating progress in vague ways we can't imagine. If this happens, then it's a fucking tsunami. The most credible and ably articiulated take I've seen on the possible future remains the beginning of Leopold Aschenbrenner's piece Situational Awareness. First he counts the OOMs and describes the unhobblings, then he describes the gate of the automated AI researcher, and finally he discusses energy use for the trillion dollar cluster.
The geopolitical consequences, according to Leopold, are enormous. The need for AI primacy could make the Manhattan Project look like a science fair volcano. If he is right, expect some form of nationalization and much more than just export control for some of Nvidia's chipsets. And, at some point, the transformational change on the order of industrialization or electrification will come.
2. Open source models will win.
...or will at least be equivalent to closed source models. Llama 3 (meta.ai let's you give it a whirl) was a step change in what was expected of open source models and it shrunk the lead of the closed source products ahead of it. It's only 6 months old now and there's been continual improvement. Claude Sonnet is amazing (and my daily go-to) and the multimodal (4o) and chain-of-thought (o1) advancements from OpenAI on top of GPT4 are indicative of the types of unhobblings we can expect in the future (as per Aschenbrenner). As training and RL techniques improve, the open source and closed source base models will converge. Meta has committed to an open source model play and they have all of the training and data resources to run that race — more on that shortly.
There may be future differences between base models, but it will be at the margins. The real improvements will sit above the base models - at the level of agentic behavior, reasoning, inference, and product construction. Both 4o and O1 use the same GPT4 as a base model now.
3. Meta has access to as much data as they need.
We know there will be a race to grab unique datasets because that race has already started. Poe has isolated Quora data for its own use and OpenAI has partnered to use the vast scribblings of degenerates that is Reddit. Zuck can match all of that with Facebook and WhatsApp. And Zuck has Instagram for video data which isn’t as good as YouTube, but still a lot. Google has a massive moat of video data in YouTube. It's arguable that YouTube is the biggest collection of possible training data for models there is, but it remains to be seen how or even if Google is willing to leverage that particular goldmine.
Suffice it to say that Meta has access to plenty of raw data.
4. Meta has the datacenter capacity for training.
As we race towards that trillion dollar cluster, there's only a finite set of places it will be able to happen. The hyperscalers are the obvious choices here, although there's some upstarts like Anthropic and even X. The Grok xAI team stood up the Colossus cluster with 100,000 H100s in Tennessee in a rather incredible 122 days. That's a couple billion dollars in training investment in one shot. And in case you weren't sure about how important energy will become, Microsoft just partnered to reopen Three Mile Island and use its nearly gigawatt of power for Microsoft datacenters.
Meta has purchased 350,000 H100s. So far. They're using these resources for training and inference both, and a lot of the inference is apparently on unknown internal tasks rather than just a bunch of users pounding tokens on a chatbot.
Meta also knows how to scale datacenters more generally. They followed the Google playbook and have 21 massive sites around the globe. They sponsor the Open Compute Project and have expanded this to AI infrastructure. They have proven their capability to scale up energy demands and manage high bandwidth. And they're clearly willing to outlay the capex required to move forward. They will continue to be a top Nvidia customer.
5. Zuck controls Meta.
You thought I was going to talk about Founder Mode here didn't you? Well, you're right. Look, I accept that the Founder Mode distinction is a binary red herring. But regardless of origin story, you still need the will, ambition, vision, desire, and control to execute. It just so happens that founders often end up with more of the necessary combination.
You could make the argument that Google parallels Meta in all the things I've described so far. They invented the transformer. They've got more datacenter capacity than Meta. They know open source. They have the data.
But they don't have a Zuck.
Look, Sundar seems like a powerful executive — he has to be to make it to where he is. But he's not Larry or Sergey. I'm not even sure Larry and Sergey are Larry and Sergei anymore. Sergei seems.. untethered rather than sharp.
And there's a whole internet literature about the complacent cultural rot that's existed inside of Google for years. I haven't seen any protests at Meta where employees stake out their bosses office. Meta has had more existential crises to deal with than any of the other behemoths. And survived them all.
Zuck is sharp. Zuck has been in the exact same game for 20 years. Zuck is the one talking about learning through suffering. Zuck has control of the board. Zuck has supervoting shares. Zuck can happily waste 10 billion on the Metaverse and then immediately pivot into the capital intensive AI race and his board says "yes sir, absolutely". Zuck is doing MMA.
Maybe that last one is not related, except to say that Zuck is still young and energetic. He has total control and he is in the game, not fucking around. If he thinks something needs doing, he's gonna do it.
6. Meta is the only hierarchy with everything under one person.
Before we do a friendly little conclusion for this, let's examine the OpenAI structure for comparison. We're going to ignore the whole non-profit/for-profit structure for now since that's been covered ad nauseum.
Sam Altman the Executive is apparently a scalpel and a bludgeon - whichever is necessary - and gets remarkable things done (including deposing of an antagonistic board and staying in power, by the way). And Satya Nadella has built a track record transforming Microsoft from the Ballmer (anti-)peak days into a modern juggernaut. These two entities have a symbiotic relationship. All the models and the algorithms and the training roll up to Altman. The dataclusters and the capital and the energy all roll up under Satya. And they kinda split the product I guess. I recognize this is a series of gross overgeneralizations with errors on the edges, but the point is: it's split.
Meta is the only place where the datacenters AND the energy AND the capital AND the data AND the training AND the software AND the algorithms AND the products all roll up in one hierarchy to one person with control, vision, ambition, will, and supervoting shares. Zuck is high agency with extraordinary resources working on the biggest question of the century. Don't be surprised if the last man standing is the guy who started out ranking hot college girls on the internet.
Now I'm not saying that just because this centralization exists, it's better. All I'm saying is that Zuck has the potential to make an outsized impact and lead a brand new technology with transformative power to the whole world. Even in the bear case, AI is going to change white collar work. And in the bull case? The world of 2050 is impossible for us to describe the same way that Ben Franklin couldn't imagine the magic of an iPhone.
This is Rockefeller level stuff.
So that's my prediction for the next decade of AI progress. There will be many Titans, just as there were during the industrialization of the late 19th century.
But one will lead them all. And his name is Zuck.