Meta's $10B+ AI Reset
For months and months, all we've heard is how much Meta's AI is "killing it." Or "crushing" in bro parlance. They were saying it so much, so often, to anyone who would listen that it was clearly a bit suspect. Or "sus" in 'how do you do, fellow kids' slang. "The best models!" Even when (non-manipulated) data suggested otherwise, as did delays. "A billion users!" Said with what was undoubtedly the backing of growth hacked data that would make even the most die-hard Google+ PM blush.
As it turns out, that crushing was sus.
You don't need actual data to know this, you just need to watch what Meta is actually doing. Which is scrambling to get their AI house in order. Which you typically don't do when your technology and products are "slaying"...
Case in point, it was interesting when news broke over the weekend that Meta was planning for a massive investment into Scale AI. Not only does Meta typically not do those kind of investments, the amount being reported – $10B+ – is outlandish even for a company as profitable as they are. For context, when Microsoft invested in Meta (then called Facebook) back in the day, they put in $240M. That money valued the social network at $15B.1 Times have changed.
Such a deal was seemingly more akin to Microsoft's investment in OpenAI. But Scale AI is a very different company than OpenAI, their focus is not on actually building foundation models, but on facilitating those builds, augmenting the data, and other adjacencies. It seems sort of like a weird business for Meta to be backing given how many fingers they have in other models (many by their rivals). And if Meta wanted the business as a way to diversify from their all-important advertising revenue, why not buy it versus investing in it? Well, besides the fact that they undoubtedly would not be allowed to...
Today, the other shoe dropped. Here's Cade Metz and Mike Isaac for The New York Times:
Meta is preparing to unveil a new artificial intelligence research lab dedicated to pursuing “superintelligence,” a hypothetical A.I. system that exceeds the powers of the human brain, as the tech giant jockeys to stay competitive in the technology race, according to four people with the knowledge of the company’s plans.
Meta has tapped Alexandr Wang, 28, the founder and chief executive of the A.I. start-up Scale AI, to join the new lab, the people said, and has been in talks to invest billions of dollars in his company as part of a deal that would also bring other Scale employees to the company. Meta has offered seven- to nine-figure compensation packages to dozens of researchers from leading A.I. companies such as OpenAI and Google, with some agreeing to join, according to the people.
Forget about a "hackquisition", this is... I can't even come up with a quip. It's just sort of wild. It sure sounds like Meta, at the direction of Zuckerberg, is potentially investing an eleven-figure sum into Scale in order to gain access to their CEO and some of their talent for this new "lab".
So Meta is sort of buying Scale's talent but seemingly not actually the company – which is what all of these would-be but can't-be acquirers say these days. Only here, I think it's probably true. This is decidedly not like those other strange deals where a startup is paid out, stripped of their talent, and left "running" with someone there to answer a phone for when the government calls to inquire about the deal. This is perhaps more akin to what Google did with Character.ai to bring Noam Shazeer back in house, but at an even bigger scale – because Scale is a far bigger business than Character.ai is and was. Unlike, say Meta's AI business, Scale is a business that by all accounts is actually "crushing". So this money, beyond what is presumably going to Wang himself, will undoubtedly be used to actually keep going. [See update on the money below]
At the same time, there have long been questions about where a company like Scale ends up in the AI revolution. They've already had to morph and shift multiple times to expand from data labeling to doing more direct work to evaluate and feed models. And those shifts will undoubtedly continue as AI continues to change and mature, with more of what Scale does going in-house at the large companies. And Wang started Scale nearly a decade ago when AI as we know it today wasn't yet a thing, it was all about computer vision back then.
This also doesn't seem as straightforward as some of the other "hackquisitions" since again, Scale isn't actually building LLMs, so what type of talent is Meta actually scooping up here? Beyond very expensive talent, clearly!
Also, will Meta get some sort of priority access to use Scale for their own models? Again, unclear. But that's not actually the point here, now is it? If it ends up as a good and helpful investment for Meta, that will simply be very expensive icing on the multi-billion-pound cake.
More on this "lab" from Kurt Wagner and Riley Griffin for Bloomberg:
Mark Zuckerberg, frustrated with Meta Platforms Inc.’s shortfalls in AI, is assembling a team of experts to achieve artificial general intelligence, recruiting from a brain trust of AI researchers and engineers who’ve met with him in recent weeks at his homes in Lake Tahoe and Palo Alto.
Zuckerberg has prioritized recruiting for the secretive new team, referred to internally as a superintelligence group, according to people familiar with his plans. He has an audacious goal in mind, these people said. In his view, Meta can and should outstrip other tech companies in achieving what’s known as artificial general intelligence or AGI, the notion that machines can perform as well as humans at many tasks. Once Meta reaches that milestone, it could weave the capability into its suite of products — not just social media and communications platforms, but also a range of AI tools, including the Meta chatbot and its AI-powered Ray-Ban glasses.
Zuckerberg aims to hire around 50 people for the new team, including a new head of AI research, almost all of whom he’s recruiting personally. He’s rearranged desks at the company’s Menlo Park headquarters so the new staff will sit near him, the people said, asking to remain anonymous discussing private plans.
This almost sounds as if Zuckerberg is Steve Jobs in the early 1980s creating the Macintosh team within Apple.2 Such a comparison would flatter anyone, but it does read like that's essentially what Zuckerberg is trying to create here after Meta's initial attempts at AI haven't really panned out. Are there going to be pirate flags?3
Zuckerberg has spoken openly about making artificial intelligence a priority for his company. In the last two months, he’s gone into “founder mode,” according to people familiar with his work, who described an increasingly hands-on management style.
The CEO’s desire to micromanage the recruitment effort is driven in part by frustration over the quality of and response to Llama 4, the latest version of Meta’s large language model to power chatbots and other services.
Uh oh, you now things are serious when someone busts out "founder mode". Back in the day, we used to know such Zuckerberg moments as "Lockdown". But again, times change.
Those missteps pushed Zuckerberg to get more involved and led to his interest in building out the new team, according to people familiar with the matter. He seeded a WhatsApp group chat among senior leaders called “Recruiting Party” to discuss potential targets for the team. Members of the chat group have been engaged in discussions at all hours of the day to identify talent.
Zuckerberg has been compiling his own list of recruits and likes to be the first point of contact during outreach. The CEO hopes that with the new bench, Meta will see improvements to its Llama models and better AI tools for voice and personalization features, the people said.
It’s not yet clear how the Superintelligence Labs group will work alongside Meta’s existing AI teams. Some employees are expected to move over to the new unit, according to people familiar with the plans.
Also of note: Zuckerberg sure seems to be changing his tone quite a bit when it comes to AGI. Whereas he used to be quick to dismiss such talk as other searching for their "AI God" and that there wouldn't be one model to rule them all, now he's... getting into that race, it seems.
As for where that leaves their commitment to "open source" (read: open weight), it's not clear from these posts. Nor is it clear if Yann LeCun, the Turing Award winner who has long overseen Meta's AI efforts (ever since Facebook missed out on acquiring DeepMind to Google back in the day) is fully on board with this as he's famously been quite skeptical of the approaches that the other AI labs are taking, at least as it related to AGI – or sorry, "Superintelligence" or whatever you want to call it.
Will Wang be at this new lab under LeCun? (It seems like no?) Or is this a full Meta AI reboot? Or actually, more of a starting from scratch, which is what it sounds like? Which again, is not something you typically do if you've been "killing it". Which won't do Meta any favors in the outside world trusting in what they're conveying. But if they win the race to AGI/Superintelligence/Whatever, I guess that won't really matter.
Will investors like this deal? Right now, with the news breaking, the stock is little-changed in after-hours. They're probably still parsing this as it is a bit weird of a deal. Meta is undoubtedly fine with that, especially if it means any government officials don't quite know what to make of it either. Certainly it doesn't seem like Meta will be easing up on the CapEx anytime soon. And they may actually need to increase it in order for this new team to catch up. Get ready for more ads in feed!
Update: More details on the deal from Cory Weinberg at The Information, including that the amount is set to be $14.8B paid to Scale shareholders for Meta to buy a 49% stake -- sound familiar? -- in the company and bring Wang and some of his team over to operate within Meta's new AI team.
1 By the way, perhaps the best deal Steve Ballmer ever did. Meta is now valued at $1.75T. Of course, Microsoft sold their 1.6% starting back at Facebook's IPO back in 2012, and by 2014 they had fully exited. Today those shares, again, bought for $240M, would be worth $27.84B.
2 A team which, yes, Jobs actually and famously co-opted from Jef Raskin.
3 To be honest, Apple could probably use some of that spirit again...