Meta invests $14.3B in Scale AI to kick-start superintelligence lab
nytimes.com464 points by RyanShook 4 days ago
464 points by RyanShook 4 days ago
The only way to understand this is by knowing: Meta already has two (!!) AI labs who are already at existential odds with one-another and both are in the process of failing spectacularly.
One (FAIR) is lead by Rob Fergus (who? exactly!) because the previous lead quit. Relatively little gossip on that one other than top AI labs have their pick of outgoing talent.
The other (GenAI) is lead by Ahmad Al-Dahle (who? exactly!) and mostly comprises of director-level rats who jumped off the RL/metaverse ship when it was clear it was gonna sink and by moving the centre of genAI gravity from Paris where a lot of llama 1 was developed to MPK where they could secure political and actual capital. They've since been caught with their pants down cheating on objective and subjective public evals and have cancelled the rest of Llama 4 and the org lead is in the process of being demoted.
Meta are paying absolute top dollar (exceeding OAI) trying to recruit superstars into GenAI and they just can't. Basically no-one is going to re-board the Titanic and report to Captain Alexandr Wang of all people. Its somewhat telling that they tried to get Koray from GDM and Mira from OAI and this was their 3rd pick. Rumoured comp for the top positions is well into the 10's of millions. The big names who are joining are likely to stay just long enough for stocks to vest and boomerang L+1 to an actual frontier lab.
I wouldn't categorize FAIR as failing. Their job is indeed fundamental research and are still a leading research lab, especially in perception and vision. See SAM2, DINOv2, V-JEPA-2, etc. The "fair" (hah) comparisons of FAIR are not to DeepMind/OAI/Anthropic, but to other publishing research labs like Google Research, NVIDIA Research, and they are doing great by that metric. It does seem that for whatever reason that FAIR resisted productization, unlike DeepMind, which is not necessarily a bad thing if you care about open research culture (see [1]). GenAI was supposed to be the "product lab" but failed for many reasons, including the ones you mentioned. Anyways, Meta does have a reputation problem that they are struggling to solve with $$ alone, but its somewhat of a category error to deem it FAIR's fault when FAIR is not a product LLM lab. Also Rob Fergus is a legit researcher; he published regularly with people like Ilya and Pushmeet (VP of Deepmind Research), just didn't get famous :P.
not affiliated with meta or fair.
[1] https://docs.google.com/document/d/1aEdTE-B6CSPPeUWYD-IgNVQV...
FAIR is failing. Dino and JEPA at least are irrelevant in this age. This is why GenAI exists. GenAI took the good people, the money, the resources and the scope. Zuck tolerates entertains ideas until he doesn’t. It’s clear blue sky research is going to be pushed even further into the background. For perception reasons you can’t fire AI researchers or disband an ai research org but it’s clear which way this is headed.
As for your comparisons, well Google Research doesn’t exist anymore (to all intents and purposes) for similar reasons.
This is why GenAI exists. GenAI took the good people, the money, the resources and the scope. Zuck tolerates entertains ideas until he doesn’t. It’s clear blue sky research is going to be pushed even further into the background
I agree with most this, I just think we have different meanings of failure. FAIR has "failed" in the eyes of Meta leadership in that they have not converted into a "frontier AI lab" like Deepmind, and as a result they are being sidelined (much like Google Research, which I admit was a bad example). But the orgs were founded to pursue basic research and I think it's not the a failure of the scientists at FAIR that management has failed to properly spin out GenAI. Of course, it sounds like your metric is "AI/LLM competitiveness" and we have no disagreements that FAIR is failing on that end (I just don't think its only important or right metric to be judging FAIR).
* Normatively, I think that it's good to have monopolistic big tech firms be funding basic open research as a counterbalance to academia and also because good basic research requires lots of compute these days. It feels somewhat shortsighted to reallocate all resources to LLM research.
* DINO and JEPA aren't particularly useful for language modeling, but are still important for embodiment/robotics/3D, which indeed seems to be the "next big thing." Also, to their credit, FAIR is still doing interesting and useful work on LLMs for encoders [1], training dynamics [2], and multimodality [3], just not training frontier models.
** GenAI took the money, scope, and resources, but not sure about the good people lol, that seems to be their problem.
[1] https://arxiv.org/abs/2504.01017 [2] https://arxiv.org/abs/2505.24832 [3] https://arxiv.org/abs/2412.14164
This is exactly why Zuck feels he needs a Sam Altman type in charge. They have the labs, the researchers, the GPUs, and unlimited cash to burn. Yet it takes more than all that to drive outcomes. Llama 4 is fine but still a distant 6th or 7th in the AI race. Everyone is too busy playing corporate politics. They need an outsider to come shake things up.
The corporate politics at Meta is the result of Zuck's own decisions. Even in big tech, Meta is (along with Amazon) rather famous for its highly political and backstabby culture.
This is because these two companies have extremely performance-review oriented cultures where results need to be proven every quarter or you're grounds for laying off.
Labs known for being innovative all share the same trait of allowing researchers to go YEARS without high impact results. But both Meta and Scale are known for being grind shops.
Can't upvote this enough. From what I saw at Meta, the idea of a high performance culture (which I generally don't have an issue with) found its ultimate form and became performance review culture. Almost every decision made filtered through "but how will this help me during the next review". If you ever wonder about some of the moves you see at Meta, perf review optimization was probably at the root of it.
I may or may not have worked there for 4 years and may or may not be able to confirm that Meta is one of the most poorly run companies I've ever seen.
They are, at best, 25-33% efficient at taking talent+money and turning it into something. Their PSC process creates the wrong incentives, they either ignore or punish the type of behavior you actually want, and talented people either leave (especially after their cliff) or are turned into mediocre performers by Meta's awful culture.
Or so I've heard.
Beyond that, the leaders at Facebook are deeply unlikeable, well beyond the leaders at Google, which is not a low bar. I know more people who reflexively ignore Facebook recruiters than who ignore recruiters from any other company. With this announcement, they have found a way to make that problem even worse.
Interesting that "high-impact" on the one hand, and innovative/successful in the marketplace on the other, should be at odds at Meta. Makes one wonder how they measure impact.
It doesn't matter much how they measure if it's empirical. Once they say the scoring system, all the work that scores well gets done, and the work that resists measurement does not get done.
The obvious example was writing eng docs. It was probably the single most helpful things you could do with your time, but there was no way to get credit because we couldn't say exactly how much time your docs might have saved others (the quantifiable impact from your work). That meant that we only ever developed a greater and greater unfilled need for docs, but it only ever got riskier and riskier to your career to try to dive into that work.
People were split on how to handle this. Some said "do the work that most needs doing and the perf review system will work it out long term." Other said, "just play the perf game to win."
I listened to the first group because I'm what you call a "believer." In a tech role I think my responsibility is primarily to users. I was let go (escorted off campus) after bottoming out a stack ranking during a half in which I did a lot of great work for the company that half (I think) but utterly failed to get a good score by the rules of the perf game (specifically I missed the deadline to land a very large PR and so most of my work for the half failed to make the key criteria for perf review: it had to be *landed* impact)
I think I took it graciously, but also I will never think of these companies as a home or a family again.
It's not that he needs a Sam Altman, but that he cannot be Sam Altman (for path-dependent reasons related to his standing in international politics),
not any advantage in virtue (or vices, for that matter)
In national politics, Sam is toe to toe with Elon,which is to say, not great, not terrible
> In national politics, Sam is toe to toe with Elon,which is to say, not great, not terrible
That’s quite the stretch, Elon is now PNG with the MAGA crowd and was already reviled by the left
PNG may not be a stretch, but so isn't the purported health of Sam's local ambitions? :)
These people should better make a lot of money while they can, because for most of them their careers may be pretty short. The half life of AI technologies is measured in months.
Meta is struggling here for the same reason Microsoft couldn’t stop the talent bleed to Google back in the day.
Even if you’re giving massive cash and stock comp, OpenAI has a lot more upside potential than Meta.
Microsoft back in the day and today still doesn’t pay top dollar. So you can’t get top talent with 65th percentile pay.
Microsoft used to have a location advantage, the Seattle area was a lot cheaper than the bay area.
They've long since lost that advantage.
This is wrong. OpenAI has almost no upside now at these valuations and there is a >2 year effective cliff on any possibility of liquidity whereas Meta is paying 7-8 figures liquid.
Metas problem is that everyone knows that it’s a dumpster fire so you will only attract people who only care about comp which is typically not the main motivation for the best people.
It's not a 2 year cliff: it's 6 months before vesting, then 2 years before you can sell.
Effective cliff. What use is vested “equity” (ppus aren’t even equity) that you cannot sell?
It means that you can keep those shares even if you leave. Otherwise the term vesting cliff would be meaningless at any startup where the shares are not liquid.
They are yours. That’s a huge difference between a real cliff and illiquid stock.
If you decide you don’t like it, you take what’s vested after the cliff and leave. Even if you have to wait another year and a half to sell, you still got the gain.
Massive difference. You can vest and move on, even if you don’t have liquidity, which most private companies don’t for employees anyway.
Except you can only sell a prescribed amount at an undetermined time. By the earliest possible sell date you have already made 8 figures liquid at Meta.
Ok, but that’s a trade off anyone who works for a private company makes, and it’s never called an “effective cliff”.
You forgot LeCunn, but yeah that guy's on its own death spiral.
No I didn’t. He is functionally irrelevant at Meta and he doesn’t actually lead anything.
Weird, Meta says it's their Chief AI Scientist [1].
But maybe they're wrong ...
You really don't understand that what is advertised on a "people" page can be different from what the person actually does?
FYI if you worked at FB you could pull up his WP and see he does absolutely nothing all day except link to arxiv.
Cool. So what does a chief AI scientist do?
ideally lead AI science, but in reality mostly pontificate on social media. One could say that is fitting for Meta though right?
Anyone know what scale does these days beyond labeling tools that would make them this interesting to Meta? Data labeling tools seem more of a traditional software application and not much to do with AI models themselves that would be somewhat easily replicated, but guessing my impression is out of date. Also now apparently their CEO is leaving [1], so the idea that they were super impressed with him doesn't seem to be the explanation.
[1] https://techcrunch.com/2025/06/13/scale-ai-confirms-signific...
Scale has also built massive amounts of proprietary datasets that they license to the big players in training.
Meta, Google, OpenAI, Anthropic, etc. all use Scale data in training.
So, the play I’m guessing is to shut that tap off for everyone else now, and double down on using Scale to generate more proprietary datasets.
OpenAI and Anthropic rely on multiple data vendors for their models so that no outside company is aware of how they train their proprietary models. Forbes reported the other day that OpenAI had been winding down their usage of Scale data: https://www.forbes.com/sites/richardnieva/2025/06/12/scale-a...
And scale doesn’t even have the best data among these vendors so I also don’t get this argument
Yeah, but they know how to get the quality human labeled data at scale better than anyone — and they know what Anthropic and OpenAI wanted — what made it quality
I wondered that.
But then huge revenue streams for Scale basically disappear immediately.
Is it worth Meta spending all that money just to stop competitors using Scale? There are competitors who I am sure would be very eager to get the money from Google, OpenAI, Anthropic etc that was previously going to Scale. So Meta spends all that money for basically nothing because the competitors will just fill the gap if Scale is turned-down.
I am guessing they are just buying stuff to try to be more "vertically integrated" or whatever (remember that Facebook recently got caught pirating books etc).
Yeah, also the industry could come up with their own Scale if they were forced to.
But probs. it just makes sense on paper, Scale's revenue will pay this for itself and what they could do is to give/keep the best training sets for Meta, for "free" now.
Zuck's not an idiot. The Instagram and WhatsApp acquisitions were phenomenal in hindsight.
> Yeah, also the industry could come up with their own Scale if they were forced to
I worked at Outlier and it was such a garbage treatment
is he not?
what about the whole metaverse thing and renaming the whole company to meta?
The metaverse will happen, IMO. The tech is just not there, yet.
Even if it turns out to be wasted money, which I doubt, he's still sitting on almost 2 trillion. Not an L on my book.
> The metaverse will happen, IMO. The tech is just not there, yet.
This seems possible, and it just sounds so awful to me. Think about the changes to the human condition that arose from the smartphone.
People at concerts and other events scrolling phones, parents missing their children growing up while scrolling their phones. Me, "watching" a movie, scrolling my phone.
VR/AR makes all that sound like a walk in the park.
“We went outside this weekend. Terrible. I wasn’t hot anymore, the smog was everywhere. House was tiny. No AI to help with conversations and people were unfriendly. I’m staying plugged in, where we can fly amongst the stars on unicorns. Some say it’s fake but I say life has been fake for a while.”
Meta has done great work on the underlying technology of the metaverse, but what they really need is a killer app. And I don't think Meta or really Silicon Valley types have the proper institutional ability or really cultural acumen to achieve it. We think back to Horizon Worlds that looked more like a amateur weekend asset flip than the product of a billion dollar conglomerate.
If it does come, it will likely come from the gaming industry, building upon the ideas of former mmorpgs and "social" games like Pokemon Go. But recent string of AAA disasters should obviously tell you that building a good game is often orthogonal to the amount of funding or technical engineering. It's creativity, and artistic passion, and that's something that someone who spends their entire life in the real world optimizing their TC for is going to find hard to understand.
Meta buys 900 AI employees here at less than $20M/head. Pretty cheap these days. Any IP the company has is a bonus.
850 people doing data labelling?
Nah - I think of Scale as "like Uber but for AI data". I.e. they match humans to AI companies wanting to do RLHF.
Could be they need a perpetual license to the data too… so they could potentially sue everyone.
Meta “will have a 49% stake in the artificial intelligence startup, but will not have any voting power”
Wouldn’t Scale’s board/execs still have a fiduciary duty to existing shareholders, not just Meta?
The prevailing theory is that Meta did a 49% deal so it didn't set off anti-trust alarm bells. In other words, the 49% doesn't give them ultimate power, but you can best believe when Meta tells them to jump, the board and the execs are going to ask "how high?".
Yes, but Meta would be able to kick the board and find another one more willing to accept their proposal as the best for shareholders.
Power struggles like this are weird to me. Is kicking the board likely to succeed at 49%? If so it feels like the control percentage isn't the primary factor in actual control.
At 49% I'm certain they would become the largest shareholder, by far. Then allying with another smaller shareholder to get majority - especially as you are Meta and can repay in various ways - is trivial. This is control, in all forms but name.
another shareholder, one of the largest, is the now-former-CEO who now works at Meta. They have full control.
There's a lot of things shareholders can do to screw over other shareholders. Smaller shareholders are at least somewhat likely to follow along with the largest shareholder, just to avoid becoming their enemies and getting squeezed out.
It's a smart purchase, it's just that I don't see how these datasets factor into super-intelligence. I don't think you can create a super-intelligent AI with more human data, even if it's high-quality data from paid human contributors.
Unless we watered-down the definition of super-intelligent AI. To me, super-intelligence means an AI that has an intelligence that dwarfs anything theoretically possible from a human mind. Borderline God-like. I've noticed that some people have referred to super-intelligent AI as simply AI that's about as intelligent as Albert Einstein in effectively all domains. In the latter case, maybe you could get there with a lot of very, very good data, but it's also still a leap of imagination for me.
I think this is kind of a philosphical distinction to a lot of people: the assumption is that a computer that can reason like a smart person but still runs at the speed of a computer would appear superintelligent to us. Speed is already the way we distinguish supercomputers from normal ones.
I'd say superintelligence is more about producing deeper insight, making more abstract links across domains, and advancing the frontiers of knowledge than about doing stuff faster. Thinking speed correlates with intelligence to some extent, but at the higher end the distinction between speed and quality becomes clear.
If anything, "abstract links across domains" is the one area where even very low intelligence AI's will still have an edge, simply because any AI trained on general text has "learned" a whole lot of random knowledge about lots of different domains; more than any human could easily acquire. But again, this is true of AI's no matter how "smart" they are. Not related to any "super intelligence" specifically.
Similarly, "deeper insight" may be surfaced occasionally simply by making a low-intelligence AI 'think' for longer, but this is not something you can count on under any circumstances, which is what you may well expect from something that's claimed to be "super intelligent".
I don't think current models are capable of making abstract links across domains. They can latch onto superficial similarities, but I have yet to see an instance of a model making an unexpected and useful analogy. It's a high bar, but I think that's fair for declaring superintelligence.
In general, I agree that these models are in some sense extremely knowledgeable, which suggests they are ripe for producing productive analogies if only we can figure out what they're missing compared to human-style thinking. Part of what makes it difficult to evaluate the abilities of these models is that they are wildly superhuman in some ways and quite dumb in others.
I think they can make abstract links across domains.
Like the prompt "How can a simplicial complex be used in the creation of black metal guitar music?" https://chatgpt.com/share/684d52c0-bffc-8004-84ac-95d55f7bdc...
It is really more of a value judgement of the utility of the answer to a human.
Some kind of automated discovery across all domain pairs for something that a human finds utility in the answer seems almost like the definition of an intractable problem.
Superintelligence just seems like marketing to me in this context. As if AGI is so 2024.
> It's a high bar, but I think that's fair for declaring superintelligence.
I have to disagree because the distinction between "superficial similarities" and genuinely "useful" analogies is pretty clearly one of degree. Spend enough time and effort asking even a low-intelligence AI about "dumb" similarities, and it'll eventually hit a new and perhaps "useful" analogy simply as a matter of luck. This becomes even easier if you can provide the AI with a lot of "context" input, which is something that models have been improving at. But either way it's not superintelligent or superhuman, just part of the general 'wild' weirdness of AI's as a whole.
I think you misunderstood what I meant about setting a high bar. First, passing the bar is a necessary but not sufficient condition for superintelligence. Secondly, by "fair for" I meant it's fair to set a high bar, not that this particular bar is the one fair bar for measuring intelligence. It's obvious that usefulness of an analogy generator is a matter of degree. Eg, a uniform random string generator is guaranteed to produce all possible insightful analogies, but would not be considered useful or intelligent.
I think you're basically agreeing with me. Ie, current models are not superintelligent. Even though they can "think" super fast, they don't pass a minimum bar of producing novel and useful connections between domains without significant human intervention. And, our evaluation of their abilities is clouded by the way in which their intelligence differs from our own.
I don't know about "useful" but this answer from o3-pro was nicely-inspired, I thought: https://chatgpt.com/share/684c805d-ef08-800b-b725-970561aaf5...
I wonder if the comparison is actually original.
Comparing the process of research to tending a garden or raising children is fairly common. This is an iteration on that theme. One thing I find interesting about this analogy is that there's a strong sense of the model's autoregressiveness here in that the model commits early to the gardening analogy and then finds a way to make it work (more or less).
The sorts of useful analogies I was mostly talking about are those that appear in scientific research involving actionable technical details. Eg, diffusion models came about when folks with a background in statistical physics saw some connections between the math for variational autoencoders and the math for non-equilibrium thermodynamics. Guided by this connection, they decided to train models to generate data by learning to invert a diffusion process that gradually transforms complexly structured data into a much simpler distribution -- in this case, a basic multidimensional Gaussian.
I feel like these sorts of technical analogies are harder to stumble on than more common "linguistic" analogies. The latter can be useful tools for thinking, but tend to require some post-hoc interpretation and hand waving before they produce any actionable insight. The former are more direct bridges between domains that allow direct transfer of knowledge about one class of problems to another.