Mark Zuckerberg — Llama 3, $10B models, Caesar Augustus, & 1 GW datacenters
What this covers
Dwarkesh Patel interviews Mark Zuckerberg on Meta's AI strategy, focusing on why the company is investing billions in frontier models and committing to open weights release rather than building proprietary moats. Zuckerberg makes the case that building social and assistant products to genuine usefulness requires something close to general intelligence—a lesson Meta learned by observing that coding-heavy training improves reasoning across unrelated domains. He argues against two intuitions: first, that intelligence and consciousness are intrinsically linked (they're not, and this separation makes AI a valuable tool), and second, that concentrated AI power in a few companies is safer than widespread availability. The conversation moves through why open source as system hardening applies to AI the way it does to security software, how energy and permitting will constrain scaling before capital does, and why the bounds of the conceivable shift over time—illustrated by Augustus reframing peace as positive-sum rather than temporary.
Zuckerberg details Meta's technical bets: custom silicon for inference that freed GPUs for training, the efficiency gains in Llama-3's small models, and the observation that scaling still has room even after 15 trillion tokens. He contests the notion of a sharp AGI threshold, describing instead a gradual accumulation of capabilities like multimodality and reasoning. He places most near-term AI harm in mundane adversarial uses—fraud, election interference—rather than existential scenarios, and emphasizes the role of focus as scarcest resource in large organizations. The discussion touches the mobile ecosystem's control problem (that Apple and Google gatekeep what developers build) and Meta's response: owning models and supporting open source to prevent a future of closed APIs. He forecasts that 1-gigawatt training clusters remain years away due to grid and permitting constraints, that future models will come through architectural step-functions rather than pure compute, and that users won't converge on a single AI but rather on thousands of specialized agents.
Zuckerberg argues that achieving general intelligence is necessary even for Meta's social use cases, and that aggressively open-sourcing frontier models—rather than concentrating them—is both good for Meta and the safer path for society, while physical constraints like energy will bottleneck scaling before money does.
- Working on social/assistant products revealed they require full general intelligence (e.g. coding training improves reasoning across domains)
- Concentration of AI in one institution is as dangerous as wide availability; open source hardens systems collectively like open-source security software
- Energy and permitting, not capital, will be the binding constraint on building gigawatt-scale training clusters
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
A concentration of AI power in one institution has the potential to be as dangerous as AI being widespread; it is probably bad for one institution to have an AI far more powerful than everyone else's, so open source helps prevent dangerous concentration.
“a concentration of AI in the future has the potential to be as dangerous as it being widespread... it's probably also pretty bad for one institution to have an AI that is way more powerful than everyone else's AI.”
People wrongly assume intelligence is fundamentally connected to life and consciousness; the current trajectory of AI suggests intelligence can be separated from consciousness, agency, and life, which is precisely what makes it a valuable tool.
“another bias that people tend to have is thinking thatintelligence is somehow fundamentally connected to life. It's not actually clear that it is... intelligence can be pretty separated from consciousness,agency, and things like that, which I think just makes it a super valuable tool.”
By analogy to open-source security software, widely deployed and progressively hardened AI keeps systems in check: a very intelligent AI could exploit security holes like a human who time-traveled with future knowledge, but open source lets improvements be broadly deployed across banks, hospitals, and government, hardening the whole world together so a weaker attacking AI fails against a stronger defending AI.
“a very intelligent AI probably would be able to identify some holes and basically be like a human who could go back in time a year or two and compromise all these systems.”
Before AI scaling hits a capital limit, it will hit energy constraints; no one has built a gigawatt single training cluster, and powering one is a multi-year, heavily regulated project involving power plants and transmission lines, so scaling can't be as simple as adding capital and magically getting better models.
“I actually think before we hit that, you're going to run intoenergy constraints. I don't think anyone's built a gigawatt single training cluster yet.”
Meta open-sourcing its server, network switch, and data center designs via the Open Compute Project led the industry to standardize on Meta's designs, building supply chains around them; volumes rose, costs fell, and Meta saved billions of dollars—demonstrating how open-sourcing infrastructure can be directly profitable.
“the industry now standardized on our design, which meant that the supply chains basically all got built out around our design. So volumes went up, it got cheaper for everyone, and it saved us billions of dollars which was awesome.”
Like Bell Labs inventing the transistor to enable long-distance calling, the immediate profitable use of an innovation differs from its long-run significance; specific products evolve and come and go, but the underlying advances for humanity persist—true of Reality Labs, AI, and open source work.
“It may be one of these things likeBell Labs, where they were working on thetransistorbecause they wanted to enable long-distance calling... if you asked a hundred years later maybe it's a different answer.”
It is easier to have wild ideas and to pivot when you're younger because you're earlier in your trajectory, with fewer prior commitments to disrupt—an innovator's-dilemma dynamic that exists in personal life as well as in companies; the challenge, per Picasso, is to remain an artist as you grow up.
“When you're younger, it's just easier to have wild ideas. There are all these analogies to theinnovator's dilemmathat exist in your life as well as for your company... You're earlier on in your trajectory so it's easier to pivot and take in new ideas without disrupting other commitments.”
For most companies at large scale, the scarcest commodity is not capital but focus; a startup is capital-constrained while working on one idea, but past a threshold you build multiple things and become constrained by what the CEO and management can oversee, so organizational capacity is largely limited by management attention.
“I actually think for most companies, of this scale at least, it's focus. When you're a startup maybe you're more constrained on capital... You're building multiple things... but you become more constrained on what you can direct to go well.”
AI will be fundamental—more like the original creation of computing than like a single new app platform—a low-level innovation comparable to people going from not having computers to having computers, enabling all experiences to be rethought.
“I think it's going to be more like the creation of computing in the first place... it's going to be a muchlower-levelinnovation. My sense is that it's going to be more like people going from not having computers to having computers.”
While existential risks are more intellectually interesting to discuss, the real harms needing mitigation energy are mundane day-to-day ones—people using models to commit fraud or harm others—and Zuckerberg expects these to dominate for the current and next couple of model generations.
“While it is maybe more intellectually interesting to talk about the existential risks, I actually think the real harms that need more energy in being mitigated are things where someone takes a model and does something to hurt a person.”
Meta first built custom silicon to handle inference for ranking and recommendation (Reels, News Feed, ads), which was consuming many GPUs; moving that to in-house silicon freed the more expensive NVIDIA GPUs to be used only for training, with a roadmap to eventually train large models on Meta's own silicon (not by Llama-4).
“we first builtcustom silicon that could handle inferencefor our ranking and recommendation type stuff... When we were able to move that to our own silicon, we're now able to use the more expensiveNVIDIA GPUsonly for training.”
Training models on coding makes them more rigorous and better at reasoning across many domains, even for users who never ask coding questions—so Meta heavily emphasized code in Llama-3 training despite WhatsApp users rarely coding.
“it turns out thatcoding is important for a lot of domains, not just coding. Even if people aren't asking coding questions, training the models on coding helps them become more rigorous in answering the question and helps them reason across a lot of different types of domains.”
There is no single threshold of intelligence; AGI is not one thing but the progressive addition of different capabilities (multimodality, reasoning, memory, emotional understanding), so AI will surpass humans at most tasks gradually rather than at a single moment.
“I don't think AGI is one thing. You're basically adding different capabilities.”
Contrary to the expectation that it would asymptote/saturate, the Llama-3 70B model was still learning even after training on around 15 trillion tokens, suggesting current architectures can absorb far more data than expected.
“We trained it on around 15 trillion tokens... our prediction going in was that it was going to asymptote more, but even by the end it was still learning. We probably could have fed it more tokens and it would have gotten somewhat better.”
There won't be one singular AI users interact with; every business will want an AI representing its own interests rather than one that might sell competitors' products, and the ~200 million creators on Meta's platforms will want AIs they own and train to engage their communities beyond the limits of their own time.
“Every business is going to want an AI that represents their interests. They're not going to want to primarily interact with you through an AI that is going to sell their competitors' products.”
The scenario of AI rapidly going from somewhat intelligent to extremely intelligent overnight is unlikely because of physical constraints; society will have time to acclimate rather than facing a sudden runaway.
“There is some set of people who are afraid of it really spinning out and going from being somewhat intelligent to extremely intelligent overnight. I just think that there's all these physical constraints that make that unlikely to happen.”
Zuckerberg worries more about an untrustworthy actor—an adversarial government or untrustworthy company—having a uniquely strong AI than about open AI systems doing something bad, because such an actor could cause significant mayhem for economic and security reasons.
“I stay up at night worrying more about an untrustworthy actor having the super strong AI, whether it's an adversarial government or an untrustworthy company or whatever. I think that that's potentially a much bigger risk.”
The mobile ecosystem's worst feature is that gatekeepers (Apple and Google) can tell you what you're allowed to build—a qualitative control more upsetting than the economic rent—so Meta builds its own models and supports open source to avoid a future where a few closed-model companies control the APIs and dictate what developers can build.
“you have these two gatekeeper companies, Apple and Google, that can tell you what you're allowed to build... There'sa bunch of timeswhen we've launched or wanted to launch features and Apple's just like "nope, you're not launching that."”
Caesar Augustus changed the prevailing conception of peace—from a mere temporary rest between inevitable enemy attacks to an actually positive-sum economy—a novel idea at the time that illustrates how the bounds of what people can conceive of as rational ways to work are historically contingent.
“The people's understanding of peace was peace as the temporary time between when your enemies inevitably attack you... He had this view of changing the economy from being something mercenary and militaristic to this actually positive-sum thing. It was a very novel idea at the time.”
Even with unlimited compute, a model is fundamentally limited by its architecture: you can improve a Llama-3 70B by feeding more data, but you can't go infinitely; reaching the next capability level requires a generational architecture step-function that no open source model currently matches.
“there is a fundamental limitation on the model architecture... when you make a generational improvement to something like the Llama-3 70B or the Llama-3 405B, there isn't anything like that open source today. I think that's a big step function.”
Future training of big models may shift toward inference-style generation of synthetic data that is then fed back into the model; Zuckerberg considers synthetic data generation to be more like inference than training, which has implications for whether training can be distributed.
“more of what we call training for these big models is actually more along the lines of inference generatingsynthetic datato then go feed into the model... I consider the generation of synthetic data to be more inference than training today.”
Most data centers are 50-150MW today; a 1 gigawatt data center dedicated solely to training a model would require energy on the order of a meaningful nuclear power plant, and won't be built for some number of years.
“I think a gigawatt would bethe size of a meaningful nuclear power plantonly going towards training a model.”
Harms differ in adversarial sophistication: hate speech is non-adversarial because people don't get better at being racist, so AI defenses outpace it; but nation states interfering in elections use cutting-edge technology that improves yearly, making it an arms race—which Meta believes it is currently winning by growing its AI sophistication faster than adversaries'.
“hate speech is not super adversarial in the sense that people aren't getting better at being racist.”
The 8 billion parameter Llama-3 is nearly as powerful as the largest (70B) version of Llama-2, indicating a large generational efficiency gain at small scale.
“The 8 billion is nearly as powerful as the biggest version ofLlama-2that we released. So the smallest Llama-3 is basically as powerful as the biggest Llama-2.”
Although Meta initially thought its social use cases (creator/business interactions, assistants) wouldn't need full general intelligence, working on them revealed that they do—because interactions like business customer support are multi-step reasoning problems, not simple question-answering—so Meta upped its investment to solve general intelligence.
“it wasn't completely clear at first that you were going to need full AGI to be able to support those use cases. But in all these subtle ways, through working on them, I think it's actually become clear that you do.”
Emotional understanding should be treated as its own AI modality—distinct from generic video or image—because so much of the human brain is dedicated to understanding people, expressions, and emotions; few in the industry focus on it.
“So much of thehuman brain is just dedicated to understanding peopleand understanding expressions and emotions. I think that's its own whole modality, right?”
Meta's development pattern is to hand-engineer application-specific behaviors (like tool use with Google/Bing) around the base model first—which is brittle and non-general—then progressively bake those capabilities into the model itself to make them general, using each hacked prototype to hone intuition for what to train into the next version.
“for anything that you're hand-coding you can unlock some use cases, but it's just inherently brittle and non-general.”
Just as ancient peoples could only conceive of peace as a temporary lull, many investors and observers can't grasp why Meta open-sources models—assuming it must be a temporary phase before going proprietary—when in fact open source in tech genuinely creates many winners.
“A lot of investors, and other people, can't wrap their head around why we would open source this. It's like "I don't understand, it's open source. That must just be the temporary time between which you're making things proprietary, right?"”
A key lesson of technology is to move things from the physical-constraint realm into software as much as possible, because software is far easier to build, evolve, and democratize—anyone can write or modify open-source code, whereas not everyone can have a data center; the metaverse applies this by turning physical presence into software-enabled digital presence.
“one lesson of technology is that you want to move things from the physical constraint realm into software as much as possible because software is so much easier to build and evolve. You can democratize it more”
Meta's decision to acquire a large GPU fleet came from being behind on Reels recommendations versus TikTok; after being constrained, Zuckerberg ordered enough GPUs for Reels/feed ranking and then doubled it on the principle that some unforeseen need would emerge—which turned out to be AI.
“we have to make sure that we're never in this situation again. So let's order enough GPUs to do what we need to do on Reels and ranking content and feed. But let's also double that.”
Meta's ratio of inference compute to training compute is probably much higher than most other companies doing AI, because of the sheer volume of the community it serves, which justifies training models on more data than is compute-optimal for training alone.
“Our ratio of inference compute required to training is probably much higher than most other companies that are doing this stuff just because of the sheer volume of the community that we're serving.”
The biggest bets people make are usually based on conviction and values rather than financial analysis, because it is very hard to connect the dots forward analytically; Zuckerberg didn't sell Facebook for $1B because he'd just build a similar company anyway.
“a lot of the biggest bets that people make are often just based on conviction and values. It's actually usually very hard to do the analyses trying to connect the dots forward.”
Meta is strongly pro-open-source but hasn't committed to releasing everything; the decision should be evaluated release-by-release, and if some qualitative change makes a model's capabilities irresponsible to release, Meta won't open source it.
“I haven't committed to releasing every single thing that we do... If at some point however there's some qualitative change in what the thing is capable of, and we feel like it's not responsible to open source it, then we won't.”
Open-sourcing makes sense as long as Meta's app-specific work remains differentiated enough; the one world where it doesn't is if the model itself becomes the product, in which case open-sourcing commoditizes yourself heavily—but Zuckerberg doesn't think Meta is in that zone.
“There is one world where maybe that's not the case. Maybe the model ends up being more of the product itself. I think it's a trickier economic calculation then, whether you open source that. You are commoditizing yourself then a lot.”
The right question for releasing a model is not what bad behaviors it could exhibit—any product can exhibit harms that can be mitigated—but rather which harms could NOT be mitigated after they appear; Meta has built a taxonomy of ~18-19 categories of social-media harm and uses AI systems to detect them.
“I think the question isn't really what behaviors would it show, it's what things would we not be able to mitigate after it shows that.”
Meta's Llama license is very permissive except for a limit on the largest companies: if a major cloud like Microsoft Azure or Amazon is going to resell the model and make money off it, they should talk to Meta and have a revenue share arrangement.
“If you're likeMicrosoft AzureorAmazon, if you're going to be reselling the model then we should have some revenue share on that. So just come talk to us before you go do that.”
Most decisions that end up looking good at Meta come from having messed something up earlier and resolving not to repeat the mistake, rather than from being far ahead.
“most of the times where we make some decision that ends up seeming good is because we messed something up before and just didn't want to repeat the mistake.”
Zuckerberg attributes his big bets (declining the $1B sale, investing $100B in AI/metaverse) to a deep constitutional drive: he feels he's doing something wrong if he's not building something new, and is fundamentally incapable of running things without building, so taking a swing at the next thing isn't even a question for him.
“I just feel constitutionally that I'm doing something wrong if I'm not building something new... I'm fundamentally incapable of running something, or in my own life, and not trying to build new things that I think are interesting.”
Meta AI, powered by the new Llama-3, is the most intelligent freely-available AI assistant people can use.
“we think now that Meta AI is the most intelligent, freely-available AI assistant that people can use.”