Mark Zuckerberg — AI will write most Meta code in 18 months
What this covers
Mark Zuckerberg speaks with Dwarkesh Patel about the near-term trajectory of AI development and Meta's positioning within it. The conversation centers on a specific forecast: within 12 to 18 months, most code written for Meta's AI efforts will be generated by AI agents rather than humans—not simple autocomplete, but systems given goals that run tests, iterate, and produce work exceeding average engineer output. Underpinning this prediction is a broader thesis about how AI will develop: not as a race toward a single superintelligence, but as a fragmented ecosystem where different models optimize for different tasks, where real-world constraints (compute, energy, human feedback loops) prevent sudden takeoff, and where Meta's advantage lies in open-source distribution and product-grounded measurement rather than leaderboard chasing.
The conversation branches into three distinct domains. On infrastructure and geopolitics, Zuckerberg argues that bottleneck shifting makes fast takeoff implausible—that as one part of the compute stack accelerates, human-dependent tasks like building electrical grids and securing permits become the constraint. He contends that export controls are working, forcing competitors to accept tradeoffs (DeepSeek's models are text-only while others have achieved multimodal capability). On business strategy, he stakes Meta's future on open-source models overtaking closed ones, warning that competitors' revealed preference before Llama existed suggests they would abandon openness if not for market pressure. On AI development itself, he pushes back against leaderboard optimization, noting that benchmarks like LM Arena skew toward narrow use cases and are easily gamed, whereas different applications—coding, companionship, customer service—warrant different models and architectures optimized for latency and inference cost rather than test-time reasoning. He also returns repeatedly to how models encode values and how distillation across safety boundaries requires careful tooling like Llama Guard and Code Shield.
Zuckerberg argues that AI will not produce a single winner-take-all superintelligence but a diverse ecosystem of specialized models, where Meta's edge comes from open source, distribution, and product-anchored benchmarks rather than chasing leaderboards or fast-takeoff narratives.
- Real-world bottlenecks (compute, energy, human feedback loops) make fast takeoff implausible even as coding automation accelerates
- Open source models are on track to overtake closed source, and Meta's strategy depends on keeping that trend alive
- Different applications (coding, companionship, entertainment, productivity) will tend toward different models and business models
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.
Language models encode values and ways of thinking about the world — an early Llama translated to French sounded like an American speaking French — and models out of China have certain values baked in that can't be removed with a light fine-tune, which is why building standards around American models matters.
“The feedback we got from French people was, "This sounds like an American who learned to speak French. It doesn’t sound like a French person."”
A real risk of using a model tied to another government is that it could embed vulnerabilities in code that the country's intelligence organizations could later exploit — you could wake up to systems that model secured being vulnerable in a way that country knows about and you don't, or it could activate a vulnerability at some point.
“can it embed vulnerabilities in code that their intelligence organizations could exploit later? In some future version you're using a model that came from another country and it's securing your systems. Then you wake up and everything is just vulnerable in a way that that country knows about and you don’t.”
Contrary to the common belief that AI will automate jobs away, the history of technology shows that taking away 90% of the work usually leads to wanting more people, not fewer — illustrated by AI making customer support economically viable: if AI handles 90% of Meta's 3.5 billion users' issues and kicks the rest to humans, the 10x cost reduction makes voice support worth offering, likely increasing support hiring.
“The common belief is that AI will automate jobs away. But that hasn't really been how the history of technology has worked. Usually, you create things that take away 90% of the work, and that leads you to want more people, not less.”
Reasoning models that consume more test-time/inference-time compute to provide more intelligence are a compelling paradigm for math and coding, but for consumer products latency and intelligence-per-cost matter more — people won't wait half a minute for an answer when a good answer in half a second is a great tradeoff.
“reasoning models that consume more test-time or inference-time compute in order to provide more intelligence are a really compelling paradigm”
AI development is more research- and model-led than product-led: you can't design the product you want and build a model to fit it; you must design the model and target capabilities first, then get emergent properties, then discover what products those properties enable — because at the end of the day people want to use the best model.
“You can't just design the product that you want and then try to build the model to fit into it. You really need to design the model first and the capabilities that you want, and then you get some emergent properties.”
Open-source benchmarks like LM Arena are skewed toward a narrow set of use cases that don't match what normal people actually do in a product, and they are easily gameable — Meta's team could easily tune a Llama 4 Maverick version to top the Arena, but the released untuned model ranks lower, so optimizing for these benchmarks leads product development astray.
“It was relatively easy for our team to tune a version of Llama 4 Maverick that could be way at the top. But the version we released, the pure model, actually has no tuning for that at all, so it's further down.”
Distillation has emerged as a powerful technique that works better than most predicted: you can take a much bigger model and capture roughly 90-95% of its intelligence in a model that's 10% of the size — not 100% of the intelligence, but 95% at 10% of the cost is a great tradeoff for many uses, and you can now distill from multiple open-source sources to combine their strengths.
“You can basically take a model that's much bigger, and capture probably 90 or 95% of its intelligence, and run it in something that's 10% of the size.”
Export controls are working: DeepSeek had to spend their calories and time on impressive low-level infrastructure optimizations that American labs didn't need to, because they are using partially nerfed chips that are the only ones NVIDIA is allowed to sell in China — and as a result their model is text-only while every major model is now multimodal.
“It’s because they’re using partially nerfed chips that are the only ones NVIDIA is allowed to sell in China because of the export controls. DeepSeek basically had to spend a bunch of their calories and time doing low-level infrastructure optimizations that the American labs didn’t have to do.”
Fast-takeoff scenarios are wrong because building physical infrastructure takes human time: gigawatt compute clusters require NVIDIA to stabilize new systems, networking, building construction, permitting, and energy supply chains, so as intelligence grows in one part of the stack you simply hit a different bottleneck — that's how engineering always works.
“Part of what I generally disagree with on thefast-takeoffview is that it takes time to build out physical infrastructure.”
A core product principle is that people are smart and know what's valuable in their lives — if you think something someone is doing is bad but they think it's really valuable, most of the time they're right and you're wrong, you just haven't yet found the framework for why it's valuable to them.
“if you think something someone is doing is bad and they think it's really valuable, most of the time in my experience, they're right and you're wrong. You just haven't come up with the framework yet for understanding why the thing they're doing is valuable and helpful in their life.”
Over 100-150 years, human society has shifted from agrarian survival — where most energy went to feeding ourselves — toward basic needs being a smaller share of energy, leading to more creative and cultural pursuits and less time working; AI will continue this arc, making the world funnier, weirder, and quirkier with richer cultural technology to express complex ideas.
“it's basically people going from being primarily agrarian — with most human energy going toward just feeding ourselves — to that becoming a smaller and smaller percent.”
The AI space is so massive that no single company with one optimization function will serve everyone best; like the early internet which produced both knowledge work and consumer apps, there will be specialization across labs — enterprise/coding, productivity, social/entertainment, informational assistants, and companions.
“I don't think there's just going to be one company with one optimization function that serves everyone as best as possible. There are going to be a bunch of different labs doing leading work in different domains.”
Fact-checking was less effective than Community Notes because it's not an internet-scale solution — there weren't enough fact-checkers and people didn't trust the specific fact-checkers — so a more robust, distributed system like Community Notes is the right approach to content moderation.
“the fact-checking thing was not as effective as Community Notes because it's not an internet-scale solution. There weren't enough fact-checkers, and people didn't trust the specific fact-checkers.”
Within the next 12 to 18 months, most of the code going toward Meta's AI efforts will be written by AI — not autocomplete, but agents you give a goal that run tests, find issues, improve code, and write higher-quality code than the average very good engineer on the team.
“I would guess that sometime in the next 12 to 18 months, we'll reach the point where most of the code that's going toward these efforts is written by AI. And I don't mean autocomplete.”
The average American has fewer than three people they consider friends, while the average person has demand for meaningfully more — around 15 — so AI companions could fill a real gap because people consistently have less connection than they want and feel more alone than they'd like.
“The average American has fewer than three friends, fewer than three people they would consider friends. And the average person has demand for meaningfully more. I think it's something like 15 friends or something.”
This year is on track for open-source models to generally overtake closed-source models as the most-used models, and notably it's no longer just Llama — many good open-source models now exist in the field.
“the prediction that this would be the year open source generally overtakes closed source as the most used models out there, I think that's generally on track to be true.”
When Meta automated ad ranking experiment generation, they found they were bottlenecked not on hypotheses but on compute and on cohorts of people large enough to make tests statistically significant — even with 3.5 billion users, AI is only useful once its average hypothesis quality exceeds the best human-generated ones already above the testing line.
“What we basically found was that we were bottlenecked on compute to run tests, based on the number of hypotheses.”
Many competitors now doing open source would not be doing it if Meta weren't — they're jumping on the train to avoid losing to the open-source trend, and their revealed preference (closed APIs before Llama existed) shows they'd abandon it; Android is the cautionary example of a project that started open source and steadily became more closed.
“would they still be doing open source if we weren’t doing it?”
The number one design principle for AR glasses is that they must get out of the way and be good glasses first — the AI is there when you want it but invisible when you don't — which is partly why the Ray-Ban Meta product succeeded, and a future of attention-competing content in the corner of your vision is not what people will want.
“Probably the number one thing the glasses need to do is get out of the way and be good glasses.”
AI monetization will span a spectrum: ads remain great for free services that need cost coverage and can even add value if ranking is good and advertiser liquidity is high, but high-compute applications like a thousand software engineering agents will warrant premium pricing in the thousands to hundreds of thousands of dollars — Meta wants both a free ad-supported tier and a premium tier.
“Not everyone is going to want a software engineer, or a thousand software engineering agents, or whatever it is. But if you do, that's something you're probably going to be willing to pay thousands, or tens of thousands, or hundreds of thousands of dollars for.”
Meta builds its own large models because no other model is exactly what it wants — even open models impose architectures and size tradeoffs affecting latency and inference cost that matter enormously at Meta's scale — but Meta won't fight with one hand behind its back and will use a model like Claude for a specific dev tool if it's better.
“There's a level of control of your own destiny that you only get when you build the stuff yourself.”
The next exciting phase for Meta AI is the personalization loop, where the AI draws on context from your feed, profile, and social graph as well as your AI interactions, eventually being able to reference things you discussed two years ago — something impossible to launch fully formed because the data must accumulate over time.
“You wake up a year or two into it and the assistant can reference things you talked about two years ago and that’s pretty cool. You couldn’t do that even if you launched the perfect thing on day one.”
Distilling language models is fraught because of embedded values, but for reasoning models limited to verifiable domains you can distill quite securely by combining code cleanliness and security filters, tools like Llama Guard and Code Shield, and extensive red teaming to check the model isn't doing anything unwanted after distillation.
“On reasoning, though, you can get a lot of the way there by limiting it to verifiable domains, and running code cleanliness and security filters.”
Content has evolved from text to photos to video as networks improved, and in five years feeds won't just be passive video — content will be interactive: a Reel you can talk to that talks back and changes, or that you can jump into like a game, all generated by AI.
“But do you think in five years we’re just going to be sitting in our feed and consuming media that's just video? No, it's going to be interactive. You'll be scrolling through your feed. There will be content that maybe looks like a Reel to start. But you can talk to it, or interact with it, and it talks back”
The Llama license is not onerous because it generally doesn't try to stop people from using the model — it only asks the largest cloud companies (Microsoft, Amazon, Google, Apple) who reach 700 million users to have a business conversation with Meta before reselling a model Meta spent billions training, which is a reasonable ask.
“if you're spending many billions of dollars training these models, I think asking the other companies — the huge ones that are similar in size and can easily afford to have a relationship with us — to talk to us before they use it seems like a pretty reasonable thing.”
More than half of communication in an actual conversation is not the words you speak but nonverbal cues and gestures, which is why Codec Avatars that render AI as a realistic person with gestures matter for making AI relationships feel real.
“More than half of communication, when you're actually having a conversation, is not the words you speak. It's all the nonverbal stuff.”
An American company's default should be to try to have a productive relationship with whoever runs the government, because dialogue is necessary to make progress on things like building the energy capacity AI needs — framed against frustration that the previous administration didn't engage with the business community.
“Our default, as an American company, should be to try to have a productive relationship with whoever is running the government.”
The most leveraged thing a CEO does in a week is probably not the same thing each week — by definition, if it were, you should spend more than one hour on it — and recruiting awesome people is a consistently high-leverage activity.
“the most leveraged thing you do in a week is not the same thing each week. Or else, by definition, you should probably spend more than one hour doing that thing every week.”
Meta AI is most used in WhatsApp, which is mostly used outside the US where iMessage dominates, so Americans underestimate Meta AI usage; this is precisely why a standalone Meta AI app matters — to build a first-class experience in front of US users where the main messaging channel isn't available.
“Meta AI is actually most used in WhatsApp. WhatsApp is mostly used outside of the U.S... people in the U.S. probably tend to underestimate Meta AI usage somewhat.”
Llama 4 is in the same ballpark as DeepSeek on text tasks but with a smaller model, so cost-per-intelligence is lower, and on the multimodal side Meta is effectively leading because DeepSeek's models are text-only.
“we’re basically in the same ballpark on all the text stuff that DeepSeek is doing but with a smaller model. So the cost-per-intelligence is lower with what we’re doing for Llama on text. On the multimodal side we’re effectively leading at and it just doesn’t exist in their models.”