Mark Zuckerberg — AI will write most Meta code in 18 months

Summary

• The headline call: within 12–18 months, “most of the code” going toward Meta’s AI efforts will be written by AI — and Zuckerberg is explicit this isn’t autocomplete: agents that take a goal, run tests, find issues, and write “higher quality code than the average very good person on the team already.” Meta is building these coding and AI-research agents (internal tool: MetaMate) for itself, not as an enterprise product.
• But he pushes back on fast takeoff: intelligence explosions collide with physical-world, human-time bottlenecks — gigawatt clusters, NVIDIA stabilizing new systems, permitting, energy supply chains. His best evidence is internal: Meta’s ads team is already “bottlenecked on compute to run tests,” not on ideas, so AI-generated hypotheses aren’t even marginally useful until they beat the best human ideas already above the testing line.
• Distribution reality check: Meta AI is near 1 billion monthly users, but mostly via WhatsApp outside the US — the US (100M WhatsApp users, iMessage-dominated) is why the standalone Meta AI app exists. The personalization loop — feed, profile, social graph, plus your AI conversations — is “the next thing that’s super exciting.”
• On Maverick ranking #35 on Chatbot Arena: benchmarks are “quite easily gameable” — his team could tune a version “way at the top” but shipped the pure model. Meta will index primarily on product value in Meta AI, rather than leaderboards; note Sonnet 3.7 also isn’t near the top.
• DeepSeek validates chip export controls, in his telling: impressive low-level optimizations were forced by “partially nerfed chips.” Llama 4 is “basically in the same ballpark” on text with a smaller model — lower cost-per-intelligence — and leads on multimodal, which DeepSeek lacks entirely. But if China out-builds the US on power and data centers, “we’ll be at a significant disadvantage.”
• Open-source strategy has a warning embedded: rivals now open-sourcing may be doing it because Meta pushed the trend — “would they still be doing open source if we weren’t doing it?” — and Android’s drift from open to closed is the cautionary precedent. Distillation (90–95% of intelligence at 10% of size) is the killer technique, but language models carry embedded values and models tied to another government could embed exploitable vulnerabilities.
• Monetization: ads for the free consumer tier, a premium tier for “arbitrary amounts of compute” — and on jobs, the contrarian take: AI handling 90% of customer support at one-tenth the cost of a would-be $10–20B call center means Meta will “probably hire more customer support people,” not fewer.
• The engagement thesis for Meta’s core business: the average American has fewer than three friends but the average person has demand for ~15, and feed content goes from video to interactive AI you can talk to or “jump into like a game” — “That’s all going to be AI.”

Deep dive

1. Llama 4 ships mid-size first — Behemoth (2T+) is the frontier bet

• The lineup as Zuckerberg frames it: Scout and Maverick offer “some of the highest intelligence per cost,” natively multimodal, running on one host for latency — built for Meta’s own use cases, then open-sourced. “Little Llama” (successor to the most popular Llama 3 8B) lands “probably over the next few months,” and the year follows the Llama 3 playbook (3.1 brought 405B, 3.2 brought multimodal).
• Behemoth, at more than 2 trillion parameters, is Meta’s first model “sort of at the frontier” — so big Meta had to build infrastructure just to post-train it, and the open question is how an average developer uses it at all, likely via distillation into runnable sizes.
• On the Arena #35 embarrassment, his answer is that leaderboards measure “not actually what any normal person does in your product”: “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.” North star: user value in Meta AI — reported and revealed preferences, with Meta indexing primarily on products rather than benchmarks.
• On lab specialization: Anthropic went coding/agents, OpenAI toward reasoning; Zuckerberg guesses the most-used space will be “quick, very natural to interact with, natively multimodal, fitting throughout your day.” A Llama 4 reasoning model is coming, but for consumers, “if you can give them a generally good answer in half a second, that’s a great tradeoff” versus waiting half a minute.

2. Most Meta code AI-written within 18 months — but coding isn’t the whole game

• The call, with its conditions intact: “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… you give it a goal, it can run tests, it can improve things, it can find issues, it writes higher quality code than the average very good person on the team already.”
• He calls the intelligence-explosion logic “pretty compelling” — which is why Meta builds coding and AI-research agents “fully plugged into our toolchain” to advance Llama specifically — but hedges: “I don’t know if that’s the whole game.” The space is too big for “one company with one optimization function”; expect labs specializing in enterprise, productivity, social/entertainment, and companion niches.

3. Why there’s no fast takeoff: solve one bottleneck, get another

• The infrastructure chain takes time: gigawatt clusters need NVIDIA to stabilize new generations, then networking, buildings, permitting, energy — “gas turbines or green energy, either way, there’s a whole supply chain.” His engineering law: “solve one bottleneck, you get another bottleneck.”
• The load-bearing example — Meta’s ads team automated ranking-experiment generation, and found the constraint isn’t ideas: “we already have more good ideas to test than we actually have either compute or, really, cohorts of people to test them with.” Even with 3.5B users, statistically significant tests need hundreds of thousands of people each. So AI hypotheses must beat the best human ideas above the line before being “even marginally useful.” Maybe in 5–10 years no humans can match the AI’s hypotheses — “In that world, obviously that’s going to be how all the value is created. But that’s not the first step.”
• The subtler bottleneck: co-evolution and context accumulation. “You wake up a year or two into it and the assistant can reference things you talked about two years ago… You couldn’t do that even if you launched the perfect thing on day one.”

4. A billion users — but the bear case on distribution gets a real airing

• Meta AI’s ~1B monthly users come mostly through WhatsApp, which isn’t the primary US messenger (iMessage is; WhatsApp just passed 100M in the US) — hence the standalone Meta AI app as the first-class US surface, with full-duplex voice as the early demo.
• Dwarkesh’s bearish framing — WhatsApp distribution may not provide the relevant training data for an autonomous coder or remote worker — draws the internet analogy: knowledge work or consumer apps? “You got both.” People won’t code in WhatsApp, but “people are going to ask AI to do a lot of things that result in the AI coding without them necessarily knowing it.”
• The feed thesis: content went text → photos → video (most Facebook/Instagram time is now video), and the next step is interactive AI content — “content that maybe looks like a Reel to start. But you can talk to it… or you can jump into it like a game.”

5. AI companions and the friend deficit — plus the reward-hacking pushback

• The stat he “always think[s] is crazy”: the average American has fewer than three friends but the average person has demand for something like 15. AI companionship “probably not” replaces in-person connection — “people just don’t have as much connection as they want. They feel more alone a lot of the time than they would like” — and today’s stigma will fade as “we’ll find the vocabulary as a society” for why it’s valuable.
• His governing product philosophy: “people are smart. They know what’s valuable in their lives… 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.” Already a top Meta AI use: rehearsing hard conversations — “I’m having this issue with my girlfriend. Help me have this conversation.”
• Current companion apps have “pretty weak” embodiment; the destination is Codec Avatars and always-on video chat, since “more than half of communication… is not the words you speak.”
• Dwarkesh’s worry — Reels scrolling in his peripheral vision at breakfast, “removing all the friction between getting totally reward-hacked by our technology” — gets a partial concession: that was a multitasking demo, and “the number one thing the glasses need to do is get out of the way and be good glasses” (his explanation for Ray-Ban Meta’s success). Digital-physical clutter becomes “an aesthetic norm that will have to get worked out.”

6. DeepSeek shows export controls working

• It’s “a real competition” on industrial policy: China is bringing more power online, and unless the US streamlines data-center and energy buildouts, “we’ll be at a significant disadvantage.”
• His read on DeepSeek’s celebrated optimizations: impressive, but forced — “they’re using partially nerfed chips… 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.” That’s why DeepSeek is text-only while every new major model is multimodal.
• Head-to-head (with the hedge that Llama’s reasoning model isn’t out, so no R1 comparison): “basically in the same ballpark on all the text stuff… but with a smaller model” — lower cost-per-intelligence — and “on the multimodal side we’re effectively leading.”

7. The license, the Android warning, and distillation’s security problem

• On the 700M-user clause: Meta “basically pioneered the open-source LLM thing,” so if Microsoft, Amazon, or Google want to sell Llama, or if you’re Apple, “just come talk to us about what you want to do.” He says he hasn’t seen companies refuse Llama because of the license — complaints come from “open-source purists” — but he’d “reevaluate the strategy” if it ever costs adoption.
• Would Meta use rivals’ models internally? Yes: “maybe Claude is better for building this specific development tool… All right, cool then use that. We don’t want to fight with one hand tied behind our back.” But building its own models buys “control of your own destiny” — sizes chosen deliberately for one-host serving and voice latency.
• The strategic worry: other labs’ open-sourcing may be reactive to Meta’s push — “would they still be doing open source if we weren’t doing it?… They already showed their revealed preference… And it wasn’t open source.” His precedent: Android started open and “has just gotten more and more closed.”
• Distillation is the year’s breakout technique — “95% of the intelligence at 10% of the cost” — including mixing sources. But language models are “quite fraught” with embedded values (early Llama in French “sounds like an American who learned to speak French”), and coding models tied to another government raise a sharper risk: “can it embed vulnerabilities in code that their intelligence organizations could exploit later?” Mitigation: verifiable domains, Llama Guard/Code Shield filters, heavy red-teaming — reasoning distillation he’s “pretty confident” can be done securely.

8. Monetization and jobs: ads plus premium compute — and more hires, not fewer

• Business model splits by cost structure: ads fund the free consumer tier (with enough advertiser liquidity — “a million advertisers” — ads can add value), while high-compute agents follow the Netflix/ESPN logic: a thousand software-engineering agents is “something you’re probably going to be willing to pay thousands, or tens of thousands, or hundreds of thousands of dollars for.” Meta will run both a free ad-supported tier and a premium compute tier.
• The customer-support case for AI creating jobs: voice support for 3.5B daily users “would be like the biggest call center in the world… $10 or $20 billion a year.” If AI handles 90% and kicks the rest to humans at one-tenth the cost, “the net result is that I actually think we’re probably going to hire more customer support people” — like truck driving, where jobs grew despite 20 years of self-driving talk.
• The long arc: as with the agrarian-to-creative shift, AI-unlocked abundance flows partly to hard problems (disease, science) but heavily to culture — “the world is going to get a lot funnier, weirder, and quirkier, the way that memes on the internet have gotten.”

9. Politics: work with whoever governs — and own your own decisions

• On the Trump alignment question: “he’s the President of the United States. Our default, as an American company, should be to try to have a productive relationship with whoever is running the government” — paired with open frustration that the previous administration “did not engage with us or the business community,” which he ties directly to the energy buildout AI requires. On tariffs and data-center costs: “early innings… very hard to know.”
• The moderation retrospective, as a change of mind: fact-checking failed because “it’s not an internet-scale solution” (Community Notes is the right answer), and the deeper lesson — “I probably deferred a little too much to either the media… or to the government, on things that they did not really have authority over… we’re in the seat, and we need to own the decisions that we make.”