In brief
Mark Zuckerberg argues that Meta should build its own AI models and open-source them, tuned for fast, cheap, personal consumer products rather than for benchmark rankings. He predicts AI will write most of the code on Meta's AI research efforts within 12 to 18 months. He says the AI field will split into many specialties, and that no single lab will dominate.
The central tension is between an expected AI takeoff and real-world limits. Zuckerberg accepts that coding agents could speed up AI research, but says power, chips, test capacity and human adoption each act as bottlenecks. He is also wary of China, benchmark gaming and relying on rival open models. He is optimistic that AI will increase demand for human work.
Facts
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Meta announced four Llama 4 models and released the first two, Scout and Maverick, which Mark Zuckerberg describes as mid-sized to small.
“we announced four models and we released the first two, the Scout and Maverick ones, which are kind of like the mid-sized models”
0:00:52 · Mark Zuckerberg · Llama 4 · 2025
Zuckerberg says Meta's Behemoth model, its first at the frontier, has more than 2 trillion parameters.
“That's going to be our first model that is sort of at the frontier. I mean, it's like more than 2 trillion parameters.”
0:02:00 · Mark Zuckerberg · Behemoth model · 2025
Meta released a 405 billion parameter model in Llama 3.1 and added multimodal capabilities in Llama 3.2.
“3.1 was when we released the 405 billion model. 3.2 is when we got all the multimodal stuff in.”
0:02:47 · Mark Zuckerberg · Llama 3 releases · 2024
Zuckerberg says Meta is building a Llama 4 reasoning model that had not yet been released.
“We're building a Llama 4 reasoning model and that'll come out at some point.”
0:04:59 · Mark Zuckerberg · Reasoning model · 2025
Zuckerberg says Meta's team tuned a version of Llama 4 Maverick that ranked at the top of the public arena leaderboard, while the released version was not tuned for it.
“it was relatively easy for our team to tune a version of llama 4 maverick um that basically was way at the top”
0:07:32 · Mark Zuckerberg · Benchmark gaming · 2025
Meta has an internal assistant called Metamate and is building coding and AI research agents around it.
“We have this internal thing that we call Metamate and basically in a number of different coding and AI research agents that we're building around that.”
0:24:21 · Mark Zuckerberg · Internal AI agents · 2025
Zuckerberg says an ads team project to automate ranking experiments found Meta was limited by compute and test cohorts, not by ideas.
“we were bottlenecked on compute to run tests based on the number of hypotheses.”
0:18:36 · Mark Zuckerberg · Ads experiments
Zuckerberg cites a statistic that the average American has fewer than three people they consider friends, while wanting meaningfully more.
“the average american i think has i think it's fewer than three friends three people that they consider friends”
0:28:28 · Mark Zuckerberg · Friendship
Zuckerberg attributes the success of the Ray-Ban Meta glasses to their use for music, calls, photos and video, with AI available on demand.
“the Ray-Ban meta product has done so well is like, all right, it's like great for listening to music and taking phone calls and taking photos and videos”
0:32:44 · Mark Zuckerberg · Smart glasses
Zuckerberg says DeepSeek used reduced-capability Nvidia chips, the only ones Nvidia may sell in China under export controls.
“they're using like partially nerfed chips that are the only thing that nvidia is allowed to sell in china because of the export controls”
0:36:26 · Mark Zuckerberg · Export controls
Zuckerberg describes DeepSeek's models as text-only, with impressive infrastructure and text results.
“DeepSeek is text only. So the infrastructure is impressive. The text result is impressive.”
0:36:53 · Mark Zuckerberg · DeepSeek · 2025
The Llama license requires companies with more than 700 million users to talk to Meta before using the model.
“your license says that if you reach 700 million people, you have to come talk to us.”
0:42:17 · Mark Zuckerberg · Llama license
Zuckerberg says the license exists so Meta can negotiate with large cloud companies such as Microsoft, Amazon and Google before they resell Llama.
“if you're going to have these large cloud companies like Microsoft and Amazon and Google turn around and sell our model, that we should at least be able to have a conversation with them”
0:40:54 · Mark Zuckerberg · Llama license
Zuckerberg cites Android as a platform that began as open source and has become more closed over time, as a caution about relying on others' open-source commitments.
“Android started off as the open source thing. There's not really like any open source alternative.”
0:46:24 · Mark Zuckerberg · Open source
Zuckerberg says French speakers who tested an early Llama translated into French said it sounded like an American who learned French, showing how models embed values.
“this sounds like an American who learned to speak French”
0:47:57 · Mark Zuckerberg · Model values
Zuckerberg says distilling a large model can keep roughly 90 to 95 percent of its intelligence in a model a tenth of the size and cost.
“95 % of the intelligence at 10 % of the cost is like pretty good for a lot of things”
0:51:56 · Mark Zuckerberg · Distillation
Meta has released open-source safety tools, Llama Guard and CodeShield, that filter inputs and outputs for models.
“whether it's like the LamaGuard open source or the CodeShield open source things that we've done that basically allow you to incorporate different input into your models”
0:53:02 · Mark Zuckerberg · AI safety tools
Zuckerberg compares costly AI services to Netflix and ESPN, which charge subscriptions because content is expensive to produce and ads cannot cover it.
“if you want to watch netflix um or like espn or something you need to pay for that.”
0:55:30 · Mark Zuckerberg · Business models
Zuckerberg says Meta's consumer AI will have a free, ad-supported tier alongside a premium paid service.
“for the consumer piece, we definitely want to have a free thing. And I'm sure that will end up being ad supported.”
0:56:51 · Mark Zuckerberg · AI monetization
Zuckerberg says Meta's default approach is to build a productive relationship with whoever runs the US government, and he criticizes the previous administration for not engaging with business.
“Our default as an American company should be to try to have a productive relationship with whoever is running the government.”
1:02:23 · Mark Zuckerberg · Politics · 2025
Zuckerberg says Meta's fact-checking program was less effective than Community Notes because there were too few fact checkers and users did not trust them.
“I just think the fact checking thing was not as effective as community notes because it's not an internet scale solution.”
1:04:24 · Mark Zuckerberg · Content moderation
Zuckerberg says almost 3.5 billion people use Meta's services every day.
“we have almost 3.5 billion people use our services every day”
1:11:00 · Mark Zuckerberg · Meta users · 2025
Zuckerberg estimates that staffing voice customer support for Meta's user base would cost roughly $10 to $20 billion a year.
“it would be like $10, $20 billion, something ridiculous a year to kind of staff that”
1:11:30 · Mark Zuckerberg · Customer support
Zuckerberg says there are more truck driving jobs now than when self-driving cars were first widely discussed, and uses this to argue automation does not simply eliminate jobs.
“There's actually more truck driving jobs now than there were like when we started talking about self-driving cars”
1:12:29 · Mark Zuckerberg · Automation and jobs