Dwarkesh PodcastMark Zuckerberg — AI will write most Meta code in 18 months
At a glance
WHAT IT’S REALLY ABOUT
Zuckerberg predicts AI-coded Meta, open-source dominance, and weirder internet
- Mark Zuckerberg discusses Meta’s Llama 4 roadmap, emphasizing open-source models, multimodality, and an upcoming two-trillion-parameter ‘Behemoth’ model designed mainly as a teacher for smaller distilled systems.
- He predicts that within 12–18 months, most of the code for Meta’s AI efforts will be written by AI agents, while arguing this will increase, not decrease, overall demand for human work and new services.
- Zuckerberg frames Meta’s north star as a ubiquitous, personalized AI assistant embedded across messaging, apps, and AR glasses, and believes AI will reshape not just productivity and research but also entertainment, culture, and online social interaction.
- He defends Meta’s semi-open Llama licensing, stresses the geopolitical stakes of AI standards and infrastructure, and reflects on content moderation, governance, and the need to avoid over-deference to government and media critics.
IDEAS WORTH REMEMBERING
5 ideasAI will soon write most frontier AI code, but within real-world constraints.
Zuckerberg expects that in 12–18 months, AI agents at Meta will generate the majority of code for Llama and related efforts—running tests, finding bugs, and exceeding average engineer quality—yet he stresses progress is bottlenecked by compute, infrastructure, and experimentation bandwidth, not just coding speed.
Open-source models are becoming the default, but Meta wants to steer the standard.
Zuckerberg argues 2024 is the year open-source broadly overtakes closed in usage, with multiple strong families (not just Llama), but he believes Meta’s consistent commitment is crucial; others may be ‘dabbling’ and could re-close when convenient, so Meta keeps pushing open models that encode ‘American’ values and secure architectures.
Benchmarks matter less than real user value and latency in products.
He notes Llama 4 Maverick can be pushed to the top of Chatbot Arena via tuning, but Meta chooses to optimize for Meta AI usage metrics, low latency, and cost per intelligence—especially for consumer and voice use cases—rather than gaming public leaderboards or chasing narrow reasoning benchmarks alone.
Distillation from giant models and even rival models will be central.
The two-trillion-parameter Behemoth is mainly a teacher: Meta intends to distill ~90–95% of its intelligence into 10x smaller, cheaper models, potentially combining strengths from multiple sources (e.g., Llama plus better coders) while mitigating security risks via verifiable domains, filters like Llama Guard/Code Shield, and intensive red-teaming.
Personal AI will be pervasive, multimodal, and conversational throughout the day.
Zuckerberg envisions people constantly interacting with AI—within WhatsApp, Instagram, the standalone Meta AI app, and AR glasses—using full-duplex voice, persistent memory, and context from social graphs and feeds, making AI feel like an always-available, personalized collaborator and companion.
WORDS WORTH SAVING
5 quotesI would guess that, like, sometime in the next 12 to 18 months, we'll reach the point where, like, most of the code that's going towards these efforts is written by AI.
— Mark Zuckerberg
If you think that 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.
— Mark Zuckerberg
I tend to think that, for at least the foreseeable future, this is gonna lead towards more demand for people doing work, not less.
— Mark Zuckerberg
I would guess that the world is gonna get a lot more... a lot funnier and, like, weirder.
— Mark Zuckerberg
These models encode values and ways of thinking about the world.
— Mark Zuckerberg
High quality AI-generated summary created from speaker-labeled transcript.
Get more out of YouTube videos.
High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.
Add to Chrome