Dwarkesh Podcast
Dwarkesh Podcast

Mark Zuckerberg — Llama 3, $10B models, Caesar Augustus, & 1 GW datacenters

Mark Zuckerberg on: - Llama 3 - open sourcing towards AGI - custom silicon, synthetic data, & energy constraints on scaling - Caesar Augustus, intelligence explosion, bioweapons, $10b models, & much more Enjoy! Watch on YouTube. Listen on Apple Podcasts, Spotify, or any other podcast platfor

Featured Speakers

Dwarkesh Patel HostMark Zuckerberg Guest

Topics Discussed

Episode Summary

Executive Summary: Mark Zuckerberg discusses Meta’s rollout of Llama 3 and the upgraded Meta AI assistant, arguing that open source AI plus massive infrastructure investment will shape the next computing platform. He explains Meta’s GPU and data-center strategy, why AGI is a gradual capability stack, how open source can counter concentration risks, and why AI will transform products, creators, business tools, and science.

Main Topics: Llama 3 and the new Meta AI rollout (Priority: 5/5): Meta is rolling out Llama 3 as the model powering Meta AI, integrating Google and Bing for real-time knowledge and adding features like image animation and real-time image generation across WhatsApp, Instagram, Facebook, and Messenger. Why Meta invested heavily in GPUs and infrastructure (Priority: 5/5): Zuckerberg says Meta bought GPUs to avoid being caught short again after needing more compute for Reels ranking and unconnected content recommendations, and to prepare for future unknown demands like large-model training. AGI as a collection of capabilities, not one threshold (Priority: 5/5): He argues AI progress is progressive: reasoning, coding, multimodality, memory, emotional understanding, and efficiency each matter separately, and together they move systems toward general intelligence. Open source strategy and concentration risk (Priority: 4/5): Meta sees open source as beneficial for innovation, cost reduction, ecosystem development, and geopolitical balance, while acknowledging limits if future systems become qualitatively too risky to release. Inference, synthetic data, and scaling bottlenecks (Priority: 4/5): Zuckerberg emphasizes that inference will dominate compute demand as AI does more multi-step work and generates synthetic data; scaling may eventually hit energy, regulatory, and data-center constraints rather than pure money limits. Metaverse, digital presence, and long-term product vision (Priority: 3/5): He reiterates that the metaverse remains important mainly for realistic digital presence and human connection, while AI is the near-term and longer-term platform shift that will touch all Meta products. Personal conviction, history, and building culture (Priority: 3/5): Zuckerberg links his decisions to a builder identity shaped by computer science, psychology, history, and a personal drive to keep creating new things rather than optimizing only for financial exit.

Key Arguments: Llama 3 is the upgrade that makes Meta AI the most intelligent freely available assistant in Meta’s view, because the model gains stronger reasoning, coding, and tool use. Open source helps Meta because outside developers can improve efficiency, build useful derivatives, and create a broader standard that reduces dependence on closed AI gatekeepers. Meta’s GPU purchases were driven first by Reels/recommendations needs, but the company deliberately doubled capacity because it expects future horizons to emerge before they are obvious. AGI should be understood as an accumulation of distinct capabilities—reasoning, multimodality, memory, emotional understanding, and efficiency—not a single magical threshold. The company believes AI will be used at industrial scale mainly through assistants, business/creator agents, and synthetic-data-driven workflows, not just chatbot-style Q&A. Training and deployment will face real-world limits: energy permitting, data-center build times, and physical infrastructure will become major bottlenecks even if capital is available. Meta is pro-open-source in principle, but would reconsider if a future model crossed a qualitative safety boundary that could not be mitigated responsibly. Zuckerberg views concentration of advanced AI in a few actors as a major strategic and security risk; broad deployment and open standards are a stabilizing force. The metaverse’s core value remains enabling presence with other people, but AI is the more immediate civilization-scale shift. Meta’s open-source model may create revenue via cloud-hosted licensing and enterprise/cloud partnerships, especially for larger models used at scale.

Data Points: Llama 3 model sizes: 8B, 70B, and a 405B dense model - Meta is releasing the 8B and 70B versions now; the 405B is still training. Llama 3 70B benchmark: ~82 MMLU - Zuckerberg cites the 70B model as leading at its scale, especially on math and reasoning. Llama 3 8B benchmark position: Nearly as powerful as the largest Llama 2 model - He emphasizes the efficiency jump from Llama 2 to Llama 3. 405B model benchmark: Around 85 MMLU (in training) - He says the 405B is already around this level while still training. Training data for 70B: ~15 trillion tokens - He says the model kept learning by the end, suggesting more data could still help. Creators on Meta platforms: About 200 million - Used to explain why creator-facing AI agents could be highly valuable. AI-related and infrastructure scale: 22,000–24,000 training clusters - He says Meta has built large training clusters alongside broader fleet usage. Total GPU fleet target: 350,000 GPUs by end of year - Mentioned as part of Meta’s large-scale compute buildout. Typical data-center scale today: 50MW to 150MW - He contrasts current facilities with the larger power needs he expects in the future. Future data-center scale discussed: 300MW, 500MW, and eventually 1GW - He argues AI infrastructure will run into energy and permitting constraints at these sizes. Social media harm taxonomy: 18 or 19 categories - He references Meta’s existing safety classification work as a model for AI safety mitigation.

Pivotal Quotes: "we now think that Meta AI is the most intelligent AI assistant that people can use that's freely available" — Mark Zuckerberg: Describing the significance of Llama 3 powering Meta AI "I think that all this stuff is going to be progressive over time" — Mark Zuckerberg: On AGI, intelligence, and why progress will come in capability increments rather than one sudden threshold "the day I stop trying to build new things, I'm just done" — Mark Zuckerberg: Explaining his personal motivation for taking long-term bets like AI and the metaverse

Implications: Meta is betting that open-source, capability-rich AI will become the default infrastructure layer for consumer apps, businesses, and creators. For the industry, the next constraints are likely energy, compute, and governance—not just model quality.

🔓 Sign Up for Unlimited Episode Search

About Dwarkesh Podcast

Deeply researched interviews

View all episodes from Dwarkesh Podcast