Mark Zuckerberg: Future of AI at Meta, Facebook, Instagram, and WhatsApp | Lex Fridman Podcast #383

In his second appearance, Zuckerberg discusses Meta's approach to AI, centered on open-sourcing models like Llama and a contrarian belief that the future holds many specialized AIs rather than one singular super-intelligence. The conversation covers content moderation, government pressure, misinformation, and the difficulty of drawing lines between harm and free expression. He details Meta's layoffs and a renewed focus on lean, engineering-driven teams, then pivots to the newly announced Quest 3 mixed-reality headset and a comparison with Apple's Vision Pro. The episode closes on existential AI risk, the distinction between intelligence and autonomy, mortality, physical training, and faith.

• Reveals he competed in his first jiu-jitsu tournament in disguise, registering under his first and middle name (Mark Elliot) wearing a hat, sunglasses, and COVID mask.
  00:02:04
• Confirms the first Llama model 'leaked' after a limited research release and says companies have generally respected the no-commercial-use terms.
  00:21:46
• Stakes out his core contrarian thesis: unlike OpenAI and Google's 'One Singular AI,' he believes people will want many different AIs for different roles.
  00:43:54
• Criticizes the pandemic 'establishment' for waffling on facts and asking for things to be censored that later turned out to be debatable or true.
  01:14:36
• Confirms a text-based Twitter competitor (code-named P92) is in the works, exploring a federated approach tied to Instagram identity.
  01:33:25
• Praises Elon Musk's leaner, more technical restructuring of Twitter and says it influenced his own org changes.
  01:41:14
• Announces Quest 3 at $499, launching fall 2023, with high-resolution mixed reality as the headline new capability.
  01:57:32
• When asked if building AGI could make him the most powerful man in the world and corrupt him, he argues open-sourcing is the safeguard against power imbalance.
  02:17:25

• Meta runs roughly 18 different categories of systems to detect violating content like terrorism, child exploitation, and incitement to violence.
  01:13:00
• Zuckerberg says he now writes about 80-90% of his code AI-generated first, then edits.
  01:00:35
• Meta's translation work expanded text/speech tools from ~100 languages to more than 1,000, and the model can identify over 4,000 spoken languages.
  00:29:11
• Cites demographic studies showing the average person has fewer close friends today than 15 years ago.
  00:53:17
• Meta has sold tens of millions of Quest devices; he estimates Apple is building only about 1 million Vision Pro units.
  02:06:56
• Quest 3 is 40% thinner than Quest 2 with 2x the graphics processing power and 40% sharper screens.
  02:03:47
• At the start of the restructuring, Meta managers averaged about three reports; he wants to move closer to seven or eight to cut information latency.
  01:49:37
• He ran a sub-40-minute Murph Challenge (1 mile, 100 pull-ups, 200 push-ups, 300 squats, 1 mile) and notes a friend beat him by several minutes.
  02:28:19
• Offers a brain analogy: a super-intelligent system could be like the neocortex, a powerful tool serving a simpler decision-making 'old brain,' separating intelligence from autonomy.
  02:14:16
• Argues entities like companies and the stock market are already greater-than-human distributed intelligences without bodies or language.
  02:24:40