scripod.com

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

Dwarkesh Podcast

Shownote

Zuck on: * Llama 4, benchmark gaming * Intelligence explosion, business models for AGI * DeepSeek/China, export controls, & Trump * Orion glasses, AI relationships, and preventing reward-hacking from our tech. Watch on Youtube; listen on Apple Podcast...

Highlights

In this podcast, Zuck discusses various aspects of AI development, including the introduction of Llama 4 and its capabilities compared to other models. The conversation also delves into broader topics such as the intelligence explosion, business models for AGI, geopolitical considerations with DeepSeek and China, and the societal impacts of AI in relationships and productivity.
08:35
Open-source models benefit from community scrutiny and diverse optimizations.
17:17
For AI to be useful, it must generate higher-quality hypotheses than humans.
29:33
Humans need tools to keep up with AGI, but reward-hacking is a concern.
35:16
Llama 4 expected to be more competitive in multimodal capabilities.
47:57
Language models encode values and reasoning capabilities.
56:51
AI services can be offered with both free and premium models.
1:01:13
Design systems for best AI models with personalization, low latency, and full duplex voice.
1:02:08
Deferred too much to media and government in the past.
1:10:27
Technology has historically led to more job opportunities rather than fewer.

Chapters

How Llama 4 compares to other models
00:00
Intelligence explosion
11:34
AI friends, therapists & girlfriends
26:36
DeepSeek & China
35:10
Open source AI
39:49
Monetizing AGI
54:15
The role of a CEO
58:32
Is big tech aligning with Trump?
1:02:04
100x productivity
1:07:10

Transcript

Dwarkesh Patel: All right, Mark, thanks for coming on the podcast again. Mark Zuckerberg: Yeah, happy to do it. Good to see you. Dwarkesh Patel: You too. Last time you were here, you had launched Llama 3. Yeah. Now you've launched Llama 4, Well, the firs...