Mark Zuckerberg — AI will write most Meta code in 18 months - Dwarkesh Podcast Recap

Podcast: Dwarkesh Podcast

Published: 2025-04-29

Duration: 1 hr 15 min

Summary

Mark Zuckerberg discusses the rapid advancements in AI at Meta, highlighting how the company is on track to enable AI to write a significant portion of its code in just 18 months. He emphasizes the growth of open-source models and the importance of user value in developing AI technologies.

What Happened

In this episode, Mark Zuckerberg returns to discuss the latest developments at Meta, particularly the launch of Llama 4. He reveals that Meta AI now boasts almost a billion monthly users, marking a significant milestone for the platform. Zuckerberg expresses excitement about the new models being released, including the Scout and Maverick, which are designed for efficiency and low latency, crucial for consumer applications. Additionally, he shares insights into the upcoming Behemoth model, which will feature over 2 trillion parameters, posing challenges for usability and infrastructure.

Zuckerberg notes that there has been a notable shift in the AI landscape, with open-source models gaining traction and potentially surpassing closed-source ones in popularity. He stresses the importance of latency and cost-effectiveness in AI applications, particularly for consumer products. The conversation also touches on the reasoning phenomenon and how specialized models for specific tasks, such as coding or math, are emerging. Zuckerberg is optimistic about the integration of reasoning capabilities with core language models, reflecting industry trends towards creating more intelligent and versatile AI systems.

Key Insights

Key Questions Answered

What are the new features of Llama 4?

Zuckerberg discusses the launch of Llama 4, which includes the Scout and Maverick models that are designed to be efficient and low-latency. He mentions that these models represent some of the highest intelligence per cost available. Additionally, he reveals that a new model, internally called Little Llama, is on the way, which will follow the popular 8 billion parameter model from Llama 3.

How is Meta AI expected to evolve in the next 18 months?

Zuckerberg predicts that AI will soon be able to write a majority of Meta's code, suggesting a significant shift in how software development is approached at the company. He emphasizes the roadmap for Llama 4 and the advancements that have been made since the previous models, indicating that this year will be pivotal for development in AI at Meta.

What is the significance of open-source AI models according to Zuckerberg?

Zuckerberg believes that this year marks a turning point for open-source AI, suggesting that these models are becoming more viable and could potentially outpace closed-source models in usage. He points out the growing number of high-quality open-source models available, which reflects a broader trend in the AI community towards open collaboration and sharing of resources.

What challenges does the Behemoth model present?

The Behemoth model, featuring more than 2 trillion parameters, poses significant challenges in terms of infrastructure and usability. Zuckerberg notes that the team is working on how developers can effectively utilize such a large model, as it is impractical for consumer applications. He highlights the need to distill the model into more manageable sizes while still providing value.

How does Meta AI prioritize user value in its models?

Zuckerberg emphasizes that the primary benchmark for Meta AI is user value. He shares that while various benchmarks exist for comparing models, they often do not align with the specific needs and desires of users. By focusing on what users actually value, Meta aims to create AI products that not only perform well but also provide meaningful and timely responses in consumer applications.