Zuckerberg and Alexandr Wang outline Meta's personal superintelligence strategy
Podcast notes say Meta trains a multimodal, memory-focused personal assistant from the product side first, gives Wonder to paying users first, and spends on compute ahead of product validation.
Great episode. Meta's $100B+ AI bet is built around PERSONAL superintelligence: an agent that knows your context, remembers it and gets things done for you. Zuckerberg and Wang on Cheeky Pint: >They see room for a consumer-focused frontier lab while much of the industry's attention goes to coding and knowledge-work automation. The model is trained for that product from the beginning. >That changes what they optimize for: multimodal perception, memory across conversations, persistence through setbacks and discretion with private information. Their evals are built around whether those capabilities make the agent useful. >Coding is central to this too. Zuckerberg's MMA agent writes its own video-processing pipeline, then has to deliver useful feedback inside a 60-second break. The consumer never needs to see the code. >Meta starts with distribution to billions of consumers, years of experience in recommendations and multimodal data. Glasses can extend that into an agent that sees and hears the world with you. Owning the models gives Meta more control over how those experiences work. >The commercial plan spans personal agents, small-business tools, commerce, APIs and enterprise access, alongside improvements to Meta's existing business. They're trying to apply the same underlying model investment across a very large platform. >Wonder is larger and more expensive to serve, so paid subscribers get it first. The goal is to bring comparable capability into the smaller Spark form factor over time and serve it much more broadly. >The infrastructure comes before product proof. Zuckerberg says Meta committed inference capacity years before knowing Muse demand. His fallback is that other customers will want the compute if Meta's own models disappoint but AI demand holds up. >His rough estimate: $50–60B to build a 1 GW cluster and ~$15B to operate it for a year. He explicitly corrects John when the annual figure gets translated into a training bill. That's all-in, it seems. We still need retention, paid conversion and cost per completed task to assess the returns, but the strategy is clearly organized around making personal agents useful at Meta's scale. They are laser focused and executing...
This post quotes another post: x.com/collision/status/2109714250226905248