Episode Summary
Executive Summary: Noam Shazeer discusses Character.ai’s vision as a full-stack, direct-to-consumer AI platform built on the belief that general models should be launched broadly and let users invent use cases. He argues AI progress is chiefly limited by compute, that the field is still early, and that hallucinations, when framed as creativity, can be useful for entertainment, companionship, and support.
Main Topics: Google lessons and the power of B2C scale (Priority: 5/5): Shazeer reflects on 20 years at Google, noting that the biggest lesson was serving a general technology to billions of users rather than optimizing for enterprise only. He sees Character.ai as following the same direct-to-consumer logic. Character.ai’s mission and product philosophy (Priority: 5/5): The company aims to provide a highly general, easy-to-use AI tool and let users create a billion use cases. Shazeer emphasizes user agency and avoiding over-prescribing the product’s purpose. Model scaling, compute, and hardware as the main bottleneck (Priority: 5/5): He argues the key challenge is not just data or model size, but the amount of computation required to train larger models longer. Better hardware and more compute are the biggest unlocks. Use cases, companionship, and human connection (Priority: 4/5): Shazeer describes surprising emergent use cases, especially entertainment, emotional support, and therapy-like interactions. He believes AI should complement, not replace, human relationships. Open vs. closed AI ecosystems and startup advantage (Priority: 4/5): He suggests startups can move faster than incumbents, while the ecosystem will likely support many winners. He also notes that scale creates economic advantages in serving models and training them. Hallucinations, creativity, and the early-state of AI (Priority: 4/5): He treats hallucinations as a feature for certain applications, especially creative and conversational ones, and argues that AI is still at an early 'invention of electricity' stage. Personal growth, parenting, and responsibility (Priority: 2/5): The conversation closes with Shazeer discussing how parenthood changed his outlook, making him more reflective about responsibility, gratitude, and what matters most.
Key Arguments: General-purpose AI should be launched directly to consumers because users, not builders, discover the best applications. The biggest constraint on better models is compute, not simply data; larger models trained longer require massively more operations. Hallucinations are not always a flaw; in entertainment and companionship contexts they can be valuable and even desirable. Character.ai should respect user agency rather than dictate what AI is for. Startups can outpace incumbents because they can ship faster and avoid protecting legacy products. The best applications of AI have not yet been invented; current products are analogous to early electricity or the first airplane. Human-to-human connection matters, and AI should ideally help socially anxious users practice rather than isolate them further.
Data Points: Google tenure: 20 years - Shazeer spent two decades at Google before founding Character.ai. Character.ai messages per day: 450 million - Used to illustrate the platform’s scale and engagement. Character.ai users: 20 million - Cited as evidence of broad consumer demand. Model training cost: $2 million - Shazeer said the model they were serving was trained last summer using about this much compute spend. Training data source: billions to trillions of examples - He described language modeling as benefiting from huge volumes of web text for next-word prediction. Adoption timeline: 1 to 3 years - He expects very cool AI developments to arrive quickly in the near term. Market yield mentioned in ads: 5.5% - A sponsor mention referenced 26-week Treasury bill yields available at the time. Potential cost savings in travel sponsor: up to 30% - A sponsor mention claimed business travel cost reductions through Navan.
Pivotal Quotes: "the number one thing that's important is how much computation you do to train it" — Noam Shazeer: Explaining what most constrains progress in model quality. "a billion users inventing a billion use cases" — Noam Shazeer: Character.ai’s core mission and product strategy. "We consider them a feature" — Noam Shazeer: His view that hallucinations can be useful in creative and conversational settings.
Implications: AI progress will likely be driven by compute, scale, and open-ended consumer usage. For builders, the lesson is to ship general tools, observe real behavior, and expect unexpected use cases to define the market.