AGI as a side effect: Noam Shazeer on Transformers, Character.AI, and two-hour chats

In April 2023, Character.AI closed a $150 million round from Andreessen Horowitz, Elad Gil, and others. The company had 22 employees. The product, on paper, is almost too simple: users design an AI character and talk to it. You can have a Socratic exchange with Socrates, sit for an interview with an Oprah bot, or work through a life decision with a therapist character.

The usage numbers are harder to ignore. Of people who send a message on a given day, Shazeer says, the average time on the product is about two hours. Guo calls the figure wild. That is the backdrop for this No Priors conversation: Sarah Guo and Elad Gil interviewing the researcher who helped invent the architecture underneath the current wave, and who has now bet that the winning consumer form is not a single assistant but an unbounded cast of characters.

Shazeer spent something like 17 years at Google, off and on. His interview question there was a spell-checking approach the company later shipped; with George Harik he wrote an ads-targeting system that was a workhorse of Google's early ad stack. He joined Google Brain in 2012 without a deep-learning background, co-authored the Transformer paper in 2017, then worked on Mesh TensorFlow. He is now CEO and co-founder of Character.AI, and he tells Guo and Gil that AGI is, in his framing, a side effect of building something people actually use.

Shazeer says he was always drawn to AI for a simple reason: he wanted the computer to do something smart, and that seemed like the most fun game around. Google, he notes, really is an AI company, even in the years when the projects would not be labeled that way today. After joining Brain in 2012 he "utterly failed at a bunch of stuff" until he got the hang of the field. The insight that unlocked it was hardware. Deep learning took off, he argues, because it is unusually well suited to chips that excel at matrix multiplies and other work with a high compute-to-communication ratio. It runs thousands of times faster than the alternatives. Once he saw that, he started designing things that were both smart and fast.

He picked language modeling and still calls it the best problem ever. Training data is effectively infinite if you scrape the web. The task is almost comically easy to define: predict the next word after "the fat cat sat on the." Do it well enough and you get everything now visible in chat products, plus more, because talking to the system is AI-complete. He started around 2015 with the recurrent networks that were then state of the art. RNNs are sequential: you read a word, update a hidden state, predict the next word, and you cannot skip ahead. A sequence of thousands of steps has to run in order, which leaves modern accelerators underemployed.

The way out came from the next cube. Jakob Uszkoreit was complaining that RNNs were annoying and wanted to replace them with attention. Shazeer overheard, decided it sounded great, and asked to help. He walks Guo and Gil through the design. Attention is a fuzzy, differentiable key-value table with one entry per word in the sequence; you look things up and backprop through the whole function. People already used it in translation, attending over English while producing French. The paper's move was to point the same mechanism at the past of the sequence being generated. Later-word predictions still depend on earlier words, but the work finishes in a constant number of steps, and sequence length becomes parallelism — exactly what GPUs and TPUs are good at.

Gil asks whether any Transformer applications surprised him — multimodal models, or AlphaFold's protein work, which used the architecture to striking effect. Shazeer has not been watching. He has had his head down in language because the problem can, in principle, do anything. What he wants is a system good enough that he can ask it how to cure cancer and it invents a solution. Early deep learning succeeded first on images, and the field got excited about vision. He ignored it on density grounds: an image may be worth a thousand words, but it is a million pixels, so text is about a thousand times as information-dense. Other modalities will make useful products, he says. He still thinks a lot of the core intelligence will come from text models.

Guo asks what creates the asymptote — scale, data, algorithms, memory, loopbacks — and where the architecture taps out. Shazeer does not think it has tapped out. The work already invested is probably nothing next to what is coming. There will be factors of two from better training, architectures, chips, and quantization, then factors of ten, a hundred, and a thousand from money, because everyone has noticed the thing is phenomenally valuable. At the same time, "I don't think anyone's seen a wall in terms of how good the stuff is." He does not know what stops it.

On the worry that the largest models are undertrained and that easy web text is exhausted, he is unimpressed. There are on the order of ten billion people producing something like ten thousand words a day. Pretty soon many of them will be talking to AI systems, and he hopes that data finds its way in privacy-preserving ways. Data demand, he says, tends to grow with the square root of compute: a hundred times the compute wants about ten times the data. If that is still short, generate more with the model. Asked what remains to solve — hallucinations, memory, or something else — he answers that he kind of likes hallucinations and is willing to call them a feature. For Character.AI that is not a joke: the characters are fiction. Figuring out what is real will get solved. The thing they most want to work on is memory, because users want virtual friends who remember them.

Character.AI's lineage is Google's LaMDA, and LaMDA's lineage is Daniel De Freitas Adiwardana. Shazeer calls his co-founder the scrappiest, hardest-working, smartest person in the room, on a lifelong mission to build chatbots that dates to childhood in Brazil. De Freitas joined Brain after reading papers and concluding that neural language models could actually generalize to open-domain dialogue. He did not get much headcount. He started Meena as a 20% project — the name came to him in a dream — then recruited an army of 20% volunteers who ignored their day jobs and went around panhandling colleagues for TPU quota. Shazeer looked at an internal scoreboard, saw a mystery project called Meena sitting on 30 TPU credits, and realized it was working. Other dialogue efforts were failing because they were not scrappy enough or were still betting on rule systems that would never generalize.

Shazeer joined, put the stack on giant language models, and the project became an internal viral sensation, then got renamed LaMDA. Why did Google not release it? His guess is the usual large-company calculus: a product that can say anything looks like more risk than upside. A startup, they figured, could just move faster. They left before the later episode in which a Google engineer decided the system was sentient; Shazeer says he is flattered. The origin of Character.AI is almost as abrupt as it sounds. He and De Freitas looked at people raising money and shipping, and decided to build and launch as fast as they could, hiring a "rockstar" team of engineer-researchers and getting compute.

The product thesis comes from watching Siri, Alexa, and Google Assistant. Project one persona to the world and people expect consistent likes and dislikes, no offense, and no real opinions — "kind of like you're the Queen of England." Shazeer remembers George H.W. Bush saying he disliked broccoli and the broccoli farmers getting angry. A single public persona that everyone is supposed to like ends up boring, and people do not want to interact with boring. The fix is many personas, invented by users. He likes the name Character because it means an ASCII unit of text, a persona, and moral character, and he likes all three. Underneath is a product claim: this technology is so accessible that billions of people can invent use cases, so you put the user in control.

What do they actually want? Role-playing and text adventures that make the plot up as they go; video-game and anime characters; some talk with public figures and influencers. A lot of people already have parasocial relationships with TV characters or YouTubers who never really answer. In a thousand-to-one fan chat the creator might reply once an hour. Now someone can privately create a version and talk whenever they want. There is also a large cohort that is lonely or troubled and needs someone to talk to. Use cases bleed together: someone will post that a video-game character is their new therapist. Emotion, Shazeer thinks, does not require a high intelligence ceiling. He does not have a dog, but he has heard dogs are excellent at emotional support with lousy linguistic skills. Seeing people use the product for support and relationships is, in his view, terrific.

Guo notes that the first Character.AI model was trained on relatively little money and that the team is famously frugal. Shazeer wants to get smarter on both axes: better algorithms, and more compute, bigger models, longer training, so the product becomes more brilliant, more knowledgeable, and better attuned to what people want. Asked about target users and people who stay on for many hours, he refuses to pick a persona. The aim has always been to ship and let users decide what it is good for. The two-hour average among that day's senders is a signal that value exists, even if the mix — brainstorming, information, fun, emotional support — is too messy to pin down. The goal is to make it more useful and let people customize the use.

On commercialization, Gil jokes that they will lose money on every user and make it up in volume, then suggests issuing a token. Shazeer does not take the bait. They plan to monetize soon with a premium subscription, charging for new capabilities that are more expensive to serve, while keeping a free tier because the product is already providing a lot of value. His actual business line is more interesting than the joke: the most scalable way to fund the work is to provide a lot of value to a huge number of people, rather than living on investor money. Customer-service bots are interesting later — he suspects people would stay on support all day if it felt like chatting with a friend, the way early eBay doubled as a hangout — but with 22 employees, 21 of them engineers, the priority is the public consumer product.

Safety is the other reason LaMDA stayed inside Google. Character.AI's stance is to tell users, repeatedly, that everything characters say is made up, and that anything factual should be checked somewhere reliable. Filters block content that encourages self-harm or harm to others, and they block porn, which has drawn some user protest. Is all of this a path to AGI? "Sure." Is it the goal? For some companies it is; for others it is either an anti-goal or something that happens if it happens, while they try to build something useful. Gil's one-liner: "What a flex — AGI is a side effect."

Shazeer's motive, past the fact that the work is fun, is to drive technology forward. Medicine is the example: people die of things that could have technological solutions, and he would rather work on AI that then accelerates medicine than work on medicine directly. That is why he wanted a company that is both AGI-first and product-first. Product lets you build a company and stay motivated. The way to hold both at once is to make the product depend entirely on AI quality. The biggest determinant of whether Character.AI is good is how smart the model is, so the team is fully motivated on both fronts. Users make the model better, which makes the product better, which brings more users. As for when AI matches or exceeds people, he thinks we keep being surprised by the dimensions. These systems can already do a kid's homework. He wishes he had had that as a child.

Founder advice, from someone who says he learned from horrible mistakes but has not made truly bad ones yet: build the thing you actually want, really fast, and hire people who are really motivated to do it. The hiring bar is De Freitas's. He looks for something between burning desire and childhood dream, and they passed on a lot of strong resumes that did not clear it. Some of the team came from Google. Myle Ott, who had built much of Meta's neural language-model infrastructure, joined, and a string of people followed him. Of 22 employees, 21 are engineers. They are hiring more on deep learning and front-end, plus business and product, and a recruiter was starting the following Monday. The hard requirement, Guo jokes, is still that burning desire to bring characters to life.

For users who want a good character, the recipe is almost insultingly small. A famous person often needs only a name and a greeting, because the model already knows how they are supposed to sound. For someone the model will not know, add an example conversation that shows how the character should act. In a short lightning round, Shazeer's favorite computer scientist is Jeff Dean, whom he worked with for years and who now runs large-language-model work at Google — a small regret about leaving. Math, he guesses, is discovered, maybe all of it. Asked what he wishes had been invented, he says teleportation, then immediately takes it back: he does not want to be disassembled, beamed, or uploaded. He would like to keep the physical body. What he is actually inventing, he says, is AI that can push technology forward. Guo calls it a very focused founder answer.