Your AI Friends Have Awoken, With Noam Shazeer

0:05Do you view all this as a path to AGI or superintelligence? Sure, yeah. And is that part of the goal? For some companies, it seems like it's part of the goal. And for some companies, it seems like it's either not explicitly an anti-goal or if it happens, it happens. And, you know, the thing people are trying to build is just something useful for people. What a flex. AGI is a side effect. Yeah, well, I mean, that was a lot of the motivations here. Because, like, I mean, my main motivation for working on AI, other than that it's fun. Well, I mean, fun is secondary. Like, the real thing is, like, I want to drive technology forward.

0:42This is the No Priors podcast. I'm Sarah Gore. I'm Alad Gale. We invest in, advise, and help start technology companies. In this podcast, we're talking with the leading founders and researchers in AI about the biggest questions.

1:01Transformers, large language models, AI chat. These are the foundations supporting today's AI revolution. And this week on No Priors, we have an AI researcher, engineer, and inventor who is a key part to these innovations and is considered one of the smartest people in AI. Noam Shazir is the CEO and co-founder of Character AI, a service that allows users to design and interact with their own personal bots that take on the personalities of well-known individuals or archetypes. You could have a Socratic conversation with Socrates, or you could pretend you're being interviewed by Oprah. Or you could work through a life decision with a therapist bot.

1:33Character recently raised$150 million from Andreessen Horwitz, myself, and others. We talk about how Noam got his start at Google, his groundbreaking AI discoveries, and what he's doing at Character. So Noam, welcome to No Priors. Hey, Alad. Thanks for having me on. Hi, Sarah. Good to see you. Yeah, thanks for joining. So you've been working on NLP and AI for a long time. So I think you were at Google for something like 17 years off and on. And I think even your Google interview question was something around spellchecking, an approach that eventually got implemented there. And when I joined Google, one of the main systems being used at the time for ads targeting was like fill and fill clusters and all the stuff, which I think you wrote with George Herrick.

2:07And so it'd just be great to get kind of your history in terms of working on AI, NLP, language models, how this all evolved, what you got started on, and what sparked your interest. Oh, thanks a lot. Yeah, just always naturally drawn to AI. Wanted to make the computer do something smart. Seems like pretty much the most fun game around. Was lucky to find Google early on and really is an AI company. So yeah, got involved in a lot of the early projects there that maybe you wouldn't call AI now, but seemed pretty smart at the time. And then more recently was on the Google brain team starting in 2012.

2:43It looked like a really smart group of people doing something interesting. I had never done deep learning before or neural networks, I guess, as it was called then. I forget when the rebrand happened. But yeah, it turned out to be really fun. That's cool. And then, you know, you were one of the main people working on the Transformer paper and design in 2017. And then you worked on Mesh TensorFlow, I think, sometime within the following year. Could you talk a little bit about how all that got going? Yeah, I mean, I messed around a few years on the Google Brain team and like utterly failed at a bunch of stuff until I kind of got the hang of it.

3:17really the key insight is that what makes deep learning work is that it is really well suited to modern hardware, where, you know, you have the current generation of chips that are great at matrix multiplies and, you know, other forms of things that require large amounts of computation relative to communication. So basically, deep learning, like really took off because, you know, It runs thousands of times faster than anything else. And as soon as I got the hang of out, started designing things that actually were smart and ran fast. But the most exciting problem out there is language modeling.

3:56It's the best problem ever because there's an infinite amount of data. Just scrape the web and you've got all the training data you could ever hope for. And the problem is super simple to define. It's predict the next word. The fat cat sat on the, what comes next? It's extremely easy to define. And if you can do a great job of it, then you get everything that you're seeing right now and more. You can just talk to the thing and it's really AI complete. And so got started around like 2015 or so, working on language modeling and messing with the recurrent neural networks, which was what was great then.

4:33And then Transformer kind of came about as someone had the bright idea. But Jakob was great that, hey, these RNNs are just annoying. Let's try to replace them with something better, you know, in the tension. And I overheard a couple of colleagues talking about it in the next cube over. I was like, that sounds great. Let me help you guys. You know, these RNNs are annoying. It's going to be so much more fun. Can you quickly describe sort of the difference between an RNN and a transformer-based or attention-based model? Yeah, sure. Okay. Okay, so like the recurrent neural network is the sequential computation where every word, you read the next word, and you kind of compute your current, you know, state of your brain based on the old state of your brain and what this next word is, and then you predict the next word.

5:18So you have this very long sequence of computations that has to be executed in order. So that, you know, the magic of transformer, kind of like convolutions, is that you get to process the whole sequence at once. I mean, it's still a function of the predictions for the later words are dependent on what the earlier words are. But it happens in like a constant number of steps where you get to take advantage of this parallelism of you can look at like the whole thing at once. And that's what modern hardware is good at, is parallelism. And now you can use the length of the sequence as your parallelism, and everything works super well.

5:56Attention itself, it's kind of like you're creating this big key value associative memory, where you're building this big table with one entry for every word in the sequence. And then you're kind of looking things up in that table. It's all fuzzy and differentiable and a big differentiable function that you can backprop through. And people had been using this for problems where there are two sequences, where you've got machine translation. You're translating English to French. And so while you're producing the French sequence, you are looking over the English sequence and trying to pay attention to the right place in that sequence.

6:34But the insight here was, hey, you can use the same attention thing to look back at the past of the sequence that you're trying to produce. And the beauty is that it runs great on GPUs and TPUs. And it's kind of parallel to like how deep learning is taken off because it's great on the hardware that exists. And this sort of brought the same thing to sequences. Yeah, I think the classic example to help people picture it is like, you know, saying the same sentence in French and English. The ordering of the words is different. You're not mapping like one to one in that sequence. And to figure out how to do that with parallel computation without information loss is like a really elegant thing.

7:12Yep. Yeah. It seems like the technology has also been applied in a variety of different areas. The obvious ones are these multimodal language models. So it's things like ChatGPT or what you're doing a character. I've also been surprised by some of the applications into things like AlphaFold, the protein folding efforts that Google did, where it actually worked in an enormously performant way. Are there any application areas that you found really unexpected relative to how transformers work and relative to what they can do? Oh, I've just had my head down in language. Like here you have a problem that can do like anything.

7:43I want this thing to be good enough. So I just ask it, how do you cure cancer? And it like invents a solution. So I've been totally ignoring what everybody's been doing in all these other modalities where I think a lot of the early successes in deep learning have been like in images. And people are like all excited about images and kind of completely ignored it because, you know, an image is worth a thousand words, but it's a million pixels. So the text is like a thousand times as dense. So kind of big text nerd here, but very exciting to see it take off in all these other modalities as well. And those things are going to be great.

8:14It's super useful for building products that people want to use. But I think a lot of the core intelligence is going to come from these text models. Where do you think the limitations for these models, what do you think creates the asymptote that all this is being built against? Because people often talk about just scale, like you just throw more compute and this thing will scale further. there's data and different types of data that may or may not be available. There's algorithmic tweaks. There's adding new things like memory or loopbacks or things like that. What do you think are the big things that people still need to build against?

8:42And where do you think this sort of taps out as an architecture? Yeah, I don't know that it taps out. I mean, we haven't seen the tap out yet. The amount of work that has gone into it is probably nothing compared to the amount of work that will go into it. So quite possibly, there will be all kinds of like factors of two and efficiency that people are going to get through better training algorithms, better model architectures, better ways of building chips and using quantization and all of that. And then there are going to be factors of 10 and 100 and 1 ,000 of just scaling and money that people are just going to throw into the thing because, hey, everyone just realized this thing is phenomenally valuable.

9:21At the same time, I don't think anyone's seen a wall in terms of how good this stuff is. So I think it's just going to keep getting better. I don't know what stops it. What do you think about this sort of idea that we can increase compute, but the largest models are under-trained? We've used all the text data on the internet that's easily accessible. We have to go improve the quality. We have to go do human feedback. How do you think about that? Yeah, I mean, in terms of getting some more data, like there are a lot of people talking all the time. I mean, why do you think we do this podcast? Right.

9:57Like there's like order 10 billion people like producing about, you know, like, I don't know, 10 ,000 words a day. I mean, that's like a lot of words that, you know, and pretty soon many of those people will be doing a lot of that talking to AI systems. So I have a feeling like a lot of data is going to find its way into some AI systems. I mean, in privacy-preserving ways, I would hope. And then the data requirements tend to go up with the square root of the amount of computation because you're going to train a bigger model and then you're going to throw more data at it. I'm not that worried about coming up with data.

10:32And I feel like we could probably just generate some more with the AI. And then what do you think are the main things to solve for these models going forward? Is it hallucinations? Is it memory? Is it something else? I don't know. I kind of like hallucinations. It's also a feature, yeah. They're fun. Yeah, we'll call it a feature. Yeah, some of the things we want to work on the most are memory because our users definitely want their virtual friends to remember them. There's so much you can do with personalization and you want to dump in a lot of data and use it efficiently. Yeah, there's a ton of great work going on and trying to figure out what's real and what's hallucinated, of course.

11:11I think we'll solve this. And then do you want to talk a little bit about Lambda and your role with it and how that led eventually to character? Yeah, my co-founder, Daniel De Freitas, he's like the scrappiest, most hardworking, really smartest guy. He's kind of been on this lifelong mission to build chatbots. Like since he was like a kid in Brazil, he's like always been trying to build chatbots. So he came to join us at Google Brain because I think he had read some papers and figured that this neural language model technology would be like, you know, something that could actually generalize and build something truly open domain.

11:44So, and he did not get a lot of headcount. He started the thing as a 20 % project where like people are encouraged to spend 20 % of their time, like doing whatever they want. And then he just recruited like an army of 20 % helpers who were like ignoring their day jobs and like actually just, you know, helping him with the system. And he went as far as going around and panhandling people's TPU quota. Like he called this project Mina because I guess it came to him in a dream. And like at some point I'm looking at the scoreboard and what is this thing called Mina and why does it have 30 TPU credits?

12:20And it was just gotten a bunch of people to contribute. And then he was really successful at this because, you know, in building something really cool that actually worked where like a lot of other systems were just like totally failing either because people were not, you know, just weren't scrappy enough or were going for like rule based systems that were just never going to generalize. So at some point I was like, okay, there are so many ways we can make this technology better by like factors of two. But the biggest thing is just convince everyone that this is like worth trillions of dollars by demonstrating some application that is clearly super valuable to like billions of people.

12:59And Lambda was this, I believe, is the internal chatbot pre-GPT at Google that was famously in the news because an engineer thought it'd become sentient, right? Yeah, yeah, yeah. So that was like a renaming of Mina. And so I guess I went and helped Daniel on Mina. We got it on some giant language models and then kind of became like an internal viral sensation and then got renamed to Lambda. And yeah, we had left before the business about somebody thought it was sentient. I'm flattered. Can you talk a little bit about just why it wasn't released? What some of the concerns were? I think just large companies have concerns around launching products that can say anything.

13:37I would guess it's just like a matter of risk versus, you know, how much you're risking versus how much you have to gain from it. So figured, hey, startup seems like the right idea that you can kind of just move faster. Yeah. So tell us about character. What's the origin story there? Did you and Daniel look at each other one day and were just like, we have to get it out there? Yeah, pretty much. We're like, yeah. And kind of noticed, hey, there are people who just go out and get some investors and start doing something. So we're just like, okay, let's just build this thing and launch it as fast as we can.

14:12So I hired a total rockstar team of engineer researchers and got some compute. One thing that comes up a lot is people say that you all have one of the truly extraordinary teams in the AI world. Are there specific things that you recruited against? Or how did you actually go about finding these people? You know, some people we knew from Google happened to get introduced to Myelot, formerly from Meta, who had launched a lot of, well, built a lot of their large language model stuff and their neural language model infrastructure. And a bunch of other Meta people followed him, and they were great.

14:47Is there anything specific that you would look for in people or ways to test for it, or was it just standard interviewing approaches? I mean, a lot of it was just kind of motivation. I think Daniel tends to very, very highly value motivation. I think he's looking for something between burning desire and childhood dream. So there were a lot of great people that we did not hire because they didn't quite meet that far. But then we got a bunch of people who were kind of up for joining a startup and really talented and highly motivated. I mean, speaking of childhood dreams, do you want to describe the product a little bit?

15:22Like you have these bots, they can be user created, they can be character created, you can be public figures, fictional figures, anybody with like a corpus that you can make up or historic figures. How'd you even arrive there as the right form for this? Yeah, I mean, basically, this is kind of a technology that's so accessible that billions of people can just invent use cases. And it's so flexible that you really just want to put the user in control because often they know way better than you do what they want to use the thing for. And I guess we had kind of seen some of the assistants from sort of assistant bots from large companies.

16:01You've got Siri and Alexa and Google Assistant. And some of the problems there are that when you're just projecting one persona to the world, people will, A, expect you to be very consistent in, say, your likes and dislikes, and B, just not be offensive to anyone and not really have an opinion. It's kind of like you're the queen of England and you can't say something that's going to disappoint someone. Or I don't know, like I remember, like, I think it was George H.W. Bush said he didn't like broccoli. And then like the broccoli farmers were like all mad at him or something. So if you're like such a public, trying to present like one public persona that everyone likes, you're going to end up just being boring, essentially.

16:47And people just don't want boring. You know, people want to interact with something that feels human, you know. So basically, you need to go for multiple personas, you know, like let people invent personas as much as they want. And kind of I like the name character because it's got a few different meanings. You know, character like, you know, ASCII character, like unit of text, character like a persona or character like good morals. But anyway, so it's I think that's just how people like to relate to this stuff. It's OK. I kind of know what to expect from an experience if I can kind of define it as a person or a character.

17:23Maybe it's someone I know. Maybe it's just something I invent. But it kind of helps people like kind of use their imagination. So what do people want? Like, do they do their friends? Do they do fiction? Do they do entirely new things? Yeah, I mean, there's like a lot of role-playing. Role-playing games are big, you know, like text adventure, where it's just making it up. As it goes, there's a lot of like video game characters and anime. And there's, you know, some amount of people talking to public figures and influencers. And I think a lot of people have these existing parasocial relationships where they've got characters they're following on TV or internet or influencers or whatever.

18:00And so far, they just have not had the experience of, OK, now this character responds. Because it's always something you can watch. Or maybe you're in a thousand-on-one fan chat or something where this VTuber will write back to you once in an hour or something. But now they get the experience of, oh, I can just create a version of this privately and just talk to it. And it's pretty fun. We also see a lot of people using it because they're lonely or troubled and need someone to talk to. So many people just don't have someone to talk to. And a lot of use kind of crosses all of these boundaries. Like somebody will post, okay, this video game character is my new therapist or something.

18:39So it's a huge mix of fun and people who need a friend and connecting with game playing, all kinds of stuff. How do you think about emotion both ways, right? Like people's relationships with characters or like what level we are at in expressing coherent emotion and how important that is? Oh, yeah. I mean, probably you don't need that high-end level of intelligence to do emotion. I mean, emotion is great and is super important, but a dog probably does emotion pretty well, right? I mean, I don't have a dog, but I've heard that people will, like a dog is great for like emotional support and it's got pretty lousy linguistic capabilities.

19:20But the emotional use case is huge and people are using the stuff for all kinds of emotional support or relationships or whatever, which is just terrific. How do you think the behavior of the system will change as you kind of scale things up? Because I think the original model was trained on not a ton of money. Like on a relative basis, you folks were incredibly frugal. Yeah, I think we should be able to make it smarter in all kinds of ways, both algorithmically and scaling, you know, get more compute and train a bigger model and train it for longer. We should just get more brilliant and more knowledgeable and better attuned to what people want, what people are looking for.

19:57You have some users that are on the service like many hours a day. Like, how do you think about your target user over time and what the usage patterns you expect to be are? We're going to just leave that up to the user. Our aim has always been like, get something out there and let users decide what they think it's good for. And, you know, we see like somebody who's on the site today is active for about two hours on average today. That's people who send a message today, which is pretty wild. But it's a great metric that people are finding some sort of value in it. And as I said, it's really hard to pin down exactly what that value is, because it's really like a big mix of things.

20:35But our goal is like make this thing more useful to people and let people kind of customize it and decide what they want to use it for. If it's brainstorming or help or information or fun or like emotional support, let's get it into users' hands and see what happens. How do you think about commercialization? We're just going to lose money on every user and make it up in volume. Oh, good. That's a good strategy. No, I'm joking. The traditional 1990s business model. It's kind of a 2022 business model, too. You should issue a token and just make it a crypto thing. No, we're going to monetize at some point pretty soon because, again, this is the kind of thing that benefits from having a lot of compute and, you know, rather than burn investor money.

21:22The most scalable way to fund something is actually provide a lot of value to a huge number of people. So probably try some premium subscription type of service where, you know, we can, as we develop some new capabilities that might be a little more expensive to serve, then start charging for them. I really like that anyone can use it now for free because it's, you know, there's so many people that it's providing so much value. I mean, it's really taken off as a consumer service in a really striking way if you look at the numbers of users and the number of hours of usage per user, which is insane.

21:54And are there any scenarios where you think it's likely to go down like a commercial setting where you have like customer service bots who provide like a brand identity around support? Or is that just not that interesting right now as a direction? I mean, right now we have 22 employees, so we need to prioritize. We are hiring definitely enough work for way, way more people. Priority number one is just get it available to the general public. It would be fun to launch in this customer service spots when people would just stay on customer service all day. They're chatting with a friend effectively.

22:30Yeah. Start with the customer support. And that actually happened apparently on some old e-commerce sites. eBay apparently was effectively a social network really early on as people were buying and selling things and just kind of hanging out because there weren't that many places to hang out online. So I always think it's kind of interesting to see these emergent social behaviors on different types of almost like commercial products or sites. But it makes a lot of sense. So you said one of the obvious reasons Lambda didn't ship immediately at Google was safety. Like, how do you guys think about that?

22:57Like, remember, everything characters say is made up. Exactly. Right. Make sure the users are aware that this is fiction. There's anything factual that you're trying to extract from it. At this point, it's best to go look it up somewhere that you find reliable. You know, I mean, there are other things, types of filters we've got there. We don't want to encourage people to hurt themselves or hurt other people or we're blocking porn. There's been a bit of protest around that. And do you view all this as a path to AGI or sort of superintelligence? Sure. Yeah. And is that part of the goal? For some companies, it seems like it's part of the goal.

23:35And for some companies, it seems like it's either not explicitly an anti-goal or if it happens, it happens. And the thing people are trying to build is just something useful for people. What a flex. AGI is a side effect. Yeah. Yeah, well, I mean, that was a lot of the motivations here. Because, like, I mean, my main motivation for working on AI, other than that it's fun, well, I mean, fun is secondary. Like, the real thing is, like, I want to drive technology forward. There are just so many technological problems in the world that could be solved. For example, like all of medicine, like, there are all these people who die from all kinds of things that we could come up with technological solutions for.