AI & Identity, from East & West, with Yohei Nakajima GP at Untapped Capital and BabyAGI Creator

I think Western identity tends to think of like, my thoughts are who I am. And so my mind and my thoughts are, are my identity. I do feel like Eastern cultures tend to treat the mind and the brain as more of a tool to control, and your thoughts as an output of that tool. And so like meditation and enlightenment is really about harnessing how to, how to leverage that tool. As I get emails, as I get texts, as I, you know, as I follow people on Twitter, as I read websites, if I could just convert all that into a structured knowledge graph of The people and entities and events, it should be easier to like identify potential people to reach out to, to connect with, introductions to make amongst people.

Right.

I, I've seen a couple of early tools of like helping people understand each other that previously wouldn't, or helping find commonalities between 2 groups of people that, that might not be intuitive. And I think if we can start to think about AI as, as a tool for that, new ideas will emerge that can hopefully create a future that's not, not as bleak as, as, right, some, some people see the future.

Hello and welcome to The Cognitive Revolution, where we interview visionary researchers, entrepreneurs, and builders working on the frontier of artificial intelligence. Each week we'll explore their revolutionary ideas, and together we'll build a picture of how AI technology will transform work, life, and society in the coming years. I'm Nathan Labenz, joined by my co-host, Erik Thornberg. Hello and welcome back to The Cognitive Revolution. Today we've got a special bonus episode with returning guest Yohei Nakajima, general partner at Untapped Capital and creator of the pioneering Baby AGI project. Our first episode with Yohei came out in October, right after he had spoken at the TED AI event, but before the talk was available to watch online. We cover Yohei's many experimental open source AI projects, as well as his investment philosophy in that episode. Today, with the talk live on the TED website, we're back to watch it together, and to explore the key themes more deeply along the way. In the first half, I could not resist digging into more of the details of how Yohei is building today, now that OpenAI has launched the GPT Store and the Assistants API. In the second half, we unpack his reflections on identity, and discuss how different cultural contexts seem to be shaping individuals and societies Responses to, and relationships with AI technology. This sort of reaction video is a new format for me, but I think it came out extremely well. And I want to thank you, the listeners, for making it possible. Johei is one of a few guests who have recently told me that they've heard far more about their Cognitive Revolution episode than they'd heard about any other podcast that they'd done. And so it was an honor to spend an hour with Johei on the morning of his birthday, for a conversation that touches on some of the smallest technical minutiae and also some of the biggest philosophical questions in AI today. Please do continue to share the show with your friends, and I hope you enjoy this special part 2 with Yohei Nakajima. Yohei Nakajima, welcome back to The Cognitive Revolution.

Thanks for having me.

So as we kind of teased in the last episode, you have recently put out a TED Talk, and now it is Available to the public. And just as a kind of fun experimental idea and a way to get a little deeper into your philosophy as it pertains to AI as well, we're going to just roll the TED Talk and then pause it as we go and kind of discuss and allow you to expand on some of the ideas that you only had, I guess, 9 minutes to deliver live, but we've got the full hour today. So excited to play it and we'll just kind of pause and discuss as we go.

Let's just dive in and take it from there.

In my office, I have a stack of index cards that I call my You Are Who You've Met book. On one side of each card is a name of somebody from my past, and on the other is one lesson I learned from them. It includes friends, teachers, enemies, real people, and fictional characters.

Okay, first question, enemies?

Yeah, I mean, you learn as much from your enemies as you do your friends, right? And when I say enemies, right, I'm thinking like high school rivals and, and whatnot. That I eventually became friends with. But I do think you learn just as much from the people that you battle with as, as, as those who are standing by you. And, and fictional characters too, right? Like, you learn a lot from— I think where we are heavily impacted by, like, the stories we watch growing up.

Interesting. Well, this will, this will make more sense as we go, so I'll keep playing.

It was a thought exercise I started in college and kept up. So on behalf of everyone I've ever met, it's nice to meet you. My name is Yohei. I'm a venture capitalist by day, builder late at night. I'm a dad of 3 kids, and I'm obsessed with my wife. Within the AI community though, I'm best known as the father of Baby AGI. For those of you not familiar, Baby AGI was the first popular open-source autonomous agent with task planning capability. You could give it an objective and it would generate its own task list executing them one by one, adding new tasks based on previous task results, and continuing until you stopped it, all in 105 lines of code. Probably due to its simplicity, it inspired developers from all over the world to start building their own autonomous agents, and that's what's been most amazing about this. It's hurled me into the center of this incredible AI community, and I couldn't be more grateful. It's what brought me here today. But as I thought about what to talk about, I kept coming back to this idea of identity. See, Baby AGI was a weekend project that unexpectedly went viral. The quick backstory is I actually challenged myself to build an autonomous startup founder, and when I shared a video online, people went wild asking if it could do more, which it could. My friend Jenny commented, bro, did you just build Baby AGI? Which is where the name came from. Relevantly, the development of BabyAGI itself has been weirdly introspective. I'm trying to get it to do all my work, so a lot of the ideation is watching it do things, thinking about how I do it better, and trying to close the gap.

I want to ask a couple questions about that. I'd love to hear a little bit more about that cycle. I think that's something that all manner of AI developers and AI engineers are kind of going through. I imagine you have some interesting best practices. And I'm also interested in kind of your latest feelings on just how far we can go with that paradigm, even on kind of today's models. And as I'm sure you're aware, quite a few notable papers put out recently using increasingly sophisticated prompting techniques, and the state of the art kind of continues to advance. But maybe for starters, take us a little bit more kind of behind the scenes of literally how do you do that work procedurally?

When I'm thinking about like building a next mod of BabyAG or any AI tool for that matter, right? After I built like a first prototype, I will run it over and over again with multiple different prompts and examples, and I'll watch it do things. And I just watch it over and over and over again until something like clicks for me when I'm like, wait, here's something I do that the, that this isn't doing. Right. As an example was like self-reflection. When I added that, I was noticing that if I ran the same objective multiple times, it was not learning from previous objectives. So I thought, how do I do that? Well, after I run— after I try something and I go through that process and it doesn't work well, I'll reflect on it and say, why didn't it go well? And so that's a step I literally just asked a large language model to reflect on it. But then I need to use that somehow the next time I'm running a similar objective. So I was like, okay, let's, let's do a reflection on the task output after an objective, the final task output after an objective, and then let's store it. And then I guess when I'm going to do an objective, I only really need that if I'm running an objective that's similar to it. So let's store all of these and embed them so that whenever I'm running the next objective, I'll do a similarity search and find the 5 most relevant reflection notes and then use that to guide the task list creation.

And that's—

I'm just like thinking out loud exactly like how I went from like, okay, let's, let's add this. And that's just one example, right? I watch it over and over again and I notice that, hey, look, if I run the same objective multiple times, it does get better at building better task lists.

Do you have any tools that you use to do this? Are you literally just printing stuff out and scrolling logs and watching it roll by?

I should probably have a tool to do it. No, I am literally just watching it do it. I guess if I keep using the tool over and over again, I'm just waiting for something to click and inspire me to want to add to it.

I'm a big believer in the importance of exactly this, just reading raw logs, whether it's in real time or just Establishing the discipline of going back to the raw outputs and also the raw inputs, especially depending on the context. You may not have full command of what those raw inputs are depending on what you're building, but just reading the raw logs, I think, is super important for just about every AI development use case. The surface area is so broad, the surprises are so weird, the strengths and weaknesses are so unintuitive in some cases, that I think that is really important. You mostly do this stuff with GPT-4, right? Have you seen a lot of difference in how things behave depending on which model you use? Like, have you tried, for example, different versions of GPT-4 or Claude 2 or Llama 2? There's kind of multiple dimensions of this. One is just like the capabilities question of how good is it? There's also a little bit of how is it RLHF'd and how different do you feel like the character of the leading models is today, if this is something that you've spent enough time with to have a sense for?

Honest answer is I don't think I do. As a matter of fact, I actually use 3.5 more than 4 for most of my auto agents. The way I see it is if I can get it working with 3.5, which is where I started, I can always upgrade to 4 if I want a better answer. But if I can get it working with 3.5, the flow, the operations, the logic, the JSON outputs, all that, I think of 4 as the super mode, the super saiyan mode of my auto agents. But practically it's cheaper and faster, which is probably why I use 3.5.

Certainly when you're developing the scaffolding and you want it to just run fast so you're not waiting for long generations. The GPT-4 generations often are north of 30 seconds depending on what you're trying to do. I know that threshold well because that's the threshold for Google Sheets and I've had a hard time in many cases, getting GPT-4 to finish in time for the Google Sheets custom functions to not just error out at the 30-second limit. I've been thinking about this kind of from the other angle recently of there's so much effort, so much investment going into scaffolding, and we're going to have these kind of systems already built, right, and developed to a significant extent, and then there's going to be a model upgrade. And so I'm kind of like, man, a lot of these things that don't work today, may really start to click into place, and perhaps rather suddenly when a new model comes online. The flip side of that would be if you are building this with a model that is less powerful than the most powerful model that you have access to, maybe you're kind of unnecessarily limiting, not yourself, but maybe you end up building more scaffolding than you need to. For example, have you had any experiences where you felt like, Actually, I didn't have to build any of that kind of stuff because if I did wanna upgrade to GPT-4, then it would just like not get stuck in those places or whatever.

I think I've avoided making my scaffolding too complex. I think I, I mean, I just love simple code and so like my, I'm constantly just trying to like simplify everything. That being said, I've probably done a little bit in the prompting where the prompting for to get the right like task list output of a 3.5 is probably more work than 4. So I probably did put a little bit of extra work into the prompting to make it work with the lower model. But I don't think it was necessarily wasted because I do think that better prompt also just makes the better models work better. But to some extent, there's probably some wasted energy there. And that is something I'm conscious with everything I build is like, how much time am I going to put into a problem that I might not need to solve eventually? And I think a good example is reading pitch slides, for example, as soon as GPT-4B was kind of you know, was coming, I dropped all efforts to look at OCR and like, you know, figuring out how to OCR PDFs. I was like, I'll just wait till GPT-4V comes out and I can just give it the whole slide. And so I do a lot of that where I'm just trying to figure out what's going to be in the next models and like try to build stuff that might take advantage of it. And so when GPT-4V came out, I had like 3 or 4 prototypes of code where all I needed to do was swap out the API because I was building towards what I thought would be in the models next.

Do you have a wishlist or a sort of taxonomy in your mind of things that you are looking for, for the next model along the lines of a vision capability where you're like, yeah, I definitely don't need to build this because it's going to happen? And maybe not even with that much confidence, but are there things that you're kind of like discrete things you could describe that you could say to the model developers, like, if it could do this, it would be a huge unlock for the kinds of things I'm trying to build?

I'm unsure. I try not to speculate too much, but I'm very curious about, at a high level, about what part of the orchestration can get sucked into the model itself, I guess, to some extent, right? If you look at like even GPT-4V, it's not just looking at the picture, but it is obviously reading the words, which means there must be some OCR, right? There's probably a couple things. And then you see Mixture of Experts coming out, um, which is almost right. If you think of like how Mixture of Experts work, you know, before I knew about it, I would have thought that would have been an orchestration problem where I can, you know, use an LLM to— and give it a couple of expert descriptions and say, I'm just Oversimplifying, but then having the model route it. But now that's baked into the model. So I'm always curious as to what parts of what we see in the orchestration eventually get baked into the model. And I don't know the answer to it, but I'm very curious. It's one way I'm thinking about it. Tool usage, retrieval, those are orchestration problems today, but do those eventually become model problems? Are those eventually baked into the models? I don't know the answer.

We're seeing, I think, Some definite suggestions that retrieval could go that way. Perplexity's online models, I don't know exactly how they work, but if I know them, I think they're pushing that direction. They're embracing the bitter lesson and trying to train more end to end. And I would maybe look at the Google paper Retro, which I think is maybe 18 months old now, making it almost ancient in current AI discussions, but still a very interesting setup there where the retrieval was part of the core system. It wasn't like a separate generate keywords and go search, or even kind of a very distinct external embedding thing that would return the document and then feed the document into context, but rather information entered into the language model in the middle layers in that So it's entering in high-dimensional space and well above the prompting. Everything's already been worked up at that point. Hey, we'll continue our interview in a moment after a word from our sponsors.

And then tools too are still, from what I can tell, it's still not part of the model, but from cloud to OpenAI, they're at least embedding tool usage as part of their API. And from a developer standpoint, to some extent, it doesn't really matter if it's in the model or the API itself. If I'm using an API and it has tools, it has tools. So that was an example of trying to build. I've been playing around with a baby AGI concept that's using the OpenAI Assistants API, which has tools, function calls built into it. So that's been an interesting experiment. It definitely decreases the cost of building, although similar to letting a model do the retrieval, you lose a little bit of control there. So it's an interesting trade-off.

What has your experience been with the Assistants API? Mine, for what it's worth, has been that the— and I've done it more on the GPT side versus the Assistants API side, although I basically understand those to be the same core infrastructure. I have found their retrieval to be not very effective, honestly, for at least the use cases that I've tried. Have you had better luck or do you have any tips or tricks?

I've enjoyed using both the Assistants API and GPT, but for neither, I haven't used the baked-in retrieval for either.

In my experience, it's been just not enough context loaded in.

I use web browsing a lot in the GPTs. A lot of the stuff I would want to retrieve is online, I guess. A lot of my work is I'm doing web research all the time. I did find that with GPTs or with assistants, you can actually just tell it to do a web browse with a site-specific search, for example. It's kind of like you built a retrieval over a website and it works well enough. That being said, I don't know if I've Yeah, I haven't really played with the actual retrieval where you upload a PDF and it does the chunking for you.

Listeners will know I've been on a bit of a knowledge graph kick recently, and the experience with the GPTs was kind of the reason for that. I've uploaded a code repo and I don't know, you have this kind of opacity problem of you don't really know how it works and you kind of have to try to fill in the gaps in your knowledge of how it works based on how it behaves. But it seems like they're chunking things pretty small. And then at least for the things I was trying to do, when you match on a small chunk and you load that small chunk into context, it just wasn't enough context. So that got me thinking about knowledge graphs of, okay, that localised search, I'm sure there's a good reason that they're chunking small to search small, but I kind of need the whole Python file from which it came in order to Because then I've run into a lot of these problems that people have been, that have probably multiple origins of like, you can implement this or whatever. And it's like, I have implemented this, but I need you to find it in the thing to load it in. So I think there remains definitely some opportunity for improvement in that system. And it is funny how little knowledge we have about just how it works.

Well, you know that I've been on the knowledge graph kick too, probably from like September or so with Instagraph and whatnot. But I have a I have a prototype, like autonomous CRM that I've been testing, which I'm testing it with Game of Thrones episode descriptions from Wikipedia. And so I'll like upload one episode description at a time. And the idea is that it's like on the backend building a knowledge graph of all the information.

This is something I actually want to do more custom development on, because I haven't seen anything that has really killed it for us, to be honest. What are you kind of storing that in a traditional knowledge graph database?

I mean, it's whatever ChatGPT suggested I use, honestly. But yeah, it was, it started with Instagraph, which was just like converting any text input into into a knowledge graph and then converted that so that I'm deduping nodes as I add new information. So I do like a search against existing nodes and then send like the 5 or 10 most relevant sounding nodes with the new node to an LLM to quickly assess whether the node exists already or it doesn't. And then if it exists, it'll map the ID to it. If it doesn't exist, it'll create a new ID. And then it does the same thing for edges. So if I just keep adding more and more kind of episode descriptions, the first one might find the relationship between Sansa and Ned Stark, but the next one will find the relationship between Sansa and someone else, and then it'll continue to kind of build on it. It's working okay. It's still, it's, I still need to prompt it better cuz sometimes Ned Eddard Stark and Eddard Stark end up as different nodes, which, you know, again, small things like that just need to be fixed. But the reason I'm working on it is like, I just, I just want this knowledge for myself, right? As I get emails, as I get texts, as I, you know, as I follow people on Twitter, as I read websites, if I could just convert all of that into a structured knowledge graph of the people and entities and events, it should be easier to identify potential people to reach out to, to connect with, introductions to make amongst people.

One of the big challenges I see with these kinds of systems is that there is a time element to obviously life, right? And just kind of with something like Game of Thrones, you could sort of treat that as like a crystallized thing and build out a graph and it wouldn't have to change. But As we move through our lives, it's more like we're going through episode by episode, season by season. And you would perhaps want a history of like, this person is aligned with this person and then they're enemies. And there's kind of a— it's not just one fixed description. Have you any kind of experiments or techniques that I might be able to borrow from on that time dimension?

I haven't added it to my knowledge graph, but when I'm doing retrievals on self-improvement, that I was mentioning earlier, I added a timestamp to it. And then when I'm doing a retrieval, I actually use the timestamp as a decay— I do a decay function. I forgot what the exact decay function is, but like the older an item is, it like multiplies it by, you know, by something less than 1. So like the similarity score slowly drops over time, um, to an embedding. So that was an interesting way to do a retrieval that took time into account. I haven't really applied it to knowledge graphs, but I guess similarly, you can imagine a timestamp being on nodes and edges and them slowly getting faded over time to some extent. Or if there's a new edge that overrides an existing edge, there might be some logic that you need to figure out there.

I love how you're scouting out the frontier of these capabilities constantly. Happy to be riding in your wake on some of this stuff. The talk's gonna move from technical to largely philosophical. Any other technical comments before we move on to the...

No, I think that's a good start. I mean, it was interesting coming up with a talk concept, right? Obviously, I was invited because of Baby AGI, but Baby AGI was such a recent thing. Actually, the TED experience was great. They hooked us up, connected us with a great speaker who helped with everything from ideation all the way to the delivery. And really, it was the ideation questions that she gave me at the beginning that really guided the topic of the TED Talk. Uh, which I felt like was a good one that tied Baby AGI, you know, and my other work, you know, with as a VC and my kind of personal history together. So.

All right, cool. Let's roll it.

I often joke about replacing myself at work with AI someday, but I'm pretty sure I'm not joking. I have an experimental chatbot called Mini Ohay that startup founders can talk to, and it sends me summaries of its conversations. Is, is Mini Ohay an extension of who I am? These seem like conversations worth having. So let's talk about identity and we shall start at the beginning.

I talk about Mini Ohay chatbots. This month I released a GPT VC associate chatbot. I just thought it was relevant because it's fresh and it's similar to what I just mentioned. But it's, I think it's had like 7,000 chats, which is extremely wild if you think about like how many, you know, conversations I have with founders each year. So it's pretty fascinating what, what AI can do in terms of like just scaling somebody's impact.

I've built the GPTs as we discussed, but I've not actually shared GPTs publicly. When you do share them, do you get anything beyond those headline numbers of just raw usage? Can you see anything about how people are using it or what sorts of questions they're asking?

No, and that's definitely a loss from a builder perspective. I encourage people to send it to me. So my GPT, the VC associate, will give founders a downloadable investment memo. And so a good dozen of those have ended up either in my inbox or DMs. There is a way to store it. And so I'm kind of working on that. But as of now, I'm not storing any of the information. So it's just truly a founder-facing tool. But the benefits are distribution, right? I did the VC Associate because I knew the marketplace was coming out. And as a result, it was listed on like the top 6 within research. So I think that drew a lot of people I didn't have reach to. And then 2 is the cost, right? I mean, using GPT-4 for a chat, for thousands of people across 2 weeks can quickly add up. But I didn't pay a dime because they're all paying OpenAI $20 a month. So it's a great MVP.

Are you specifying what model it uses? And then is that something that only paid users could come in and use? What if somebody's not paid, but you had it set on GPT-4? Do they just get the 3.5 version?

So as of now, I don't think you can even set it. I think my, I mean, and if you can, I haven't found it. So I'm just using GPT-4. All of mine use GPT-4. So people who are not paid users cannot use the custom GPTs. It's a very interesting dynamic because what may have been a startup previously with like a signup page where they might actually get email addresses, OpenAI is capturing some of that market by opening this GPT store and converting those potential users of other startups into just OpenAI subscribers.

So that's pretty significant usage that you've had. I mean, 7,000 obviously is not a massive number in the grand scheme of the world, but if you figure all of those people had to be subscribers, and I don't think we know what the paid subscriber count is for ChatGPT, but if it were a million, you would be approaching 1% of paid users, which is not small. I mean, that's really like for these little— I mean, there's so many of these platforms over time. have launched their app store and very few of them really work. Would you say this experience leads you to be bullish on the GPT store overall?

Not really. I'm definitely approaching it with a cautious, this might go away. For all I know, OpenAI might just shut it down like they did plugins. I'm kind of approaching it in a way that I would be okay with it. I think it's a tough place to build a business, but regardless, it's a great experience to See how people are using it, see how people are engaging it. Interesting to notice, like, cuz I've had a similar tool that was a standalone website that was free that people could chat with, that Mini Ohay that I've pushed on Twitter versus, you know, the one that's on GPT is getting just significant more usage. So there's definitely, you know, it's interesting to compare both approaches having done both. On one hand, I mean, I think from OpenAI standpoint, I, I, I, and I can't confirm that this is what they're doing, but it seems like a great way to collect insight into how people are using The different models, how people are prompting it. And if they can take advantage of all the data they're collecting from the GPT store into future models, it puts them in a strong position.

When you said that you can send information to yourself, does that amount to setting up a tool so that you would basically add a function and the function would be like, this is to send the results of this chat back to the developer?

So the version 1 that's out right now, it's just using Code Interpreter to store the investment memo as a text file locally. And it just keeps adding to it and then just gives a downloadable link. The next version I'm working on does the exact same thing, except instead of using Code Interpreter to store it into a text file, it's using a custom Airtable tool to store the appropriate answers into the specific Airtable columns. So as people chat with it, it's getting the same information, but yeah, instead of a text file being loaded, it's storing into a specific Airtable so that I can give an Airtable— give a link using a no-code tool that shows their report, but then it's sitting in Airtable on my end. And then also I have a separate tool called Deal Flow Digest that matches founders and investors, which is all in Airtable too. So once I get that working, I think in theory, I could get the GPT VC associate to offer to send the memo to other investors, and then just start making intros automatically.

Hey, we'll continue our interview in a moment after a word from our sponsors. When you call a function or a tool from inside of GPT, like to say write something to Airtable, it still has to, if I understand correctly, what it's going to send, what it's going to write, it still has to generate that on a token-by-token basis, right? Is there any— there's nothing yet where it's like chat.history where it could sort of reference the current context as a variable so as to not have to copy all the earlier tokens using the same language model mechanism. I haven't seen anything like that, but it seems like that would be a very natural thing for them to start to add onto this too, right?

Yeah, that is interesting.

Yeah.

If you could to some extent figure out a way to embed variables into the API, into the function API calls.

I've seen this in Code Interpreter in my own work where it's like every time it wants to regenerate something or reference something, it still is in that generation mode. And I kind of want the code interpreter to be like, whatever, airtable.send, chat.messages. And if it could just generate those 8 tokens but send the whole context, that would be powerful. Otherwise, you're kind of duplicating everything in context, plus waiting a lot longer, burning CPUs or GPUs, I should say.

That's a good question. I mean, I guess to some extent you almost have to treat The reference as a token itself.

Yeah. And that starts to get a little more native too, right? In terms of when we were talking earlier about internalising tools, it's maybe not a critical distinction, but you imagine a future in which there's these shorthand tokens that it can generate to kind of refer to its own context or its own sort of environment, whereas today it doesn't This feels like one of those things where I would think about how it would be done in orchestration a while and then assume that the model will do it and just not put effort into doing it on the orchestration side. All right, let's go back.

In the beginning of all life, there were single-cell organisms which became multiple-cell organisms and the cells started to specialize. Some cells were better at taking light and converting it to energy. Some were better at storing it and some were better start using it to propel. And as these organisms became more complex, they self-organized into complex intertwined groups of cells and subgroups within them. This pattern is common. Baby AGI started as 100 lines of code, 200, 300. It became multiple files, then it became multiple folders with multiple files within each. Startups and organizations as well start with 1 or 2 people doing everything. They bring on an HR person, they bring on a marketing which becomes an HR department, a marketing department, and then smaller departments and groups within each. See, in the context of an organization, the concept of identity is a bit more clear. It's our role within that organization which hopefully is to some extent defined, or at least the expectations are. Our personal identities though are a bit more complex, as we're part of more than just one organization, but many overlapping groups of people, a couple, a family, a local community, a country, mankind. And within each of these groups, we have a unique role, which is to some extent defined by our words and actions, or more specifically, the impact of those words and actions on said group of people and the people within them. And bringing it back to the top, we're not just the impact we have on others, but the impact others have on us. We're ultimately a conduit of mass energy, and ideas that we amplify or suppress with our everyday choices. And each of these choices impacts all the different groups we're part of, and that impact in turn defines the many roles that make up who we are. We are complicated.

You were asking about my philosophy, I think right before we started, of whether there's some Eastern philosophy. I definitely think there is, right? I think the goal here was to really implant the idea that like, that we're not individuals, right? I think the concept of identity in today's world, a lot of people think about like, who am I? make it really like, you know, internal versus I wanted to get people to start thinking about identity as, you know, your place in a, in a larger system, I guess.

Yeah, there's a lot I think to unpack there. Maybe for starters, just give us a little bit more about your background. I mean, this is a, you know, just the, the habits of mind with which we grow up, I think certainly inform the way we approach some of these big questions. So I'd be interested to hear a little bit more about kind of your personal biography and then I have a lot of questions which I don't think, you know, we can fully answer. I wouldn't expect any individual to fully answer, but about just how different societies are going to relate to this technology.

No, I think that's a great point. So yeah, I was born in Japan, but I grew up in Seattle, but in a Japanese household, speaking Japanese at home and going to Japanese school on Saturday. So when I was, when I was a kid, I had my like American school persona and then my Japanese school persona, which was different because I didn't think about it at that point, but there were different groups of friends with no overlap and speaking completely different language. It was only after I grew up that I started to really appreciate how much philosophy is embedded into language, specifically Japanese language too, because you're using Chinese characters where the characters themselves have meanings. And then sometimes you'll combine characters with different meanings to talk about something completely else. A good example is A common way to greet someone is genki desu ka? Like, are you genki? Is how Japanese people greet someone. But the characters for genki are gen is root and ki is energy. So it's like, how is your root energy? Is like the most casual way in Japanese to ask someone how they're doing. And I think that's a very philosophical way to ask somebody how they're doing. But I think that's just one example of Japanese philosophy embedded into the language itself and that you're just using on a day-to-day basis. After growing up in the U.S., I went to Japan for high school, which is when I found out that I actually wasn't Japanese. I thought I was until I went to a country full of Japanese people and realized I was the most American person there. And so it was a pretty rough initial landing in a Japanese high school, but over the course of 3 years, I acclimated and I think there was a lot of kind of personal growth there. And then kind of moving back to California for liberal arts school, I got to kind of become re-American to some extent and get to be that part of myself, but I definitely carried a lot of What came from Japan and Seattle and LA are also very different culturally as well.

So do you think that background has an intuitive impact that may ultimately become consequential in terms of how people understand language models on first contact? For me, and I think for most who have just kind of the traditional Western background, it is natural to think of ourselves primarily as individuals and then this thing is like this kind of weird language model, and I've got to tell it what role to play, tell it what individual to be. Right. That's like one of my prompting best practices is you are X. But I don't spend a tonne of time thinking about it as sort of a representation of a collective. And I wonder if when we had Balaji on, he made a really interesting point about how there are these Hindu gods that have a bazillion faces. And so it's The construct of something that is like multiple within one and can kind of play all these different roles is maybe more intuitive. I wonder if there's a similar kind of difference in how this stuff feels to people who have this cultural background. And also wonder how that may relate to some policy decisions. Right. Like, we know that— I don't think this is necessarily baked at this point, but from what I understand, Japan is taking a very permissive approach to all the data is going to be allowed to be trained on. And that in some sense also is kind of a more collective-first decision, right? You don't have rights in Japan, or at least it seems like it's shaping up where you don't necessarily have rights to assert that your data, your creations can be excluded from this collective synthesis project. Rather, the collective has the right to to subsume all this stuff. I'm out of my depth here, but I'm fascinated by it. What do you think?

I think you're touching on a couple of interesting things. I think one, language-wise, there's so many geography-related things. First of all, most language models are strong at English and not in a lot of other languages. So then there's the whole, what's the gap that it creates if there's countries that have AI and countries that don't? But on the more, I think what you're talking about, the permissiveness is kind of how we view— I think the brain, and I can't speak for all Japanese people obviously, but I do think that Eastern philosophies tend to see the mind more as a tool versus a self. I think Western identity tends to think of like, my thoughts are who I am. And so my mind and my thoughts are my identity. I do feel like Eastern cultures tend to treat the mind and the brain as more of a tool to control, and your thoughts as an output of that tool. And so like meditation and enlightenment is really about harnessing how to leverage that tool.

Yeah.

And again, I'm not speaking for all people, but when I think of artificial intelligence, it's just another tool, much like real intelligence is a tool. Artificial intelligence is also a tool. So I think even coming back to, can AI be creative to some extent? It's whether it's my brain or an artificial brain that's creating it. To me, it's a tool that's creating both things. But if your philosophy, if you think of your mind and your brain as your identity, and its ability to create something as unique to you, then I can see having a machine do it be kind of threatening, if that makes sense. But I already saw, I guess to some extent, I already see my brain as just a meat machine. So the gap is not as big philosophically.

Interesting. Somebody asked the other day on Twitter, if you could go back and tell your much younger self just one high-level fact to guide your future from that point, what would it be? What I came up with was next token prediction is more powerful than you can possibly imagine. And I think you are not your thoughts is another one that I would strongly consider giving to my earlier self. I do feel like that is a major realization for many people that is not kind of something we grow up with by default here in, I'll say in Detroit, Michigan, where I grew up. So do you think that also has a— I mean, your last comment there around things being threatening, I wonder to what degree you think that the intuition— I have another show I'm trying to put together on AI in China, and it's like, good God, what a topic, right? I mean, I can't possibly even really begin to wrap my arms around it. But the stylized fact, right, or the sort of received wisdom is that people in China, maybe people in Asia more broadly, are like, less fearful of AI. First of all, I wonder if you think that's even true. Possible recent counter evidence where the premier of China said there should be a red line in AI that we don't cross that kind of bucks that narrative a little bit. So I don't want to take that narrative for granted, but I wonder if that feels intuitively true to you and if it feels connected to this identity, collective versus individual question as well.

It does to me. I think Japan is much more of a collective Japan as like the identity, right? You see people following kind of social rules and people think of themselves very much as as part of a country that works together. And I think when you identify as a collective group of people, then your output is the output of the collective versus so so AI just AI becomes like when you think of groups of people, the tools they use become part of that identity versus if you think of yourself as an identity, all tools are kind of external to you almost by default, although you can kind of change that perceptually, everything outside of your physical body is no longer part of you. Versus if you think of yourself as part of a collective, then all the tools that everybody in my group has access to is part of something that I benefit from. I think you instinctively think about tools.

Definitely fascinating.

I talk about ants a little bit later, but I'll just touch on it because it's relevant here. I mentioned the studies on how if you observe an anthill or bee colony, it acts like a neural Like a neural network where each ant is a neuron and their interactions collectively represent an intelligence. What I didn't get into the talk is that a lot of these studies, when they're looking at the interaction of ants, they're also looking at the trails ants leave as part of the collective intelligence of communication. So if you look at a single ant, you would not consider the trail they leave as part of that ant. But when you think of the ant colony as an identity, then the trails they leave, the tools they use to communicate with each other, become part of— That collective intelligence. So if you look at me individually, my computer is not part of my intelligence. But if you look at human society as a collective intelligence, then our computers, the technologies, the internet all become part of that identity of that collective intelligence. And so you could just as well make the argument that when we build AI, it's making us stronger versus it's something external to us inherently.

I also wanted to ask about just the experience of using language models in Japanese. I don't know how much you've done that, but I've kind of only seen aggregate statistics that show, as you'd expect, strongest in English. You can look at the power rankings, and it's, I think, pretty intuitive that it's like the most common and similar languages to English tend to be kind of next. And then as you get toward more distant and certainly more rare, fewer speakers languages, then you see weaker and weaker performance. Is there anything you could kind of describe the experience of using the OpenAI models in Japanese? How much different is it? How much worse is it, perhaps? And what do you think that means in terms of the future of more localized models coming out of Japan specifically?

The first time I tried Japanese, I was pleasantly surprised that it did well. I don't use Japanese too often myself, so I can't say I've tested it at length, but I did do some testing with BabyAGI. One of the challenges I did find initially when I was testing with BabyAGI was that I'm wrapping prompts and all those wrappers were in English. So even though I was trying to get it to do Japanese, it would sometimes switch to English partway through, which was a little frustrating from a testing perspective. But I guess to some extent that is a challenge. If you think about building or doing orchestration on top of language models, how do you make the orchestration multilingual is an interesting challenge. I don't think I tested it enough to notice that like Japanese is weaker than English. I think the kind of areas it hallucinates, you, you know, were, and again, I can't think of anything specific, but whenever you're thinking about a language model, I'm, I'm thinking about like what's been written about a lot. Those are things it's probably good at. What are things that haven't been written a lot? So I probably tested a couple, couple, a couple of each and kind of had the expected results where if I asked something that's written about a lot, it would be pretty good. If I asked about something that wasn't written about a lot, it might be a little more finicky.

I mean, I know there is a certain emphasis and there's some interesting new startups being built in Japan. When you see countries around the world making their own kind of AI investments and saying, we want to have our own— I don't know if we would go as far as like a national champion or whatever to be the sort of Japanese large language model company, for example. Do you think of that as like a good idea, something that seems wise for them, something that is of practical importance to their users, something that is of national strategic importance so they can't be cut off from the technology or just have the development capability. I don't really have a theory of this, but I wonder how you understand those kind of almost like state— they're like near state-level investments, right? Certainly at the level of policy.

I mean, I can't speak to the individual decisions that are being made, of course, at these companies, But I mean, if we're using large language models to automate a lot of things and really take advantage of it, I mean, especially I think when it comes to policymaking, politicians or government understanding the thoughts of people, taking all the comments and calls that people are getting and summarizing and really understanding what's going on, language models are powerful. But if you want to start relying on it, you don't want another country to have the ability to shut you off against it. So it does become a national security issue to some extent, I think. Yeah, I think it does make sense if you have the talent internally to start playing around. I'm assuming Russia, right? I haven't seen a big funding announcement, but Russia, I'm sure, has stuff. India, right? I mean, all the different languages in India, I would suspect there'd be a model company coming out of India. Again, I don't know if they need a $100 million valuation, but at the same time, if you're going to get your country to start relying on a model, do you really want it to be outsourced to a company outside of your country? It's a good question.

Is tokenization a huge pain point in Japanese? I've seen this explored a little bit in some Indian languages where it's not just a tax, but it's like a massive difference in cost structure because one character in an Indian alphabet might be like 8 characters of Unicode or whatever. And that leads to just straight up order of magnitude more costly inference, even if performance is the same, which it's also probably not. But I assume that's got to be kind of an issue in Japanese too.

Definitely not as efficient from what I can tell. Again, this is a little bit outside of my strong suit, but I think to some extent, our optimization has been largely built around English itself. So maybe there are better approaches, like even slightly altered approaches that might perform better on a language set that has thousands of characters versus 26 characters.

Jan Leike, the head of alignment research at OpenAI, once tweeted about their Instruct model that they only trained it in English instructions and then found that it generalised to follow instructions in other languages as well. And he was like, we still don't understand that. And that was at the time that they launched it. So that's wild. It's wild that it happened. It's wild that they didn't have any real understanding of it. And then also, kind of flip side of that is in terms of jailbreaks, sometimes simply translating your command, do something bad, right, that the language model is expected to refuse. Sometimes merely translating that into another language aside from English can be enough to circumvent the refusal behaviours as well.

So there's a lot of little interesting details Interesting, because most of the model's ability to refuse is from RLHF, which means it's been tuned to basically refuse certain types. But yeah, it's mostly probably in English.

Yeah. So translation is one of several obfuscation techniques that have been shown to, at least some of the time, of course, all this stuff is on kind of a continuum, but sometimes can get you around the refusals. All right, should we go back?

The challenge is when we describe ourselves or explain our decisions to ourselves or others, we only refer to a few of these parts. And this is because language, while powerful, is at best a linear representation of a massively parallel world. Even the way you're perceiving me right now is in parallel. You're not just hearing the words I'm saying, you're noticing my facial expressions, my tone of voice, and even reflecting on your past experience to make better sense of all of these. But if you were asked to describe my talk, you would be forced to string together words one by one by one, and that's not easy. Taking complex parallel ideas and representing them in a linear fashion is no walk in the park, and relevant to today, large language models are fantastic at this. And one of my excitements for them is that they can help us collectively better understand complex ideas, many of which may seem like magic, myth, or mystery today. But I digress. There's a few studies that discuss how if you observe an anthill or the ants within them, they act like a neural network, where each ant is a neuron, and their interactions collectively represent an intelligence that has a greater comprehension of the colony's intent and environment than the capacity of a single ant. And this is comforting to me. As what this suggests is that I too perhaps am part of something that's simply too complicated for me to fully ever understand, like the free market.

Is there more context to that as to why everybody laughed at that?

No, I mean, I think it was just a good example of a complicated system that I am part of that I won't fully understand. I think, I mean, in any discussion, you kind of have to oversimplify it to make a point. To some extent. So like something as complicated as the free market, there's tons of theories on how it works, right? Economics is a whole study on it, but all of them are ultimately incomplete because perhaps the free market itself is something that's more complicated, too complicated for us to fully understand as individuals. And I think that's okay. I think it was just a good example. Cool.

Let's keep going.

And this is in line with my limited understanding of the universe. It seems like if you talk to an expert in almost any field, They'll say something along the lines of, the more you learn, the more you realize there is to learn. Identity is the same. Understanding one's identity was never meant to be a destination. It's not something to be achieved or accomplished. It was always meant to be a continuous journey, to be explored, experienced, and even enjoyed. And not a lonely one either. Our identities are intertwined overlapping, and shared, and therefore this continuous journey of self-exploration was always meant to be a shared one. In the first age of Enlightenment, profound changes in thought emerged from political revolution, scientific discovery, and a new embrace of individuality. These changes ushered in an era that valued reason and liberty, challenging old ways of thinking. And today we're going through a similar tectonic shift where the digital and social revolution is creating an increasingly interconnected and intertwined world where old paradigms no longer apply. Just look at the world around us. And with the emergence of AI, we have a newfound ability, if not reliance, on communicating with an intelligent collective voice. If you think about it, talking to a large language model is perhaps the closest thing we have to chatting with our collective conscious, a topic worth exploring much further. See, AI isn't just reshaping our tools, but our very understanding of ourselves. And herein lies the opportunity to see AI not just as a technology, but also as a lens through which we can gain deeper insights into one another, fostering greater understanding, and empathy. With it, we have the potential to bridge divides, to bring together our diverse stories, and to weave them together into a rich collective tapestry or knowledge graph. In this emerging era, our individual narratives are vital, but it's our shared values and dreams that bind us. Harnessing the power of AI, let us seize this opportunity to collaborate like never before, ensuring that our combined voices echo with harmony and purpose. Let's build incredible things. Let's be good to one another. Let's solve some real problems. Of course, there will be challenges in many of them, and face them we shall. Together, let us embrace the promise of this new enlightenment, where the fusion of AI and our shared spirit ignites an era of unparalleled unity and progress.

I love the end of that. One thing that is striking to me routinely is just how little concrete positive vision there is for the future these days. And I wonder, there's multiple reasons that this could be the case. Some will just say, Well, our society has lost its way. We're not having kids either. There's some sort of general kind of systematic sickness that is kind of the root of all this. Another angle on it would be the future has become so uncertain and so hard to see into at all that it's just very hard to do any of this sort of future imagineering. But I do think it is something that we're really lacking right now. And so I applaud you for beginning to take on the challenge of what is a positive vision for the future of AI. Definitely encourage other people to start to wrestle with that for themselves more too. But is there anything more that you can give us? I mean, you started the conversation there. Is there anything you can continue?

When I was putting together a TED Talk, I've been a TED Talk fan for so long, and I've Rewatched many of them. Like, I wanted to do a TED Talk that I could— that was going to still be relevant, you know, 10 years from now, that like my young kids could watch before they go to college and I could talk to them about. So that was an interesting kind of like approach I had to it. I think accepting not understanding is— is not understanding everything in full is, is a key part of like accepting the way things are, is my personal philosophy. Is that when you try to really understand everything and understand anything, you, you simplify everything into a framework. And then when— but it's simplified, so then something will happen that doesn't fit into that framework, and you're confused, and that creates angst to some extent. And so these frameworks are helpful in communicating and helping understand. But it's so important to remember that this, like any, any sort of understanding you have of the world is a simplified version, to some extent for like computational efficiency. And that like, it's something that needs to be constantly adjusted as new information comes in. And patterns emerged that weren't part of the initial understanding or framework. And so I think the high-level lesson here, something for my kids that I was trying to embed, was it's okay not to understand everything. You'll continue learning. And I think this applies to identity when it comes to leveraging AI. AI is fantastic at, I'd mentioned, taking complex ideas and lining it up in a linear fashion, representing them, which to me seems like an incredible opportunity to, right? I've seen a couple of tools, early tools of like helping people understand each other that previously wouldn't, or helping find commonalities between 2 groups of people that you might not, you know, that might not be intuitive. And I think if we can start to think about AI as a tool for that, new ideas will emerge that can hopefully create a future that's not as bleak as, right, some people see the future.

In very practical terms, I've done just a couple little experiments with this. One was a fitness group chat, AI chat mediator playing the role of your virtual trainer, but just kind of there to encourage you, check in. Did you do your thing today? How many reps? Tomorrow, let's see if we can get a few more. Pretty simple, but it seemed like it worked. I also tried a dispute resolution AI between 2 neighbours. Fence is out on my property. You got to move it. It's my property. Well, can't we just agree, leave it there for now? Whatever. Minor stuff, right? But trying to help people who are not seeing eye to eye reach some sort of productive resolution. The level of connectivity between people is obviously dramatically up. Like, social media allows us to communicate with anyone, anywhere, anytime, for better and for worse. We're not always great at that. The discourse is not necessarily always super healthy. If I'm reading you right, it sounds like there's a vision of sort of bringing AI into that mix and trying to have it do some of this work of bridging gaps, bringing people together, reinterpreting perspectives. Is that kind of what you're imagining?

Yeah. And I wanted to kind of contrast, right? Again, it was an artistic choice to bring up the First Enlightenment to some extent, but it was really about individual voices, right? Versus listening to just what you're being told. But when everybody has an individual voice and everybody's connected, then suddenly it becomes a stream of basically a Twitter feed with no algorithm is basically what you end up with. So you do need to use AI, which has already been used in newsfeeds to surface information. But that is the opportunity, right? To connect people that might be relevant to each other. And we're already seeing this today. When you use Twitter and you click on see similar posts, and you find someone else that posted something relevant to a Twitter post you liked, and you follow that person, that is to some extent an AI helping you connect with someone that might have information similar to you. Again, the way it's done, the intention behind it all ultimately drives the result. But the opportunity, I think, is clearly there. Another example I wanted to touch on before I forgot was that It's going to touch on a sensitive subject, but a letter written by AI, which was an Israeli and Palestinian. There was 2 letters, each writing about why the other side was upset from their point of view, which is really interesting to just read through, right? To not read it as your own voice, like read it as the voice of someone on your side, but to read like when you hear someone else on the other side explain why you were mad, it's helpful. So someone was using AI to do that. I thought it was really well written. Again, it's not like that exercise is going to solve it, But I think exercises like this spread the right energy to look for solutions in that direction.

Yeah, I mean, I think at any scale, whether on the biggest conflicts or even just in the smallest moments, the notion that AI could help us understand one another better, bridge the disconnects in our understanding, depolarize the discourse on the margin perhaps, and help people just be better to one another and and collaborate more effectively certainly is a beautiful vision. So I love the fact that you're starting to plant that seed, and I certainly hope that something like that comes to exist because definitely we could use it. Yeah, and it's also just a beautiful notion, I think, for the future. So I love that.

Yeah, and I think I mentioned in the talk that I see myself as ultimately a conduit of ideas or mass and energy as well. So when you're given 10 minutes to To have to have a stage, right? It's— there's a choice on what message, what values, what vibe I want to spread in those 10 minutes. And it was a conscious choice to try to spread a positive one.

Well, I appreciate it. We've added a full hour of commentary to your 8 minutes of time on the TED stage. Anything else you want to touch on before we break for today?

No, this was fun. We kind of just dove into it. But for people not familiar, I run a venture fund called Untapped Capital. Being a founder, I think, is also one of the most ultimate, like, self-reflection journeys. And it's something, you know, I see it in myself as I'm building my own fund and trying to figure out who I am in the context of the VC ecosystem. But I think founders in their own industry are doing the same. So yeah, I just wanted to touch on that. I think the concept of identity is important for all of us. And I welcome thoughts and thoughts and ideas. And a lot of people have sent me some interesting books and research papers. So I welcome all of that as well.

Cool. Well, thank you for being here today, Yohei Nakajima. Thank you for being part of the Cognitive Revolution.

Thanks for having me.

It is both energizing and enlightening to hear why people listen and learn what they value about the show. So please don't hesitate to reach out via email at tcr@turpentine.co, or you can DM me on the social media platform of your choice.