E31: Where are the Moats in AI? With Nathan Labenz and Erik Torenberg

Nathan and Erik analyze the moats of the most powerful companies in AI. The paradigm-shifting technology has led to a flourishing open-source community with market share. Yet, the big players have key competitive advantages that can be examined from many different angles. LINKS: Nathan’s Twitter thread on AI moats that sparked this discussion: https://twitter.com/labenz/status/1654853321876815872 Read about the 9 moats at length in our newsletter: https://cognitiverevolution.substack.com/p/the-leaked-google-memo-and-a Leaderboard for AI rankings: https://lmsys.org/ -- We're hiring across the board at Turpentine and for Erik's personal team on other projects he's incubating. He's hiring a Chief of Staff, EA, Head of Special Projects, Investment Associate, and more. For a list of JDs, check out: eriktorenberg.com. RECOMMENDED PODCAST: The HR industry is at a crossroads. What will it take to construct the next generation of incredible businesses – and where can people leaders have the most business impact? Hosts Nolan Church and Kelli Dragovich have been through it all, the highs and the lows – IPOs, layoffs, executive turnover, board meetings, culture changes, and more. With a lineup of industry vets and experts, Nolan and Kelli break down the nitty-gritty details, trade offs, and dynamics of constructing high performing companies. Through unfiltered conversations that can only happen between seasoned practitioners, Kelli and Nolan dive deep into the kind of leadership-level strategy that often happens behind closed doors. Check out the first episode with the architect of Netflix’s culture deck Patty McCord. https://link.chtbl.com/hrheretics TIMESTAMPS: (00:00) Episode Preview (01:36) Where are the moats in AI? (07:31) Open source vs. closed (15:24) Recommendation: The AI Breakdown Podcast (16:39) Sponsor: Omneky (20:46) AI Safety (33:00) Which players are going to win? 9 Moats. TWITTER: @CogRev_Podcast @labenz (Nathan) @eriktorenberg (Erik) SPONSORS: Shopify is the global commerce platform that helps you sell at every stage of your business. Shopify powers 10% of ALL eCommerce in the US. And Shopify's the global force behind Allbirds, Rothy's, and Brooklinen, and 1,000,000s of other entrepreneurs across 175 countries. From their all-in-one e-commerce platform, to their in-person POS system – wherever and whatever you're selling, Shopify's got you covered. With free Shopify Magic, sell more with less effort by whipping up captivating content that converts – from blog posts to product descriptions using AI. Sign up for $1/month trial period: https://shopify.com/cognitive Thank you Omneky for sponsoring The Cognitive Revolution. Omneky is an omnichannel creative generation platform that lets you launch hundreds of thousands of ad iterations that actually work, customized across all platforms, with a click of a button. Omneky combines generative AI and real-time advertising data. Mention "Cog Rev" for 10% off. This show is produced by Turpentine: a network of podcasts, newsletters, and more, covering technology, business, and culture — all from the perspective of industry insiders and experts. We’re launching new shows every week, and we’re looking for industry-leading sponsors — if you think that might be you and your company, email us at [email protected]. Music Credit: MusicLM More show notes and reading material released in our Substack: https://cognitiverevolution.substack.com

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Nathan Labenz?

  So mote one, GPT 3.5 Turbo is the best value in the LLM game today. Moats number two is branding and just trust. Moats three is the feedback loop that they have. You know, nobody has the volume of LLM usage that OpenAI has. Four is pricing power. You know, $2 per million tokens. You got to be using a lot of tokens. Five, we kind of talked about again, privileged access to cloud compute. Six, GPT-4 itself. You know, they're using GPT-4 in really interesting ways that are going to give them advantage. Moats number seven, team and talent density. Satya said to their head of research at Microsoft, how the hell did they do this with a couple hundred people? We've got all these people and like, how are they kicking our butts so much?

• 0:44Nathan Labenz?

  Moats eight, insane distribution and partnerships. The customer list is growing rapidly. Number nine, yeah, network effects. You know, if I have something that's working with OpenAI and then I'm thinking about exploring something else or switching, first thing I'm going to try is the exact same task. And if it doesn't work, then I'm going to be like, oh, this kind of sucks. 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 Torenberg. So let's get into it. You wrote this thread on moats.

• 1:29Nathan Labenz?

  People are asking you a ton about moats. Where are the moats? How should people think about moats? The best kind of comparable vision that I think is intuitive, that seems likely to be right in terms of how this, you know, broad AI, you know, kind of utility intelligence market plays out over the next few years is it probably ends up looking quite a bit like cloud. And you might even say like, in the limit, it sort of could be cloud, right? If, if, you know, the algorithms are all kind of trending toward commoditized or there's like, you know, lots of stuff is getting open source, open source is making all this progress,

• 2:10Nathan Labenz?

  then, you know, eventually like maybe the same stuff runs everywhere and it's, uh, you know, it's, it's just how many computers do you have kind of determines, you know, how much AI. Utility you can deliver. I don't think it's quite that simple, but it does seem like it's kind of shaping up on somewhat similar lines where, you know, in cloud, you have kind of big tech oligopoly has half-ish of the market. When you look at like AWS, Microsoft, Google, Oracle, you know, whoever you, wherever you want to draw that line on who is kind of, you know, the top tier cloud providers, they have about half. And it seems like if you were to just ask like how, you know, what is the breakdown of

• 2:50Nathan Labenz?

  where inference is served over the next few years, it seems like it's probably similar. You know, this time it's probably Microsoft and OpenAI that are in the lead in terms of, you know, they've got the best product. And they've certainly got the cloud, you know, to back it up. So they probably lead and then, you know, Google DeepMind, which is also their, you know, their cloud is powering Anthropic is probably next. And then AWS is a little behind at the moment, but you know, has hugging face and has plenty of stuff. Plus they just have like tons of people already using their cloud. So, you know, for, for many kind of homespun, you know, open source, you got to host it somewhere. So it's going to be, you know, a natural place.

• 3:36Nathan Labenz?

  And then you've got long tail, you know, besides that, and that can be like still on those same public clouds, but you know, different service layers. It could be, you know, your own on-prem solution, if you're an enterprise, you know, that's supported by wherever, whoever. And, you know, even on consumer devices, like it seems, you know, there, we're seeing a lot of trends toward that as well. So, you know, in kind of keeping with my theme of like everything, everywhere, all at once. I think there's no reason to suspect really that like any significant pools of compute would not be used to power AI over the next few years. And so that market structure seems like it probably lines up. You know, if for that not to happen,

• 4:21Nathan Labenz?

  somebody would either have to like run away with things dramatically, which doesn't seem like it's currently happening, like OpenAI is ahead, but it doesn't seem like they're, you know, light years ahead. And they're like maybe eighteen months to two years ahead of open source, which is to say like what they were offering commercially eighteen months to two years ago is about what open source has got to now. Not to say that open source will catch up to where they are in the next two years. That remains to be seen. Uh, GPT four is going to be, you know, harder to match. You know, it takes more compute if nothing else um than, you know, GPT three was to match. Cost of GPT three quality models is now under five hundred thousand dollars. If you go to Mosaic from scratch, you know, bring your own

• 5:07Nathan Labenz?

  trillion tokens of whatever data you have and you can get to GPT three uh quality, smaller model, more intensive training, but under five hundred thousand bucks. But yeah, meanwhile, GPT four is still rumored to be a hundred million. Uh, that's not going to be, you know, so easy for the open source folks to to replicate. But anyway, you know, it seems like it's it's kind of headed towards shaping up like the cloud market. You've got the oligopoly that kind of combine, you know, the hardware, infrastructure and kind of services layer, which in this, you know, for now is like the models themselves. They're kind of the the core differentiator there. And then you've got, you know, a ton of other things too that kind of make up the other half of the market.

• 5:49Nathan Labenz?

  And so, you know, just like AWS, you know, certainly has moats in general and compute, like all these big players have moats to varying degrees. And, you know, I think that's ultimately pretty obvious, you know, that that whole like no moats thesis, it's like,

• 6:08Nathan Labenz?

  I don't know, I just, I think that that hasn't really held up over the last couple of weeks under deeper scrutiny. And, you know, it's the negation of that memo isn't true either. Like, I think that that memo had a lot of things going forward in terms of the flourishing of what's happening in the open source community. But it's not, you know, it's still not the case that like everybody doesn't need open AI anymore. You know, it's it's far uh far from the case, you know, the the folks that are trying to implement this stuff in medicine. They are not that cost sensitive, you know, they want the best thing. And open AI has it, you know, and Google is like, you know, trying to get on their level and and certainly seems like they they can uh or can get real close.

• 6:51Nathan Labenz?

  So, you know, that's just not something that anybody in the open source community has any credible. Claim against, you know what I mean? And you can sort that out at the level of like a quick demo. You don't need a, you know, deep evaluation process. You can literally can just use common sense and ask some questions to GPT four or, you know, to like the new Med Palm two out of Google. And, you know, you'll, you'll know, you'll taste the difference between that and like, you know, whatever the latest woolly animal, you know, named model might be. Again, there's a lot of modes. We could break it down from a lot of angles, but that position is like, I think, pretty secure for the foreseeable future.

• 7:31Erik Torenberg?

  Zooming out, it's interesting to think about when you have a new kind of platform shift or technological breakthrough, there's some that have led to more sort of open source or open source would have bigger market share and some that have led to the opposite. It's interesting to think about the kind of characteristics that make either happen.

• 7:48Nathan Labenz?

  Yeah, it'll probably continue to evolve as well. And, you know, I'm certainly no scholar of kind of earlier software waves. Certainly haven't studied any of those with anything close to the intensity that I've been studying the AI wave. But, you know, so much depends right now on context. You know, people ask these questions like, why? So how would you decide? That's a question I've been getting a lot. All right. Let's say you are you, you, you're maybe at Waymark or you're in some other context. Like, how do people decide whether they're going to use an open AI or an open source model? And I do think like most of the, in most contexts, it does seem pretty clear that you do perfectly well with an open AI product.

• 8:31Nathan Labenz?

  At like Athena, for example, you know, where I'm the AI advisor, the explicit strategy, if we're going to try to do some new task, you know, and set up some automation to power some task, it's always just start with GPT four. There's no reason, you know, to start anywhere else. It's like clearly number one. It's number one on all the leaderboards. It follows instructions the best. It's not that expensive. It is a little slow, which can be a little cumbersome when you're trying to test stuff out. But, you know, it's the fastest path to. You know, just the raw, the application of this raw intelligence to your problem to like demonstrate whether or not you can even automate this task. You know, GPT four is going to give you the fastest path to that.

• 9:13Nathan Labenz?

  And of course, there are some things that can't do that. Maybe you could fine tune a model for them. And that's a whole other scale of project. So that could be one reason, you know, you could, you might be able to dial something in better. If it's a narrow application and you're really kind of focusing in on a use case. You know, that could be one, there are reasons to go away from GPT four. That that is one of them. But you know, the, the reasons that people tend to think more often, I think are really not great reasons like. Cost, you know, it's like, it's pretty cheap, actually. Uh, and it's especially cheap if you consider a total cost of ownership. Because I can get in there and, you know, take most common tasks. And get the AI performing reasonably well on it.

• 9:54Nathan Labenz?

  In, you know, with a, with a few rounds of prompt engineering, often in like. Thirty minutes or less. And, you know, what is the cost of the person's time? Like, you know, how many, how many runs of this? Task automation do you need to be planning for? Before it is going to pay back even to just spend a handful more hours looking for a cheaper option, you know, I mean.

• 10:20Nathan Labenz?

  What is the end, you know, of your task? But most of, most tasks, people are not trying to scale to the billions. So, or even to the millions. You know, so you think about like the, the cost per token on a. GPT four, three cents per thousand tokens in, six cents per thousand tokens out. You know, maybe just for simple math, average that to four thou, four cents per thousand tokens. That's four cents per two pages. Of text processed. And, you know, the kinds of tasks that we're setting up are like. Can we take a long transcript of a call that we have with a client? And convert that into a.

• 11:01Nathan Labenz?

  You know, condensed, information-rich profile of the client that we can then put in the assistant's hands as they're beginning a relationship to give them, you know, as much context as we possibly can. And, you know, that cashes out to however many pages, right? But it's not that many pages. We might be able to power this for a quarter. You know, it used to take four hours. So we're saving a ton. That's, I mean, when people get very obsessed with these cost questions, I'm often like, most people don't care if they're saving 90%, 95% or 99%, you know, especially if it's like all kind of new. You know, the cost I just don't find to be that compelling. One, there are some applications, you know, where if you're doing kind of high-volume

• 11:46Nathan Labenz?

  generation or and you have a low hit rate. This is another thing I've been kind of highlighting for people, you know, with Waymark. We have a pretty tight process. And, you know, we've been doing this before generative AI. So we make these, you know, advertising, marketing videos for small businesses. Typically, they're like 30-second commercials. They might run on TV. We've been doing this, you know, before generative AI. Now the generative AI writes the full script, picks all the images, you know, layers on the voiceover, but we still had that structure, you know, even before, you know, GPT-3 popped up. So the ratio of like, how many of your generations are ultimately useful to people?

• 12:29Nathan Labenz?

  Is a, is a reason that you might end up wanting to go, you know, towards some cheaper model. So like Jasper, for example, they're now on the Mosaic homepage as a client. And, you know, they've been an OpenAI client, obviously from the beginning, right? Helping with all this marketing copy. You know, they're probably spending quite a lot. So for one thing, they have like some but real budget, you know, that they can maybe take a chunk out of if they can fine tune their own stuff. But then I also suspect that a big part of what's going on there is because a lot of their stuff is like fairly open-ended, like, you know, give us two words and generate a LinkedIn profile or whatever. I'm guessing they're seeing a lot of generations. And, you know, people are maybe just kind of using it in sort of a rifle-through sort

• 13:13Nathan Labenz?

  of way, you know, seeing it kind of like what, uh, Zuhail told us with Playground, where people are making, you know, 10% of their users are making more than a thousand images a day. It sounds like that's kind of the pattern that's going on at a, at a Jasper to some degree. You know, and that's a, that's a big contrast to like with way markets, basically like one in three, maybe one in four, maybe one in five, you know, probably varies a little bit by cohort of the videos that we generate ultimately gets rendered and downloaded. So people are not like sitting there rerunning it that many times. And so like the economic, you know, relationship between the cost of that generation and what they're getting for it, you know, where they're going to go spend thousands of dollars on a TV campaign, you know, that ratio is totally

• 13:56Nathan Labenz?

  fine. And then when you get to task automation, you know, your hope is that you're going to get to basically like use all of the work. So, and that's like fairly feasible, you know, when we do these transcript to, um, you know, client profile workflows, the goal is like not to have to run it a bunch of times, you know, and, and from what we're seeing, it looks like, yeah, it's basically going to work. You know, our workflow is going to be immediately after the call. And again, this used to take a couple of days and whatever, and people get bogged down and the things would take hours to write. Now you've got, you know, basically instantaneous

• 14:35Nathan Labenz?

  transcription, you feed that into the AI process that chunks that, when you know, the biggest limitation with GPT four, biggest two limitations are context window is still limited to 8,000 and it's kind of slow. So, you know, when you're chunking these long transcripts into, into bits in order to summarize them in order to then have, you know, a kind of unified summary that you can process into whatever format you want. It can be like a little bit of a, you know, friction point. It can be a little bit slow, but it is the kind of thing that we're basically targeting. Like you should be able to use a hundred percent of this output. Like we're not seeing any instances where it's like, what is this unusable garbage, you know?

• 15:20Erik Torenberg?

  Hey, we'll continue our interview in a moment after a word from our sponsors.

• 15:24Nathan Labenz?

  Hi, everyone. I wanted to take just a moment to share another podcast that I've been enjoying recently, the AI breakdown. As anyone in AI knows, the pace of progress and all the new releases are relentless. I call myself an AI scout and I work overtime to keep up. But these days, even I can't keep track of everything. The AI Breakdown helps me make sure that I don't miss anything important by curating news and analysis daily. Host Nathaniel Widmore, aka NLW, quickly highlights the top stories of the day before going deeper on a single topic of interest. Episodes are usually 15 to 20 minutes, and he releases them every single day. Now, it's not easy

• 16:09Nathan Labenz?

  to keep up with a daily release schedule and still maintain your sanity. So I really appreciate how NLW maintains a curious posture and avoids rushing to judgment. A big part of the reason I'm inclined to recommend the show is his willingness to sometimes say, "I don't know." I think listeners will find the AI Breakdown to be a great complement to the long-form, deep-dive interviews that we create, so I encourage you to check it out. The link to the AI Breakdown with NLW is in the show notes.

• 16:40Erik Torenberg?

  Omnikie uses generative AI to enable you to launch hundreds of thousands of added iterations that actually work, customized across all platforms with a click of a button. I believe in Omnikie so much that I invested in it, and I recommend you use it, too. Use CogRev to get a 10% discount.

• 16:58Nathan Labenz?

  The human review layer at the end is more to be like, "Okay, I just talked to this person for 90 minutes. Does this thing, you know, miss anything that seems important?" You know, and we have, we do see issues like that where it's like, "Yeah, this person said this, and that does seem pretty important, and the AI didn't, you know, quite pick up on it as a key point, didn't represent it in the final thing." So it's not like it's a flawless process, but you can go from, you know, just having got off the call to like, looking at a draft of your final document in a couple of minutes, and then, you know, read through it, figure out what rings, you know. Mostly, it's good, like, but, you know, there might be something that doesn't quite ring true, or something that was missed, and then you're kind of done, you know. Still in that whole process, it's like, the cost of the AI is very, is very small. You know, we had a person that

• 17:43Nathan Labenz?

  sat there and talked for 90 minutes. We're gonna have, like, all these, you know, things going on.

• 17:48Nathan Labenz?

  There's no reason to, you know, try to save a couple more pennies at this stage on that process. Like, it's all about driving the quality. You know, we wouldn't even consider an open source model. You know, there's just no, no reason for it. Now, we're also probably not going to be that huge of a customer because we don't have that many clients coming in. But you start to extend this to other things where there are, you know, points in the business where there's real scale. The next one downstream is client matching with the candidates. And they may be onboarding a hundred clients a month and you may have like 200 candidates in the pool that have already been, you know, filtered and you've gone through a whole wringer. I think they hire like sub 1% of like initial applicants.

• 18:32Nathan Labenz?

  To do every analysis there is like 20,000 of these sort of comparisons. You know, this, this client profile and this assistant profile, like, is that a good match? It's not easy to scale up to 20,000 of those with human power. You know, in practice, I think historically you would more kind of look for, you know, look until you found one that seemed good and kind of have to stop there because it just isn't time, you know, to do the fully exhaustive version.

• 19:04Nathan Labenz?

  But the AI can, you know, can run overnight and do that more exhaustive version. And, you know, it's still going to be cheaper than the human, you know, powered thing. And it might even be better. I wouldn't get on quite yet and say it's going to be better. But I think it can match, you know, pretty readily. And the prospect for it to be better is definitely there as well. And to make it more, you know, just more personalized, more, um, you know, to then turn around and be like, all right, now write a, you know, blurb introducing this person to the client. You can enhance the client experience in ways that, you know, don't even necessarily.

• 19:44Nathan Labenz?

  I mean, not even necessarily about the fundamentals, but just like, let's get this thing off to a good start by kind of introducing this person in a really nice personalized way. Yeah, anyway, in all that, cost is not an issue. I can tell you that.

• 19:54Erik Torenberg?

  Do you see the gap? You mentioned open source about two years, you know, behind. Do you see that gap widening or shrinking over time? Like, what do you predict?

• 20:02Nathan Labenz?

  I don't know. It probably goes back and forth. I think it'll probably a yo-yo effect. I mean, the the open source line is rising all the time. And then you probably have more like punctuated, big next releases from the key players. The perceived public gap probably closes and then widens, you know, again, and it won't be obvious like necessarily is the overall gap, you know, given what they have internally that the rest of the world doesn't know about. Is that like shrinking or narrowing? I bet would be pretty hard to tell still for a while. And it seems like they're they are probably going to cluster like it's there's a weird tension right now between the fact that

• 20:45Nathan Labenz?

  the leading companies, namely OpenAI and Google DeepMind, Anthropic, are pretty clearly like coordinating on some very meta level, you know, and and kind of saying very agreeable statements, you know, to each other about how like, you know, what's most important is that we do this safely and that, you know, everybody can benefit from it. And, you know, Google's CEO Sundar just said, the only race that, you know, matters is the race to safety or something like that. We should find that exact quote. That's not quite what it was. But he's, you know, specifically trying to say we are not going to get into an AI race, which, you know, I really view honestly as

• 21:29Nathan Labenz?

  great validation for the entire AI safety movement and, you know, even specifically the EA AI safety movement because.

• 21:42Nathan Labenz?

  It's very easy to imagine a counterfactual where people running the cutting-edge AI companies were just totally dismissive of big-picture AI risks. And the fact that you've got basically all three of the leaders who have a real, you know, aware, like a demonstrated awareness of the, you know, and seem to be pretty credible in taking seriously these, you know, tail X-risk scenarios. You know, I don't think that that necessarily happens in a, you know, counterfactual world where there's no EA movement or where there's no like Eliezer. Their influence seems like pretty clear there.

• 22:27Erik Torenberg?

  I certainly agree. I think the question there is, you know, actions speak louder than words. And how much will the actions match? Like, if they were in a race, would they be acting any differently? Will they act any differently? And how so? Like, or how extreme? Because in some ways, if you're a little cynical, you could say it's, you know, it's reverse psychology in this really effective way, which is what, you know, say it's not a race as a way to potentially throw it off competitors or not, you know, threaten competitors, but also on regulatory side. Like, we have this weird dynamic where for other technologies, you know, like, let's say crypto, people said, hey, this is like the next great thing, like, everyone needs this. And then on the regulatory side or media side, you see response of, hey, actually, maybe this is bad

• 23:09Erik Torenberg?

  or maybe this isn't great or it's too powerful or it's too dumb or whatever. But with AI, it was really interesting to think that the hearings and thinking about kind of the response of media and regulatory, it's like people's, you know, I think some accuse them of marketing, like by emphasizing the, like, how dangerous this thing is. It also emphasizes how powerful it is. And some people in media and, you know, sort of government are saying, hey, like, it can't do as much as you say it. Like, they're kind of down, they're more scared of social media and misinformation and the complaints they have on social media, teen depression, et cetera, than they are on AI. Like, I would have expected people to be much more concerned than they are. It seems that most of the concern is happening from within tech. And I wonder if that's

• 23:54Erik Torenberg?

  Unconsciously, a form of reverse psychology in the sense that because tech is so internally worried and about it, maybe that encourages them because they're like somewhat anti-tech or somewhat contrary to tech to be less worried.

• 24:08Nathan Labenz?

  I think, you know, Sam Altman's statements about what kind of regulation they recommend have been pretty clear and not, you know, not always interpreted in good faith. So, you know, he's literally explicitly said, we think the regulation should apply only to the leading companies that are doing the biggest models, you know, the biggest compute budget, you know, training runs, and we don't want to interfere with companies, open source, you know, research, startups, et cetera. And that seems like a very reasonable position for the, you know, for the regulatory bodies

• 24:51Nathan Labenz?

  to take, and I hope that they take something along those lines. Like, that seems pretty sane. So, yes, there could be some amount of, you know, hype marketing in that, but really, they don't market at all, you know, they like, they don't need to market. The business is, you know, the phone is ringing off the hook. When I talked to Mosaic, you know, ML guys recently, Jonathan and Abhi, they were like, yeah, we can't run long. And they also said, we're starting, he's like, if you want to be a customer, you better hurry up and call us because we are starting to get to the point where we're making serious, like, tough decisions between research and customers.

• 25:34Nathan Labenz?

  And basically, you know, they're not going to cut research to nothing. They know they can't do that. So, you know, it's like, you're going to, you might have a waitlist at Mosaic, you know, pretty soon. So.

• 25:46Nathan Labenz?

  You know, you, the, the level of spend that you have to commit to OpenAI to get some, you know, some commitment from them of some attention to you has risen dramatically. When I bought it, you know, just over a year ago, and we were already a retail customer, like on the API and fine tuning models and stuff. But as we're starting to get more serious and I, I was like, I kind of want, you know, a tutor or like a consultant, you know, especially on the inside of OpenAI. You know, what do I have to do to like, get you guys to like, take it, you know, every other week call with me? It ended up being at the time, 2,500 bucks a month was the like, service package that we bought into. And now they don't even offer that.

• 26:31Nathan Labenz?

  And, you know, to get that kind of, you know, account manager or whatever, you're at least into six figures, if not like a quarter million, um, you know, kind of upfront commitment that you're going to spend that with them to have them like, take your calls. So I don't really think he needs to go in front of Congress and call for regulation as a marketing strategy. You know, they could, they also like downplay their shit. I mean, you want to, if you want to say that they, you know, are hyping everything, look at the launch statement that Sam Altman put out on GPT four day. One of his very first things was, it seems more impressive at first than it really is, which is, is true, you know, and is honest. That is an honest statement.

• 27:15Nathan Labenz?

  But, you know, you don't really see tech CEOs downplaying their major launches on launch day. And yet they did that in a very like, clear way, you know? Sam's, I say, I don't know if Sam Altman, but Sam A's, uh, you know, tweet, it's, it goes something like, here is GPT four, it's our most capable model, yet like all of our models, it still makes stuff up. It's still, you know, has major weaknesses and it's, you know, appears more impressive at first than it ultimately is. Like, that is not, you know, how most, uh, you know, that's, that doesn't sound like a hype cycle to me, or like, uh, you know, a, a marketing ploy. I think they really mean it, you know?

• 27:58Nathan Labenz?

  I, I, I think that at the end of the day, I think that's becoming increasingly clear. Now, you could really ask, and I was thinking about this earlier today, too. You know, you read their governance statement from this last week and you get down to the bottom of it. And the last point that they make is we think that it's counterintuitively dangerous to not develop this stuff. Because the fundamentals are all going this way. And that means that it's increasingly easy to develop these, you know, powerful systems. And if we don't kind of keep the gap between what's possible and what is, you know, what exists like somewhat narrow.

• 28:43Nathan Labenz?

  Then we may have these like sudden, super disruptive events in the future if somebody, you know, achieved some breakthrough unexpectedly. And all of a sudden, you know, it drops into an unprepared society.

• 28:54Nathan Labenz?

  And it's like, okay, I don't know if I really. Follow that to the same conclusion that they, I mean, it almost kind of seems like they're saying. Somebody's going to develop. AI so powerful that it's dangerous and better us than them, whoever, you know, for, I guess. Basically all, you know, other thems that it could be. I don't necessarily want to sign on to that or endorse that. Um, but I do think, you know, everything they're doing is pretty consistent with that. They do seem to be trying to rise to the occasion. And. You know, they definitely could be jamming stuff out faster than they are. They even, you know, they have that part of the charter too, which.

• 29:39Nathan Labenz?

  I mean, people, obviously at this point, don't trust, you know, them on necessarily anything, but it is in the charter that they will. Uh, combine forces with another effort that they believe is, you know, credibly close to AGI rather than compete with them. You know, that's a hell of a commitment to have made, you know, whatever in 2015 or whenever they came up with the charter that, um. You know, at that time, it must've seemed like a pretty distant, uh, worry. But, you know, they've, they've had that online for a long time. That is not like they just released that. The statement that's been there for years. Yeah, going back to your like with the open source and closing, you know, the opening and closing of the of the closed and open source gap.

• 30:20Nathan Labenz?

  I do think we'll see probably these releases again come in some kind of cluster. Where, you know, GPT four and kind of poem two, you know, came in a pretty tight window. I would bet that in the future. That will continue to happen as a reflection of the fact that leadership at these companies is genuinely concerned with, you know, the the possibility of an AI race genuinely doesn't want to create that dynamic. And so we'll kind of talk to each other behind the scenes a little bit and say like, you know, can we can we kind of ease into this stuff together? So we're not, you know, you're going to have an next generation thing. We're going to have a next generation thing.

• 31:02Nathan Labenz?

  If we kind of bring those online around the same time, we can, you know, the oligopoly is not going to get too disrupted. Microsoft and Google don't have to like, you know, battle to the death. They can both kind of advance similarly. And, you know, that's probably a really good thing as long as, you know, they don't collectively still like love the world in the process. You know, but it does seem like definitely way better than a different dynamic where they're like sworn enemies, you know, won't talk and, you know, are trying to one up each other, which we saw a little bit of that earlier this year, but it seems like they've kind of, you know, Google has not taken that bait. I think to their, I think they will be vindicated in the end for saying like, all right, you,

• 31:46Nathan Labenz?

  you go ahead, but we're going to do it when we're ready and, you know. So far, it doesn't even seem like they've lost significant market share. As far as I can tell, like maybe a couple of points, but not like, you know, they're still dominating Bing in terms of just raw search volume. So, you know, it seems like even from a shareholder standpoint, people were not that were, you know, they, the consumer wasn't that ready to switch where like, you know, corporate requirements would dictate you like must ship immediately. So I think their judgment has has been actually pretty good and vindicated. And, you know, again, I think that's by extension of another vindication of the kind of EA AI safety crowd because. You know, he didn't, you know, I don't think he grew up thinking about this kind of stuff.

• 32:30Nathan Labenz?

  So he needed some, you know, people intellectually to have done that work to be able to tap into it. You know, if there's no literature, I don't think these people are coming to these, you know, some of them might, but like, Soodar probably not, you know, he on his, on his own, totally, you know, totally organically.

• 32:48Erik Torenberg?

  I agree. It's, it's very influential literature. Marx too is very influential literature. I'm, I'm just teasing. Let's go deeper on sort of the respective kind of assets. You know, we're talking about NBA the other day where, you know, we're talking about like the Nuggets chances versus the Boston versus the Heat. We talk about their relative strengths and weaknesses and, and kind of where they're at now. Let's do something similar with, with, with, with some of these players that we're talking about in terms of like, how, how do we see things playing out based on what are the, what are the, you know, the facts on the field, so to speak, in terms of people's relative strengths, you know, weaknesses, opportunities, et cetera.

• 33:24Nathan Labenz?

  We'll just run down the quick nine moats that identified for OpenAI and then we can kind of compare that to Google. Most of them are basically the same and OpenAI is just a little stronger, you know, in, in a number of these categories, it seems like right now. But also you could make the argument that, you know, that Google is stronger in, in some of the, maybe the most important categories as well. So moat one, GPT 3.5 Turbo is the best value in the LLM game today. So, you know, earlier I talked about how like GPT four is generally a huge savings over a human powered process. And, you know, if I'm, if I'm new to AI and trying to make stuff work, there's really no incentive for me to go

• 34:03Nathan Labenz?

  anywhere else. Well, if I were going to go anywhere else, they happen to have a 60x cheaper model that is instead of, you know, four cent, no 50x, I guess, whatever. Instead of four cents per thousand tokens, it's point two cents per thousand tokens, which is $2 per million words. And that's what powers the ChatGPT free tier.

• 34:26Nathan Labenz?

  And I said it's the best value in the utility L M game today. I think that's probably still true in there's a new leaderboard that I've started to follow called L M sis dot org. Where they literally have like chess style head to head battles between language models and keep track of elo ratings. So users go in, you get, you know, for the same input, you get the response from different models, you choose the winner, they keep the score. And G P T four is, you know, at the top of the power rankings, Claude is number two. And Claude instant has, which is kind of the the anthropic answer to the chat G P T, has actually taken third place recently and put turbo into fourth place.

• 35:06Nathan Labenz?

  So arguably, you might say at this point that Claude instant could be, you know, even slight edge over G P T three point five turbo for the best value in the game today. But either way, you know, they both have great value. They're both, you know, they're both cheap, they're easy, they're fast, they can do a ton of stuff, they can handle the marketing, you know, copy tasks for the most part. You know, they can return formats pretty reliably. They can't do the advanced stuff of G P T four, you know, it's this is like the but the difference between like bottom ten percent on the bar exam and top ten percent of the bar exam. That's like, you know, three point five to four is that is that leap? But you know, still bottom ten percent of the bar exam, you can do a lot of stuff.

• 35:51Nathan Labenz?

  You can process a lot of data, you know, you can. You know, you can organize a shopping list, you know, there's plenty of things that you can do, you know, without being quite powerful enough to pass the bar. So that's a great product, you know, and it's it's still better than all of the open source imitators. And in fact, all of these open source things, I mean, not all, but like a lot are using G P D output, you know, from Open A I and just like imitate training on that. So, you know, they are like getting somewhat close on some test domains. But, you know, they're not really that close to even, and the leaderboard will tell you this, you know, and that's a blind head to head.

• 36:31Nathan Labenz?

  You know, user call the winner process, you know, they're not that close even to, you know, the second tier.

• 36:41Nathan Labenz?

  Five percent, you know, price of GPT four version of OpenAI's products. So just having that means like, there's not that much opportunity for people to come in and like, steal share, you know, on pure price reasons. They have not left the door open there all that much. You know, so people could do it for maybe other reasons, you know, control, fine tuning, you know, ideology, you know, data, not wanting to send data out over certain, you know, boundaries, whatever. But, you know, they've, they've got a great commodity product in today's world. I'd say, you know, again, it's pretty clear at this point, OpenAI and Anthropic are the two leaders in that category.

• 37:25Nathan Labenz?

  Mode number two is branding and just trust. The, you know, for all of the complaining that has been posted on the internet about how like,

• 37:39Nathan Labenz?

  chat GPT is, you know, to this, to that. You know, I mean, and you hear it from all sides, right? It's, it's, it's simultaneously, depending on who is, you know, trying to embarrass it, it can be both, you know, too like white supremacist and too woke. Uh, you can kind of, you know, see that almost regardless, I think of perspective, you can see like unwanted biases or unwanted behaviors in it. But the alternative in the open source world is just way worse, you know, if you are a corporate customer. If you want, you know, a sort of radical free speech experience and you're, you know, not a major company, then you can go the open source route and do, you know, whatever you want.

• 38:21Nathan Labenz?

  But if you are the kind of business that is thinking about maybe putting a chatbot into your product experience somewhere, you know, you don't want it to.

• 38:33Nathan Labenz?

  Get too adventurous, you know, I was joking with a guy at OpenAI. I was like, yeah, these, you know, the sort of open source radicals like really overestimate the corporate appetite for large language model adventure. You know, nobody wants their own, you know, personal Sydney experience. Nobody wants that kind of embarrassment. And when you can get, you know, at $2 per million tokens, good quality, pretty reliable service, you know, from OpenAI, like, then it's kind of burden's kind of on you to figure out like, well, why did you do something, you know, why would you or why did you do something different? So I started saying that, you know, people used to say

• 39:14Nathan Labenz?

  nobody got fired for going with IBM. And now it's like, that might be true for a few of these top players. Because your AI might still embarrass you, but at least you can fall back on like, look, I used like the industry standard, like these guys, you know, they spent six months, you know, working on safety of GPT four, like, what do you want me to do, boss? Like, if we're gonna use this kind of shit, there's gonna be some risk, but I, you know, I kind of made the safest choice I could with OpenAI or Anthropic, like, you know, so again, that's a moat, right? And Google's, I think, gonna get there too soon. Certainly, they have the like, I think, you know, trust and kind of gravitas that a corporate buyer would believe them that like, they do,

• 39:54Nathan Labenz?

  you know, have good standards in place. And, you know, again, when it's so cheap, like, why am I gonna go, you know, put myself as a, you know, a CIO or whatever in a position where I don't really know where this, you know, exactly how this open source model was trained, or by whom, or like, to what degree it's really been battle tested. Like, I'm gonna do that to save a little money. And by the way, am I even really saving money? Depending, you know, maybe, but I'm certainly gonna be putting more man hours into it, you know, than I would if I was just doing, you know, the simple thing. So.

• 40:32Nathan Labenz?

  You know, I don't know. It sounds kind of tough. Mode three is the feedback loop that they have. And this is where OpenAI is currently quite out in front. Depending on exactly what you think about, you know, some of the alternative strategies, but you know, nobody has the volume of LLM usage that OpenAI has, you know, that ChatGPT has. You know, Bing's doing some decent volume I've seen, but not as much as ChatGPT. And they're of course, you know, powered by GPT four anyway. So they're getting, you know, this data, they now have terms of service that say if you use them via the API, that they will retain your data for a short time, then delete it and not use it in training. But the free tier of ChatGPT by default is used in training.

• 41:16Nathan Labenz?

  And you can opt out of that. Now they just added that as well. But you know, by default, you're opt in. And if you want to opt out, you can opt out. But they're getting more, you know, raw usage and feedback data than anybody else. And they have a well honed product development process, you know, that is that is humming. And others are finding that, you know, kind of hard to reproduce. Even somebody like Bing, you know, comes out and finds out, wow.

• 41:47Nathan Labenz?

  You know, we didn't expect that. And you know, some of those things you're like, I don't, you know, there's, there were some really pretty flagrant breaks in the, in the Microsoft process. I've, I've gone down that rabbit hole. They tested this thing for months in other parts of the world. They had users report in their forum. Such behavior from the bot has ultimately graced the cover of the New York Times. They failed to detect it. Cause like one part of the organization wasn't talking to the other or whatever. The, the Microsoft employee that responds in the forum, last I checked, this was also online in the Microsoft forums. The person who responds, who works at Microsoft seems to not know about

• 42:32Nathan Labenz?

  the AI powered search experience at all. And it's like, doesn't even know what the person is talking about, who is saying like, your chatbot is accosting me. And You know, that's all kind of documented as of like late 2022. Then, you know, they launched the thing in 2023. And then you have things that are as simple as like you got the date wrong. The AI got the date wrong. The user corrected the AI on the date. And that was enough at launch for whole things to go off the rails, you know? And I think it's a huge, people should be very clear on the difference between like a jailbreak, you know, where you can trick ChatGPT into saying a bad word or whatever.

• 43:15Nathan Labenz?

  And, you know, and that's a problem. But I've never seen ChatGPT turn on a user, you know? And that's a whole different level of like failure of alignment, you know? Failure of control when the user is trying to break your control measure versus like failure of basically, you know, of decent engagement with the user in the first place. So qualitatively different thing. All that story, you know, is to say the product feedback loop is pretty important. And, you know, it's, this stuff does not just like magically cohere into a well behaved AI by accident, or, you know, even with like moderate effort. It seems to take a lot of effort.

• 44:00Nathan Labenz?

  So OpenAI is just like crushing it in that regard. Anthropic, again, has a unique approach with their constitutional AI system where they use kind of self critique and like synthetic data to basically try to get, you know, the same level of control as if they had, you know, the user base. And it seems to work quite well. So that may be something that other people could, in fact, in some ways they are ahead in terms of like their safety profile, not their raw capability profile.

• 44:31Nathan Labenz?

  But in terms of their safety profile, in some ways I've seen them be ahead. I've also heard from others that maybe that's not the case. So it's probably mixed, you know, the surface area is so huge that you can be ahead and, and behind at the same time in different areas, and that's almost certainly the case. But nevertheless, it does seem to work quite well. They also do a pretty good job of avoiding hallucinations. At one point, I think they did a better job than, than OpenAI was, but again, with GPT four, you know, they've, they've improved a lot. But like you could previously see with, with ChatGPT versus Claude, you could kind of see these situations where. ChatGPT would still, I asked a question about like a property tax in one Massachusetts town one time. And it made up a rate that was not the real rate.

• 45:15Nathan Labenz?

  It had all the structure and conceptual analysis of property tax generally right, but it made up a rate at, you know, the key moment that was false. Whereas Claude said, for example, if the rate were, you know, it just gave like a nice reasonable round number. And so, and there, there are some ways in which it does seem like the anthropic method is, you know, at least competitive, if not in some ways, you know, maybe superior. So those two seem to be kind of at the forefront of that. Actually, character AI, you know, has a really good feedback loop as well. They're kind of a dark horse in this whole game, playing a bit of a different game. They're the, they're the only, the one that comes to mind is like having a real, you know, tight at scale product feedback loop.

• 45:59Nathan Labenz?

  Google doesn't have it yet. They should be able to get it real quick. Whether or not they can really, you know, iron out all the sort of, you know, sand off all the barnacles on their sort of product flywheel. I guess still remains to be seen. As of now, they've not done it. And if you look at the leaderboard, Bard is down the leaderboard. Not because, you know, they don't have the raw horsepower, but because, at least as far as I can tell, they have not shaped the product in the same way that the, you know, the best in that category have. You know, they can match on things like medical question answering.

• 46:44Nathan Labenz?

  You know, GPT four and Medpom two are both like expert level on these benchmarks. So they can definitely get to the point where, you know, they can do high level stuff. But they don't have this like unified product that kind of does everything and, you know, knows how to handle all these different situations. It's like Medpom is dedicated to medical question answering and it can't help you, you know, with. Whatever random search queries you have or you know, just be your like general sort of, you know, chat advisor, it's, it is specialized and, and you know, fully end to end fine tuned for that domain. And that arguably could be good. You know, you, you, somebody might say, well, geez, do we really need these like all purpose AIs? Maybe like a bunch of specialist

• 47:28Nathan Labenz?

  AIs could be, you know, arguably a better overall architecture. I think Google is definitely kind of playing that both ways. Like they're going to have their general open chat bot that you can talk to about anything, but they're also going, they're taking Medpom two to like hospital group and stuff like that. You know, they're not, um, they're not even messing around with some base, yeah, some base model. They're going to only take, you know, the, the good stuff. So four, I kind of talked about in my, in this thread, four is pricing power. I think we've probably covered that, you know, at enough length. You know, $2 per million tokens. You, you got to be using a lot of tokens. And another way I was thinking about this was, uh, this is actually kind of interesting math. If you

• 48:13Nathan Labenz?

  were to try to compete at that price point, you would have to serve a hundred billion tokens to pay for one employee. They make $200,000 in revenue for a hundred billion tokens served. So I'm, you know, rounding down from the, uh, you know, we're so back San Francisco AI, uh, salaries there to 200,000, calling that one employee. Um, by the way, you know, whatever electricity, you know, compute costs you have, uh, you know, Microsoft

• 48:55Nathan Labenz?

  gets its share out of that too, right? Out of, out of that, uh, $200,000 for a hundred billion tokens. I mean, a hundred billion tokens is like a ridiculous amount of stuff. You know, like they're training these models on like a trillion tokens. That's like a kind of general ballpark of like what a, not like a GPT four, but like a, you know, a llama type, good open source project might train on today would be like roughly a trillion tokens. So you're talking about like 10, you know, and people also talk about that as like the scale of the whole internet. So you're talking about like generating. Text on, you know, maybe 10% of the, you know, the scale of the internet for $200,000. Like just who needs that many tokens, you know?

• 49:40Nathan Labenz?

  It's going to be hard to like, build that many businesses competing at that level. You know, I just, I don't see how you, you know, how you build another world-class team that can, you know, they can hit at that level only to then, you know, be like, okay, cool. Now we've, now we're here. Now who needs 100 billion tokens? And we need to, by the way, we need to find a lot of people who need, you know, 100 billion tokens each to have a chance. You know, I don't know, it's tough.

• 50:06Nathan Labenz?

  The volume of that really kind of blows my mind. I think there are things, you know, you'll see like,

• 50:12Nathan Labenz?

  you know, just mass summarization, you know, just, just summarizing everything, just processing everything, you know, quality. It just, this stuff is a, it's cost effective for quality control, right? I mean, every time you've ever been on hold and said this thing is being called, being recorded for quality control purposes. Like now it is cheap enough that you could actually implement, you know, that the quality control on literally every call perhaps. So, you know, there will be, I don't think we've by any means like,

• 50:42Nathan Labenz?

  exhausted our imaginations when it comes to what are we going to use these tokens on? And 100 billion, that's input and output also, it should be noted. So like, just scanning through tons of shit, you know, and just kind of post-processing information. I think it's going to be a huge trend. There's a lot of tokens in that. But still, you know, how many, how many businesses can compete at that kind of commodity layer? When it's that cheap, it seems really hard to, you know? I don't, I don't know if I'm a VC, I'm not sure I want to invest in that. Five, we kind of talked about again as well, which is privileged access to cloud compute. You know, you can only serve as much of this as you have compute. There are starting to be compute shortages.

• 51:23Nathan Labenz?

  I've heard stuff from a friend at Google who has said, you know, demand for compute is now starting to kind of bind a little bit. They've made one of the biggest, you know, decade-long investments in computing infrastructure. And they're like, you know, starting to have to ration some stuff. It sounds like at Google. The same was true from what we heard from OpenAI in like middle of last year. They were like, you know, We have to make some choices on our product line because, you know, even with our partnership with Microsoft, we we just cannot scale our access to compute as fast as we would like.

• 52:05Nathan Labenz?

  So there were there were a couple of different, you know, things that were happening at the time, but DALL-E two blew up. And that was like, they had kind of decided, yeah, we got to kind of delay the launch of some other product to, you know, just put all the all the resources behind this. And, you know, again, these are, that's the Azure Cloud that they're building on, right? So like, your your moat is pretty apparent there, you know, where they are already hitting capacity constraints. After, you know, hundreds of billions of capital investment. I mean, that's about as self explanatory as it gets as a moat, I suppose. Like, you're gonna have a hard time accumulating anything on similar scale. And, you know, that doesn't mean you can't build a business out there, but it definitely means they have a moat.

• 52:51Nathan Labenz?

  Six, GPT-4 itself. You know, they're using GPT-4 in really interesting ways that are going to give them advantage. If you believe that there is a, and Anthropic has kind of said some similar things. By the way, that going back to five, Anthropic partnered with Google, you know, Hugging Face partnered with Amazon, like all of you. Cohere, I forget who they partnered with, but somebody. All of these, you know, kind of leading labs that don't have the compute, you know, they're all entering into strategic partnerships for it. So there's kind of a musical chairs game going there that, you know, isn't necessarily one to one, but how many, you know, preferred model providers, you know, does AWS ultimately take on? Probably not that many, I would guess, right?

• 53:34Nathan Labenz?

  Anyway, so coming back to GPT-4, if you wanted to make a case for why are the leaders going to run away with it? And widen the gap between themselves and open source. Maybe one of the best answers would be that they already have these advanced models that allow them to scale all sorts of things. Were previously really hard to scale. And OpenAI has given us a little bit of a glimpse into that with their recent interpretability publication where they.

• 54:04Nathan Labenz?

  Kind of use GPT four to look at itself and try to figure out what are the neurons doing within, actually, I think they might have been looking at GPT two. I think they were using GPT four to look at GPT two, but still, you know, kind of looking at like, you've got all these neurons, you don't know what they do. So how do you figure out what they do? Well, they basically run a bunch of text through the model, keep track of what is making, you know, each individual neuron highly activated, and then kind of pull that out and look at it in batches. In our Tiny Stories interview, which we just recorded, the, it's really apparent there. They have these small models, and when they do that process on these smaller models, the concepts actually kind of

• 54:48Nathan Labenz?

  jump off the page and are very apparent. And you can see like, oh, this one seems to be responding to like animals because it like, wait, dog and cat and, you know, bird and like, okay, like I see a pretty clear category there. Like this thing fires when there's an animal. And they showed a bunch of, you know, examples of that. As models get bigger, that stuff gets more messy, hard to figure out. You know, some of the concepts remain quite clear and interpretable, but others, you're looking at this like, okay, so this and this and this and this all causes this to, you know, activated at a high level. Like, I'm not really seeing anything here that is super coherent or like, you know, an obvious concept that I could label. But they're using GPT four to automate

• 55:32Nathan Labenz?

  that process. And, you know, just because there's even in GPT two, there's already like a billion five neurons or whatever. So or parameters different than neurons, fewer neurons. But still, there's a lot of, you know, there's there's plenty. So, you know, you just scale that. How else are you going to scale it? Right? So they, you think about that kind of thing and just the mega scale that they can apply to kind of enriching these data sets, cleaning data sets, you know, the next, they do, they were talking about like, it's not all just about scale, it's also about quality. Well, how are you going to clean your data set? Right? You're going to probably go, you know,

• 56:10Nathan Labenz?

  Crunch through with G P T four. Anthropic has kind of lent some credibility to this notion with their like leaked, I think accidentally with their leaked pitch deck. Which said something along the lines of, we think the companies that fall behind in the like twenty five, twenty six cycle, maybe never catch up. And you know, if you're like, well, what the fuck does that mean? It sounds kind of ominous. I mean, I think it is kind of ominous, but. If I had to interpret it, I would say it's that the models themselves become this engine of advantage. That if you don't have access, you know, you can't perform the next, you know, level of research at the same. You know, at the same pace.

• 56:53Nathan Labenz?

  And there, you know, you could, I do generally believe that the like. Calls for regulation are pretty sincere, but that is also maybe where things like start to kind of. Diverge and you're like, oh, man, you know, if certain things are required. You know, you like, you have to perform some, you know, exhaustive check.

• 57:13Nathan Labenz?

  And you can do it with G P T four, you know, at like, at one level of scale, whereas if you don't have G P T four, you can't do it. You know, then that becomes kind of interesting challenge. Maybe you could imagine a regulation where. You know, they are required to kind of share certain, you know, certain capability with other. Developers or something along those lines. So it's not like they can, you know, control the whole stack, but. As it stands, by the way, their terms do not allow. The all of these models that have been trained on chat G P T output. Basically, they all violate the, the OpenAI terms. So if you actually did want to go and commercialize that and you're like, oh, look how smart I am. I took this open source model, trained it on chat G P T output. Now here's my business.

• 57:57Nathan Labenz?

  They'll, they could just straight up sue you and probably win. Uh, because you, you know, just took a bunch of their now. The good follow up question there would be, what about humanity as a whole, assuming OpenAI for having taken all of our shit and running the, you know, the first. Trading process that created, you know, the model in the first place. Like, is it, how is, how can it be that they're allowed to take all of the, you know, human data? And create a model and then prohibit you from taking from their model to train a downstream model. That does seem a bit weird. And, you know, I'm not sure that that exact position is ultimately going to be tenable because. For multiple reasons, for one, it just sounds kind of insane. You may have a hard time defending it. And second, you know, if you do want to say, you know, we're not trying to like, slow down

• 58:41Nathan Labenz?

  research, but research comes to a point where it kind of depends on this like, very rapid, high-quality processing of information to try to, you know, build good reliable data sets, or, you know, have some, China has just put out these guidelines, right, that are, people are basically like, they're impossible to meet. And the reason that they, you know, they're, China does not appear to be racing into an LLM future. They may be racing into a, you know, AI for military future, you know, I don't know about that. But in terms of like, putting chatbots online, they do not view that as the space race right now, as far as I can tell. On the contrary, they're more worried that it's gonna like, talk about Tiananmen Square,

• 59:23Nathan Labenz?

  or whatever, and they don't want that. So they're like, you as a developer are responsible for your shit, you know, that, I would really recommend, there's a great Cineca podcast on this recently with a couple of guests, we can find the link, where they, you know, a couple of China scholars, like, do the reading of what the CCP has said about this, and they have issued statements. And they put these standards out there that are like, not easy to meet, like, your data has to, the data that you use to do the training, has to, you know, be like, reliable, or like, you know, have quality, I mean, they're using adjectives, right? So what does that even mean? It has to not violate anybody's, you know, intellectual

• 1:00:06Nathan Labenz?

  property claims, which, you know, that legal regime, I don't think it's sorted out in China either. It has to be, it has, your data alone has to like, meet these quality standards per their, you know, statement. And then everybody's like, well, that's impossible. How could we have web-scale data that meets those standards? And the answer is, if you own a cloud and you have GPT-4, then you can do that data cleaning next time. And they're already to a point where they could even, you know, probably just do the next model trained on like, almost pure synthetic data, by just kind of taking what is real, maybe transforming it, you know, filtering it and transforming it into something totally synthetic, taking all that synthetic

• 1:00:49Nathan Labenz?

  stuff, doing the training on that, and then being like, look, we're clear because Here's the entire data set that we trained GPT five on was all now, you know, there is still that link in the chain, right? It was all kind of made by GPT four, which was in turn made by, you know, with whatever. And eventually it does get down to like, obviously, human, you know, data was the, was the, you couldn't have got here without it. But I do see some potential for that kind of dynamic where it's like, okay, there's a new standard. You know, your data has to be squeaky clean or whatever. And it's like, huh, shit. Now there is kind of a lock in effect because nobody outside can really do that unless they have this model.

• 1:01:32Nathan Labenz?

  You know, like speed factor to kind of power that, that kind of thing. And, you know, are they going to share that? You know, as of now, they basically say no, uh, per the terms, but maybe they could be required to, maybe they could change their minds. But anyway, there's, there is some, you know, this is why GPT four is a moat. Because, you know, it, it does have qualitatively different ability that like, you know, they might even be able to use to accelerate their own work. And, you know, they don't have, as of now, nobody is forcing them to share that, to accelerate others' work. Uh, and they also, they've reduced to the logits, right?

• 1:02:13Nathan Labenz?

  They not reduced, but in the past, you could, maybe even still with the earlier models. But with like GPT three, you know, you go in and use the API. When you get that API result, they would give you not just the one token that was chosen, but up to, I think, the top five most likely tokens with the like percentage that each was assigned in that prediction step. And then under the hood, they've actually generated a number for all 50,000 plus tokens in the vocabulary. So they don't have to do any extra work to do that. They're, they're, they're doing all that work for all 50,000, you know, candidates anyway.

• 1:02:55Nathan Labenz?

  Picking one, which could be the top one, or could be like semi randomly chosen. But then they would return to you, you know, here's the top, however many choices. And that was really useful if you wanted to like study the model. It was also really useful if you wanted to train a imitator model. Cause it's way more information to say, you know, the token was the, that's one thing, but then to say, Top token was the at forty seven percent, you know, next was A at thirty two percent, then was and at, you know, nine percent, and then, you know, whatever. The you can learn a lot more from that, you know, much, much deeper level of disclosure. And they've now closed that off.

• 1:03:38Nathan Labenz?

  There is no, uh, logits returned with G P T four. So that's kind of the, you know, raising of the drawbridge a little bit. You could still, you know, get your G P T four outputs and try to train on them. But, you know, it's they've made it that much more difficult to do it than it used to be. Mode number seven. Team and talent density, you know, whatever, they're they are definitely absolutely killer team. Uh, there was just a news story where apparently, uh, Satya said to their head of research at Microsoft, how the hell did they do this with a couple hundred people? We've got all these people and like, how are they kicking our butts so much? And, you know, that's a probably complicated question to answer, but no doubt they do

• 1:04:22Nathan Labenz?

  have extreme talent there. And I've seen it in kind of every part of the organization as well. Um, you know, the the folks that we interacted with when we're in that twenty five hundred dollar a month, uh, you know, consulting engagement. You know, all very, very good. Uh, you know, the business contacts, like they know their technology in a way that you just can't sustain. I don't think if you get really, really huge, the business guys that I've talked to, they're like, they don't have to check with the team, you know, they know what's what. Uh, so I do think it's just extremely, you know, strong organization, top to bottom. Mode eight, you know, insane distribution and partnerships, you know, the customer list is

• 1:05:07Nathan Labenz?

  growing rapidly. Here's a list of customers recently announced for OpenAI. Intercom, Wix, Morgan Stanley, Shopify, Khan Academy.

• 1:05:22Nathan Labenz?

  Atlassian Zoom Brex. That was just one little thread, you know, from the, one of the business guys there. You know, they've also got a huge partnership with Bain. I think they just formed another consulting, you know, partnership with another kind of global consulting firm. You know, they're already in the door at basically all of corporate America, you know. So again, like you can chase, you know, you can try to sneak in that door before it closes behind them. But like those sales processes are in process, if not already closed, you know, and moving on toward model customization or what have you. Number nine, yeah, network effects. I mean, this one's a little bit.

• 1:06:06Nathan Labenz?

  It's definitely not as much network effects. You know, we talked about this with the Elad and Sarah a little bit back. The network effects in like social media certainly way stronger than network effects appear to be in AI. Sometimes people will ask me like, how much lock in is there? And I'm always like, you know, honestly, there's not much lock in. We, when we run stuff on OpenAI, we make an API call to OpenAI. You can, that code is a few lines, you know, even like a one liner. You could flip to Claude in two seconds. You could flip to, you know, some other model in two seconds.

• 1:06:46Nathan Labenz?

  It's really not that hard, even with the fine tuning, you know, you can fine tune on OpenAI's platform. They don't let you download your fine tuned model and walk off with it. You still have to pay, you know, you're essentially building it to rent. But once you've done that data set, you can certainly take your data set and go run it on an open source model. And that's something I've, I've considered recently with Waymark just on pure curiosity, really. We're not really trying to save money at the moment, but I'm thinking, geez, these open source models, they're kind of getting there. Maybe it would be worth just like taking our, you know, data set that we currently use on OpenAI and just running it against one of them and see how it goes. Maybe, you know, if it was comparable and we alluded to this earlier, didn't really get into it that much.

• 1:07:31Nathan Labenz?

  We have a high degree of developer control, you know, our task is super defined. It's always the same formula where we're saying, here's the video script structure that you have to follow. Here's some information about the user. Here's the user just said at runtime and your job is to like spit out a completed version of the script structure that you were provided. That's how it works. When we fine tune into that, we're not supporting chat. We're not, you know, helping you write haikus. We're not doing anything else. It's that the developer control there means like, we can be pretty confident that, you know, things can't go too far off the rails. If the language model starts to malfunction, the application just errors, you know, it doesn't have, it doesn't like attack the user or, you know,

• 1:08:14Nathan Labenz?

  the user won't even see the output because like it'll just break. So that developer control and the kind of, you know, predictability of the task is such that we could do a fine tuned model, even without like all the safety, you know, bells and whistles and niceties that we get would, you know, we can't do and, and we do get and value from OpenAI, uh, but we could kind of live without them because of the definition of the task.

• 1:08:41Nathan Labenz?

  But it's still not hard to change, you know, we, we, you can just flip around, you know, any way you want, even, even on the fine tuning side. So is there super strong network effects? Not really. Um, the biggest things are, are kind of maybe like social, you know, in the sense that everybody's kind of introduced to AI with these products and like the techniques to use them and the prompt engineering and the tools all get, you know, kind of built for OpenAI first. Usually they are built in a provider neutral way, pretty quickly, if not right out of the gate. But like, nobody doesn't support OpenAI with their first release, right? Of a new library or a framework or whatever.

• 1:09:21Nathan Labenz?

  It's always going to be OpenAI on launch, maybe OpenAI only, maybe, you know, others included too. Uh, but there is some kind of just gravity there that's like, you know, it's not, that's, there's a reason that's the last on my list of notes, but, you know, it, and it also kind of forces others into somewhat of a following position too. It's like, you know, if I have something that's working with OpenAI and then I'm thinking about exploring something else or switching. First thing I'm going to try is the exact same task. I'm just gonna literally copy and paste into the other thing. And if it doesn't work, then I'm gonna be like, oh, this kinda sucks. And if it could work, if I like re-prompt engineered it, you know, or read their prompt guide or whatever and did it their way,

• 1:10:05Nathan Labenz?

  then like, that's nice, but I don't really, you know, ideally, it's gonna work on the first time and all of those companies are gonna kind of feel that pressure. They're not gonna want to, you know, make you, I mean, you can imagine being the, you know, the CEO at a competitor, right? You're like, if you had a prompt, like, well, people need to read our prompt guide and then they'll know how to use our AI. It's like, no, they're not gonna do that. It's not gonna work like that. You have to make it easy for them. You know, if you gotta reduce that friction, so that in their effort to kind of reduce that friction, they end up, you know, in kind of a following position, I think, uh, pretty often. The, you know, the plugin architecture is something that seems like it's gonna kinda go that way.

• 1:10:48Nathan Labenz?

  And is another area where, you know, you can see potentially a, some of these things where like having the product dialed in, having the feed, you know, having the data scale, you know, to power a feedback loop. The plugin architecture, in theory, it's highly portable. But are other, so other people can adopt it. Microsoft is that they're going to adopt the same basic plugin architecture that OpenAI introduced. So probably everybody's gonna kinda have and have to be able to support that in some way. But there's a lot of ways that you could be worse at supporting that than them. And now you're just kind of trying to play their game, but you're playing from behind. It's gonna be hard to like leapfrog them at their own game. It's gonna be hard even to catch up with them at their own game.

• 1:11:34Nathan Labenz?

  And meanwhile, you're not developing your own game, right? So that's, um, I think that's gonna be tough. I think for some, for a lot of these reasons, companies like, character and inflection with their new Pi AI. I think those are notable exceptions to all of this analysis, or at least possibly, because they are trying to do something different. When you go talk to Pi, it's not like,

• 1:12:03Nathan Labenz?

  you know, it's not like a sort of Butler style, you know, what can I do for you? Here it is, you know, hope it's helpful. Uh, it's a much more. Open-ended like exploratory dialogue. And you know, that's that may emerge as a totally different lane. That is currently kind of unfilled. I don't think many people go to chat GPT for companionship. And they almost kind of discourage that in their approach, you know, it'll chat with you, but it's like it frequently reminds you that it's an AI in ways that are not like super conducive to, you know, an immersive experience.

• 1:12:42Nathan Labenz?

  Whereas these other ones, you know, they don't deceive you about being an AI. I think they're they're all like pretty defensible, you know, product designs from what I've seen so far, but they do engage you in a different way that

• 1:12:58Nathan Labenz?

  you know, they could just end up being a different product category in the end.

• 1:13:01Erik Torenberg?

  Hey, you know, we spoke to Elad and Sarah, you know, a couple of VCs and how they're approaching it in the space in terms of like what to invest in, what not to invest in. And I'm curious, given, you know, you we just ran through these moats. And that's what people are talking about when they talk about moats. It's like what what's investable and what what's not investable.

• 1:13:18Nathan Labenz?

  Okay, I think I think the sophisticated analysis right now generally agrees that like Salesforce should be fine. Because they've got a ton of shit built out and a ton of distribution and a ton of contracts and a ton of everything. And they can layer on a co-pilot before you can build, you know, a worthy rival from scratch, you know, with no matter how, you know, smart your use of GPT four. So that I think is happening, you know, and Adobe is in kind of a similar position in a way that is more on the, you know, visual side than the text, although it could be both. You know, there there's been a ton of amazing stuff, but nobody's really canceled their Adobe accounts.

• 1:14:02Nathan Labenz?

  And now Adobe is announcing their own amazing stuff. And, you know, it seems like if people have to pick, I just saw a really interesting thread from this guy named Riley on Twitter who said, you know, mid journey is awesome. It's always, you know, it's it's been totally groundbreaking. But if I was going to tell somebody what to learn today, I'd tell them to use the Adobe Photoshop because. Midjourney is kind of this like, you know, black box. You get things out of it. They they can be awesome. But like Adobe, you can really make what you want, you know, and they have both the generative layer and the like dig in and do shit layer now. And so, you know, again, they should be fine. Like it's it's going to take a long time to build up a credible rival to Adobe.

• 1:14:44Nathan Labenz?

  I would be very wary of a thesis that was like, you know, pick an incumbent and be like, we're gonna beat them because we're gonna use AI. I don't think that's going to work. They all know about AI and it's not that hard to implement. Uh, and in fact, in some ways, it's easier when you have a big platform because you also have extensive documentation. So like the existing models, you know, already have a decent sense of a lot of these, you know, mega platforms. Like you could probably go ask, and I mean, you people see even in the marketing copy thing, right? You can ask GPT, chat GPT to like write you a tweet or write you a LinkedIn post. It knows what the difference is between those in like general form and tone.

• 1:15:26Nathan Labenz?

  Uh, you know, a social media company that doesn't exist yet, it's not going to know how to do that. And the same thing is kind of true for like whatever, you know, technology you're pursuing. You know, if it's all well documented out there, like you could probably get pretty far with, you know, with the chat GPT. Uh, just off the shelf, you don't even have to do any special training.

• 1:15:47Nathan Labenz?

  And I think they will, by the way, like Salesforce is going to have their own co-pilot that's not like default, you know, GPT four, uh, you know, it'll be better than that. But I don't see any reason, you know, they've got documentation, like probably a trillion tokens of like documentation and forum posts and tickets and all that shit, you know? They, I mean, they they have plenty to work with. So. Yeah, I guess bottom line there, it seems like mostly the incumbent should be fine and mostly I would be pretty skeptical of a thesis that's like. We're going to take on this incumbent and we're going to win because we're going to use AI and that's how we're going to win. I just don't see that path for most things unless you really felt like an incumbent was like flagrantly dropping the ball.

• 1:16:29Nathan Labenz?

  You know, um, I think that the more kind of exciting or, you know, hard to predict stuff would be then what we also kind of talked about with them, which was like, what are the new things that just couldn't ever have existed before where there is no incumbent and it's a totally new market? And that, like, I guess so far we've got AI chatbot as like a new category. We don't really have that many new categories yet. We've also got like text to image, you know, is kind of a new category. So there are a couple of things that are kind of like, holy shit, you just couldn't do that before that have already come online. But I think that's, that stuff is mostly still in the future because again, like GPT four has only been out for a couple of months.

• 1:17:10Nathan Labenz?

  You know, and probably a lot of the people that come up with those like crazy ideas won't even be the, you know, the AI developers, you know, in the same way that like, you know, Uber was, was developed in a different way than, um, you know, the iPhone itself, you know, and you probably, Travis, uh, you know, probably couldn't have done the iPhone, um, but like a company like Apple, you know, could never have done what Uber did. There probably is some kind of similar dynamic. What gets built on this new platform that's just like radically different? You know, who knows? Like one thing I am keeping my eye on is this kind of improved discourse space.

• 1:17:53Nathan Labenz?

  You can even kind of connect this to crypto, which I've, you know, seldom do, but like the smart contract, right? Can you take a language model that is kind of a dispute resolver? You know, cannot like cryptographically guarantee that you are in fact getting the model that was agreed upon, you know, at the time the contract was signed and like, let it resolve stuff and like, you know, handle, you know, disputes and issue refunds accordingly or whatever, you know, kind of manage little escrow accounts. That kind of stuff I do think should happen and the cost of, you know, the possible unfairness or the like. You know, AI going against you to me seems like ultimately people just accept that because it's going to be so much cheaper and better than anything else. Like what are you going to,

• 1:18:38Nathan Labenz?

  your alternative in today's world is go to small claims court or something. Good luck with that. You know, you might as well just sign on to some sort of crypto, you know, consumer arbitration scheme and, you know, it's better than than, you know, any alternative, even if it's not perfect, or you know, even if you don't get the justice you kind of want every single time. Stuff like that, I think, will be big. I do think these like new relationship paradigms are going to be, I'm not excited about that myself. You know, I don't want an AI friend and I'm, I'm kind of,

• 1:19:11Nathan Labenz?

  it does, you know, this is like the max, you know, extent of like amusing ourselves to death potentially. So I don't know that that is ultimately healthy for individuals or society. But maybe it could be, uh, you know, maybe we just haven't seen the like final right version of it yet. You know, when we talked to Eugenia from Replica, I was like, man, the revelation here is how bad a lot of people need companionship, you know, because to get as far as she had gotten pre GPT three is like, that's a common, that's not even an AI story. That's a society story. And now you're going to inject an AI story into that society story. And it seems like it easily can go, you know, a hundred X.

• 1:19:56Nathan Labenz?

  I don't see any reason that like the percentage of people who have an AI friend does not totally explode. Uh, and maybe it even could be good. Like, you know, for a lot of people, it probably would be good. Where it starts to be a problem is where, well, who knows, it could probably start to be a problem in a lot of ways. But for me, I kind of think the last thing I want is that to crowd out my real relationships, you know, which I probably already, you know, spent too much time podcasting and thinking about AI relative to, uh, you know, what I might ought to be doing. And, uh, you know, I don't want to make that any worse with some like AI friend in my pocket. And but yet I bet that, you know, just like social media, you know, is kind of addictive, uh, not all good, not all bad, but definitely like crowds out other activities.

• 1:20:41Nathan Labenz?

  I think this is probably also kind of unavoidable. And so, you know, and when Sarah mentioned that, you know, she was like, you look at the like, usage of some of these things, like, I think she mentioned maybe both character and replica, you know, you can start to see it. So from an investment standpoint, that could be really good. I would, you know, my concern with investing in replica right now would be, you know, not, and I don't mean to cast aspersions on it, but I would be less worried that they'll be able to grow users and more, you know, like, do we have the right vision here for, you know, what it kind of impact we're going to make? And again, that's not to suggest, you know, that their vision is, is wrong, but certainly, you know, questions are probably coming up faster than they can answer them. You know, I think that was definitely a

• 1:21:26Nathan Labenz?

  takeaway from that episode, right? It's like, this was, this was all happening pretty quickly. And, you know, even as much as she has been in this game, like, it didn't seem like, you know, she was fully prepared for it. And again, not to blame her, it came up on us all pretty quickly. But she just is the one that happens to be in that seat, you know, so that's, that's a challenging seat to occupy. You know, again, from an investment standpoint, where do you go invest?

• 1:21:54Nathan Labenz?

  Google's going to be tough to beat, you know, and Microsoft and their, their distribution, you know, all these hospital systems are all a customer of Microsoft or Google already, right? They, they all get docs from one of these two places, you know, or like Epic, you know, the, again, the distribution seems like it's going to be hard to beat. You know, when I think about Neil Kosla and Quray, I'm like, I do expect them to be successful. I do think they will grow. You know, they've already also got some distribution because they've been in the game for a while. And, you know, they didn't, this was not a company that was started to like, take advantage of GPT-4. So, you know, I think they will be on a good path, at least for a while. But that probably

• 1:22:36Nathan Labenz?

  means like their past investors do well. It's still not super clear that like their, you know, next round investors do super well, because again, the prices get low. So how do you make, how do you make that much money? You know, I don't know how you make that much money. I don't know, it's just tough. You know, the prices keep dropping, prices dropping 97%. And like, the biggest corporations in the world are kind of ready to run at or near cost for the foreseeable

• 1:23:04Nathan Labenz?

  And they already have all the, you know, distribution. I don't know how much share, you know, you can really hope to get. So, you know, what's going to be transformative there, you know, is it? I I kind of sense that like. It is more like task automation for a while than it is. You know, a totally different thing. Now, you can imagine somebody that could bring forward the like. True AI doctor, no humans involved, you know. Taps into all your sensors, there could be like a totally transformative form factor. That is just next level and is like, shit, that's so much better than even like. A thorough application of GPT four through our.

• 1:23:49Nathan Labenz?

  You know, current system. Hard to kind of imagine what that looks like, but you can at least conceive of it. But when it comes to like. Reviewing the doctor notes and like, you know, making sure that everything's double checked and, you know, was there any possible drug interaction? Like that stuff seems like it happens in the current. Systems for the most part, you know, because it's really easy to integrate those APIs wherever they need to be. Integrated, so. You know, I don't know, I wouldn't. I wouldn't take a job right now as a salesperson for, you know, kind of. Uh, doc.ai that's like new and on the scene and kind of like. Look, we've, you know. Done whatever exactly, but there's an interesting, uh, there was a company that just came out. Hippo, was it Hippocratic.ai?

• 1:24:34Nathan Labenz?

  They said they've trained only on. Medical, like trusted medical information.

• 1:24:42Nathan Labenz?

  And, you know, wherever they, I'm assuming they get that out of existing, you know, medical record systems. And they they call out that, you know, other systems that are trained on like the whole internet plus, you know, maybe specialized. Have wrong shit in them. And that's, you know, one way that they wanted to differentiate themselves. That seems like it could be true, could be, could be like a real advantage, though. I don't know, you know, and Neil had a couple of interesting points where he said like. You know, if the patient says they had cereal, that probably also means they had milk and like you need that kind of general knowledge. That's probably not represented in these medical systems. You know, to. Figure out all those associations. You know, he also said very Silicon Valley, you know, if the if the person went to Burning Man, they probably inhaled a lot of dust.

• 1:25:27Nathan Labenz?

  And you aren't going to know that if you're like just trained on this medical data, I wouldn't think. The Mosaic guys also said like, almost all their customers could train on a combination of open source and proprietary. Maybe they're doing that too, but they seem to be saying they're not. So anyway, that's one big question that I have. And another one is, you know, they they take a very interesting position saying, we don't believe language models are safe enough for diagnosis. So we're focusing entirely on non-patient facing, but like back office shit, which is plenty big enough market. And, you know, again, I think they'd probably be successful. Do they dominate the space? I don't know. It feels like in the end, no.

• 1:26:07Nathan Labenz?

  I have to guess that like, the big, you know, the Microsoft epic, whatever kind of complex, ultimately still takes most of that market and it's tough to, you know, to get your little green shoot up through the hornet's nest of all the, you know, all this bullshit that you have to push through.

• 1:26:27Erik Torenberg?

  This is great. I'll I'll I'll let you go. Thanks for a great conversation and we'll talk next week. Omnikay uses generative AI to enable you to launch hundreds of thousands of ad iterations that actually work, customized across all platforms with a click of a button. I believe in Omnikay so much that I invested in it and I recommend you use it too. Use Cogrev to get a 10% discount.