Transcript
0:00 A really good first principle heuristic I have is you know... -I love a first principle heuristic. -is in a fully kind of like AI native employee should have about a GPA's worth of token consumption per year. And so the place that I always go to is how many, you know, leading edge GPUs are deployed. It's it's, you know, somewhere between 5 to 10 million right now in the US, of Blackwell's. And how many knowledge workers are there?
0:26 There's, you know, by some measures, around 100 million, let's say. And so if you think that every knowledge worker needs a GPU to actually fully automate all their tasks and have an agent running at all times, then we're massively short of that need in the future. Hello and welcome to another episode of the Odd Lots podcast. I'm Tracy Alloway and I'm Joe Weisenthal. Joe, there are still a lot of questions about the business model of AI.
0:58 Yeah, and we're not going to deal with all of them in this particular episode. But one of the biggest questions, and I think it's been on people's minds, basically ever since the the deep sea scare, back in January 2025 is this idea of open source. Yeah. Versus some of the closed frontier models. Yeah, absolutely. Right. So one of the questions, I suppose another way to frame it is will frontier or near frontier intelligence be commodified. Right. And then you can't charge any high premiums on it.
1:28 And then you get on top of these problems. This was the core anxiety of the moment. So it's like if you're selling chips then that's great. They run any model. But for the big American closed sort of companies, specifically anthropic and open air or some big questions right there. Right. You have to justify the expense of all this CapEx investment and all of that, and you want to charge that premium for a premium model. And in the meantime, we've seen a lot of Chinese open source models in particular, that seem to be getting pretty close to the frontier on some of their benchmark exercises.
2:01 A lot of people will say they're distilling the Western models and cheating, basically, but they're still getting there. Or the other accusation is that they're benchmarked, so that basically they're really good at taking a specific test as opposed to then being as good, you know, then you put them in general deployment. There's a lot of accusations about really early is good at like six months behind or a year behind. Also, are they actually being trained on Nvidia chips that are, you know, somewhere in Thailand or something, in which case, or in Malaysia, in which case, maybe the near frontier capabilities aren't as shocking?
2:38 Very many. A lot of questions about this though. So I thought for this particular episode, one interesting thing to do would be maybe talk to someone who's in the business space, and by that I mean someone who has an AI company or is using a yeah, their company, a practitioner of AI models, and talk to them about how they're thinking about open versus frontier. You really saw this after the hugging face stuff, and you saw this from the startup community specifically, which was this like, sure, maybe there are some safety issues related to cybersecurity, whatever.
3:14 But don't let this be used as a pretext to clamp down on open source. And this sort of like, you know, the deeper layer reading is like, if you're a startup and you're building on AI in some way, you like price competition, you like downward price pressure. And so, you know, it's very interesting to hear, like from the companies who are building, who are using AI to build a like, how much value can you get from like a slightly a cheaper Chinese model.
3:42 Yeah. Or when is that premium for true frontier capability really worth it. Exactly. So we do in fact have the perfect guest we're going to be speaking with Daniel Saedi, he is the co-founder over at Minerva. So Daniel, thank you so much for coming on Odd Lots. Thank you so much for having me. Why don't you go ahead and tell us your Bridgewater before Minerva, right? That is right. I was at Bridgewater for just about five and a half years.
4:05 Wow, what did you do there? I did a mix of things, at first I was actually in sales, so I was helping raise money for Pure Alpha and other hedge funds. And then the last two and a half years, I was helping, oversee our trading of financials. So banks, REITs and insurers, and then started a company called Minerva. After that, what does Minerva actually do? So we are an AI company focused on marketing. So we use models to build agents to automate everything from pre campaign planning all the way down to execution.
4:37 And what that practically means is you're a marketer. You have a bunch of repetitive workflows to actually go from, I want to sell XYZ the thing on Facebook to somebody. We automate everything that's required to actually get that campaign off the ground and monitor it in real time. And let people focus on things they actually care about, like telling stories and coming up with slogans. So, we use both open source and closed source models, in that process.
5:01 And, yeah, happy to dive into how we actually do that. This is perfect. And yeah, we've never, as Tracy mentioned, we haven't had this version of the AI conversation before, but I think it's really important because there are a lot of no fans. There are a lot of you guys, right? There are a lot of companies built. No. You know what I mean? Yeah. Building some, looking at some space in the economy and building some harness, some agent or whatever that can take like foundation models, whether it's the latest from ChatGPT or from OpenAI and anthropic or the Chinese competitors, and either use them as the intelligence that powers the thing, or you use them to, like, build out your product.
5:40 It's coding. So like you're in the market for it. Yeah, you have you have the whole menu of options and you're like really in the market for all of them potentially. Yeah. And that's exactly how we're thinking about it is, you know, I think as companies have gone from software to enable people to do workflows to AI, which is okay, I'm just going to do the whole workflow, itself for the person you're looking at it from.
6:02 Okay. How much does it cost to actually build the thing, like you said? And in which case, you know, we're still using the frontier because you want to save as much time, you want to save as much money. You honestly don't want your developers to be spending time dealing with with loops and with mistakes. So we're using cloud and we're using Astra, and we're using the rest of kind of the frontier while we're actually building the product.
6:24 And then when it comes down to something that's repeatable, where, okay, I just need to get it to a level where customers happy with the task being completed. You get into a different set of economics where you care about cost per completed task, which is a lot different than, you know, looking in the media where everyone cares about cost per million tokens, it's purely an input output thing. For us, it's all about, hey, a marketer has these sets of tasks that they're trying to do that's either, you know, creating an audience.
6:51 It could be creating a set of taglines for an advertisement. And so for us, it's how do I make sure that I get the minimum acceptable task completion for the lowest amount of costs? Because everything outside of that is our margin. Tracy, by the way, I hope our producers aren't listening to this. So I hope one of them is sat right there in the corner. So I hope our producers aren't listening to this, and hopefully they're just like going to tune out for a moment.
7:18 But anyway, yesterday I went on, TBN and they sent me the raw, the raw video file and I just out of curiosity, I was like, so they sent me this Google Drive where the video was, and the Google Drive had Gemini built into it, which I didn't realize. And so I just asked again, close your ears. I said, what was the most interesting section of this conversation? And it was like from 711, to 840. And then, I just showed you videos and can you just like, clip that out and download? The little version worked really well.
7:50 And it was like, you know, there's sort of like basic like tasks, like, works pretty well. How many tokens did you hear? I didn't I bryner was very quick. Oh can I ask a really basic now? But I only say that because we're talking about marketing. And these are the types of things that marketers might want to automate and stuff like that. Anyway, a very basic question. I said, you use open source models in the beginning, but maybe that's not accurate.
8:12 Maybe you use open. Wait, can you talk about what the difference actually is between open source versus open wait? I think the two are really used synonymously right now, to mean the same thing. And they're not the same thing. The main difference is open source means that you're disclosing all of the code, all of the data, the exact recipe that was used to train a model. And that's not what the Chinese labs are doing. What they're doing is they're releasing the weights of the model.
8:37 So essentially the end artifact. And now that's much different than OpenAI and anthropic, where I send in data to an API endpoint, I get data back. I have no idea how I got the data there. The Chinese labs are at least saying, okay, you can download this model and you can now host state yourself essentially. And that opens up a whole different set of possibilities. But at the same time, other than the papers they release, I have no idea how they actually went from, you know, slop of data in an architecture to an actual artifact bullet.
9:05 Sorry. Just to be clear, the also. So if you have an open weight model, you can run. If you have a bunch of chips and you have access to power, you can run it. Right? So it's like I could like download the clean weights. Yeah. And run it on my own computer or run a lot of get do. But also another element that I think, a lot of startups care about or include in other companies, including your former employer, Bridgewater is the ability to fine tune open weights.
9:33 John, that can you talk a little bit about? I am personally still unclear whether fine tuning for local data is actually going to be economically valuable, like some people claim. But where do you stand on that? I actually think like, this is the primary benefit to the application layer right now is the ability to fine tune on top of those openings and tune the weights to your. That's correct. Okay. And the reason that this creates the value and kind of the advantage for companies like ourselves is if you think about what actually makes these models good in the first place, is that they're trained on these hosts of data that the labs are buying.
10:11 And us as startups, we might have access to some amount of data from our customers. But the truth is, we don't have these like billions of dollars of data that that the labs have. So we have to get crafty and we have to basically say, okay, what is a way that we can start to compete with the labs or compete with technology that's made by the labs and still get the same performance. And that's where fine tuning comes into play, because it basically says, okay, look, I have a certain amount of data that's either processed via my company or in our case, it's our customers actually allow us to train on top of their data.
10:44 And what that basically allows us to do is take a model that's trained to be generalized and allow it to be specialized in a very specific task. So we do this, for everything from, you know, a user might come on to our application and say, I want to find everyone who works at OpenAI and likes football and build me an audience of those people. And now what we used to do before was you would use a OpenAI model to interpret that request, turn it into SQL, pull it from the database, show to the user it was slow.
11:13 We'd have to use a very expensive model. And so if we just use Quinn out of the box, it would do roughly as good as the OpenAI model, but not nearly as, as as well as, let's say, like Astro or Soul. But then you're paying for the fact that it's not that fast. And so what we did is we said, okay, we have a few thousand of these labeled examples where you have a very complicated query, and then you have a very, you know, right result.
11:36 You grade those and then you fine tune the model and essentially what you're doing is you're telling the model, hey, focus extensively on this set of results, try to reproduce them as well as you can. And now we have a model that's faster than the closed source models for our use case is better as in the sense that it, you know, takes complicated queries and actually produces the results that our customers want. And now we have what we call like a very small advantage now and now.
12:03 So that's what we do in our case. If you think about some of the larger application layer companies like Harvey, for example, or cognition, these guys are taking all of the data that they have that makes them, you know, differentiated in their distribution mode. And now they're saying, okay, I can approach or even do better on the tasks that my customers care about by using our data and fine tuning. Kimmie, in this case, for both of them.
12:25 And now, okay, they're able to have better results at cheaper costs. Now, you have an advantage that you can start to charge a margin for. And that's really the benefit of the open weight models for for American companies. Just to make the benefits crystal clear, can you sort of order there? It sounds like you're describing a bunch of advantages or potential advantages. So you have control over your own data. Yeah, I guess you have some technical flexibility.
12:51 Maybe you have some additional protection against like a big vendor deciding that they suddenly want to charge you a bunch more for tokens, that sort of thing. Like, can you prioritize those different? And cost is an obvious one as well. Prioritize those for us in order of importance. Yeah, I think the I think the most important thing is cost. And what you have to do for as a business is essentially do the calculation of I have to pay some upfront costs and then some operating costs.
13:19 And is that cheaper than paying OpenAI and anthropic? And in many cases it's true. In some cases it's not true. And that's that's priority number one. Number two I think is the sovereignty and control issue. And you see Karp and Satya talking a lot about this, I imagine this is going to be a huge issue in 27 as companies realize they're they're sending their data out to anthropic in OpenAI, and they're not getting anything in return for it.
13:42 They're just getting simply access to the models those companies do claim pretty seriously that like, there is I know, like, I don't I haven't looked in on this a little while. I know there was some controversy with the fable in data retention. They do claim quite strongly that they don't train on corporate client data. Yeah. And they're not. So it's like, okay, because they're known like, you know, if you're a hedge fund or whatever, you're uploading and stuff and like if they were training on it, they never do claim that, right?
14:12 Yeah. And and I think other than that there's no evidence that like that's been violated. Yeah. I don't want to be on this podcast saying there's any guidance that's been violated. I think the difference that I'm highlighting, though, that Karp talks a lot about, is you give all of this data, you get your answers back. The model doesn't get any better based on your use case. I see. And so what sovereignty gives you is the ability to say, okay, if I'm running gimme a bunch of times on my data, maybe I should be training it on my data as well so that it gets better.
14:40 And then I think third is the competition, right? And in that you're no longer beholden to a single vendor, even in the case where you're havi and you might not even be using tenant as a model to your customers, but the fact that you can go to anthropic and OpenAI and say, hey, I have this home, this homegrown model, you have to now engage in value based pricing for me because it's no longer no automation or automation.
15:04 It's very good automation or potentially very, very, very good automation in the case of of the frontier model. So it gives the, the leverage back to, to the smaller companies in a way that we didn't have in the early part of the AI boom. Once you fine tune a model, how like durable is it? And to what extent can it carry over. So you know, Kimi releases like I don't even know what version we're on anymore, but they release a new version of the model.
15:31 And do you need to like upgrade your existing one or how does that work? Yeah, I think that's the biggest decision point right now for for founders and for companies like myself, is that is part of the ongoing operational costs of fine tuning a model. It doesn't simply transfer over. I can't do I can't say, okay, listen, I, I fine tuned Kimi three, Kimi 3.1 weights are coming out this week. Actually let me just transfer it over.
15:55 And so there are folks doing research in that space. I am much more of an applied than a than an AI researcher. So I can't really give an update on on how that's actually going. But from talking to other founders, this is this is really the the problem is that, okay, I'm going to have to I can use the same data set so that part is transferable. If I build a really reliable, durable set of answers that are expertly graded, I can at least fine to it.
16:20 But to get to your durability point, I have no idea how they actually got to those weights. So whether or not it's going to converge to an optimal set of weights with the same training data set, I don't know. And so that's part of the uncertainty and the risk that you have to put into your calculation as you're making this decision. One of the things that people look at, they say look, the the open, more open source models are cheaper.
16:42 First of all, you know, when we hear open source is especially prior to the AI era, we thought like kind of free, right? You just like download the software and there's no such thing as free AI, period. I mean, there is kind of like with the Google Gemini, but by and large, like somebody paying for it, someone you have to pay for it. And it's a service and there's compute and electricity that's substantial with every use of it, etc..
17:07 Can you talk a little bit about like what pricing really looks like once you fine tuned the model and then you have to you still have to find electricity or you have to find a data center or find some service like a model. I don't know who you use or whatever, but like talk to us about you have to then be the host, right? You have to find all that. How what are the economics look like there?
17:27 Yeah, I think the, I think the biggest thing that's changed in the last year from when Deepsea came out to now is, is you have companies like model where we use fireworks. Very simple. Similar billboards in SF. Yeah, very similar concept in that if we a year ago had read the Deep Sea paper and were like, okay, we want to fine tune our own model and we want to host it, there's so much that we don't think about every day that we would have to think about, like, okay, I have this different model.
17:54 I need to make sure that it's on enough concurrent GPUs. I need to make sure that the, you know, the actual software that the GPU is running on, whether that's Sgt Lang or BLM, can actually support, you know, the amount of concurrency that I need to actually serve that. So you actually would need a team of inference engineers to be able to do this in the past. And so that's a a big reason why you haven't even seen anyone really attempt this, even though llama has been open source for two years now.
18:19 So I had only been restricted to really like research examples and large companies like Uber that that had even attempted this before. And so now what you can do is you can go to a company like fireworks or thinking machines or modal, like you said, and you can say, okay, hey, I really want to fine tune Quinn. Yeah. I want to here's my example kind of training data set that I have. In our case, it would be the audiences that marketers want and the results that the model supposed to output.
18:47 And it's a very simplified process now that's become quite commodified, where I can pay, you know, in the case of fireworks, it's I'm going to pay around 7 to $9 for per million input and million output tokens, or the output tokens in this case would be the actual changes to the model weights. And there are people out there who have have fine tuned models that perform really well for as low as $90 on their final training run.
19:11 So the economics of collapse, I think the most notable economics I've heard that are comparable. I want to release any of our own numbers right now. Is that Harvey on their tenant model, spent about $6 million in compute. To get to the absolute frontier of legal tasks. You compare that to. You know how much it costs, probably to to run the new Fabo model on. And we're off by, you know, 2 or 3 orders of magnitude.
19:35 It's fascinating. I for what it's worth, the only reason I said model is because I have fine tuned the model on model. Like that's like I but it wasn't an alum. It was, Bert, which was an open source classifier from Google that was came out in like 2019. Yeah. But like I've done the whole process and I know like I have note to your point about, look, I don't know where the differences between like 2026, 2025, it's just so easy.
20:01 Yeah. Like just sort of like and I imagine if you had a little so. Well I have no technical expertise. But imagine if you have someone with a little bit of technical expertise, it must be extraordinary how easy it is and what they can do. Yeah. It's just like, I don't know anything and I like fine tuning the model. It's crazy. Yeah. I think where the technical expertise, it's cheap. I think it's cheap to actually do the training, which is what you tend to see in the AI industry as a whole over the past three years.
20:27 There's anything that's related to the actual chips, to the actual power. You start to see lots of players emerge, and then you start to see competition, and then you start to see it not necessarily become commodified in the sense that the margins have gone down to zero, because Fireworks and Modo make quite a boatload of money, actually, in making it easy for folks, both of them are sitting at $1 billion in error, and I imagine their unit economics are actually quite good.
20:50 But the fact that I can go to anyone and actually do it, and in the way that you just described, the fact that you're able to fine tune Bert is a testament to what's going on in the space, and that things that can be commodified are being commodified, the things that aren't commodified yet. And what actually separates a good fine tune from a bad fine tune is the data and the actual evals that are put in there.
21:10 And that's why you have these companies like McCaw and, and Serge. And then, you know, the the reason that the labs are paying them billions of dollars is the alpha per se is, as you would say, is in the data right now, and it's structuring those things well. And so that's a real decision point for a company like ours, is do you have enough good validated data where the model can actually learn from it? Or are you just going to get noise?
21:34 This was going to be my next question. But talk to us about acquiring data now is how difficult it is to everyone always talks about, you know, original data sets and things like that. How difficult are those to find? Oh, so I think in our use case, it's, it's we've tried to go into an industry where the data is quite plentiful and marketing and that, you know, you have data for that. You know, companies are sitting on call logs, are sitting on advertising data, they're sitting on ads.
22:00 And the actual creative process that was used to make those ads, you have data on, you know, website activity and where are your users actually spending time? And all of that can kind of be constructed into what, you know, the market is now calling an RL environment, which you an RL environment is simply, you can think of it as like a simulation where you're trying to replicate the environment in which an AI is actually doing a task.
22:23 So in our case, it's the actual act of acquiring a customer. And so the simulation that you're trying to create is, you know, I show an ad to an AI, does it actually get the AI to click on the link to go to the website to actually buy the thing? And that's what your data set is supposed to represent. And so for us, it's actually quite easy because quite plentiful. And folks have been collecting data in the marketing space for, for a long time because, you know, we had all of those AI optimized kind of, algorithms back in the, you know, programmatic days 15, 20 years ago.
22:53 And ever since, companies have done a good job of keeping data there for more specialized tasks from from folks I talked to. Let's take like construction, for example. It's like, man, it's it's really, really hard to find good data sets. Okay. I have the blueprints for this building, but what are all the comments between the architects and the general contractor and the actual client that led to the changes in the blueprint? And so that's what the AI actually needs to do to generalize to do these tasks.
23:20 It doesn't need results. You know, your teachers used to tell you, I don't care if you got the right answer. I need you to show your work. Yeah, that's what we need for the AI and so it's really hard to find that data, I think as a, as a way to see what the labs care about. You can go to these companies websites and see what they're paying experts in the space. So if you go to the Micro One website, you'll see which one.
23:40 It's called Micro One. Okay. Yeah. Micro one is, the big procure of company data right now. And so I'm, I'm sure you guys are on Twitter and you saw heard of it. Yeah. You saw a few weeks ago everyone was talking about, you know, data brokering. And I'm making $100 million a year selling data to the labs. Micro one is probably the most legitimate company in the in the company data space. And so for a while, they're spending anywhere from 250 to 500 K per complete data set from a company.
24:08 And they were paying people 50 to $100,000 just for referring the companies to them. So that kind of tells you how in demand good company data is. Yeah. I mean, I also do is that, you know, Google bought Spirit Airlines for $10 million, and then you just get this treasure trove of corporate emails and calendar data. The only thing worth saving from spirit was the data. It's a little bit disturbing. You got to wonder, how many zombie firms are out there that would be worth more if they just liquidated and sold all of their emails and corporate data to a lab?
24:41 Who I mean, who knows? How much can we get for all blowouts? Transcripts? Priceless is priceless. One of the concerns that you sometimes hear with open source models is cybersecurity. The idea that, like, okay, well, maybe you don't know the exact weights, but people can like look at the source code or I don't really understand it, but there's a vulnerability there that you don't necessarily find with the frontier models. How are you thinking about that? Yeah, I in preparation for this episode, one of the things that I did was actually play around with, the newest deep sea model, and it's uncensored form.
25:17 And, and that's really the biggest risk, I think, personally, to having open, open weight frontier models out in the wild. And that basically what somebody did to the deep Seek model that was released is they turned off the neurons that tell the model to say no. And so, oh, this is great. Yeah, yeah. This is what's called an obliteration grotto. It's such a good word. It's a great word. It's a great word. Iterated model. Yeah. It's come on, an obliterated model.
25:42 I mean, it's pretty crazy playing around with it in that, you know, the model is probably not as capable as mythos back in February of this year, but it's up there and it's really cheap to run. It's about $0.30 per million or blended. So input output tokens. So it's about, you know, ten cheaper than sonnet which is, you know, about ten cheaper than fable. So it's things like 1/100 of the cost of able to run and you could tell to do anything.
26:10 And so if you see all these hacks that are going on right now, the, the theory on the internet is people are running swarms of these deep seek obliterated models. And that's really the risk that the the labs have talked about, which is, hey, there's actually like no risk to the Chinese getting your data from this. Let's just quell that kind of, notion. Now, the the risk that they're concerned about is, is not data going to the labs.
26:32 It's not, you know, vulnerabilities from the actual models. It's the fact that I can just turn off a few parameters in this model. And now suddenly I have this PhD level, you know, Elliot Alderson, if you watched Mr. Robot, I have a thousand of these Mr. Robots now running in the cloud, and I can tell them to hack anything. And that's the real risk. And what's scary. And we're seeing it in real time. And so I don't know what the right policy, is to, to, to have here whether that's, you know, you have to KYC before you can download the weights, but we have to reach some sort of medium where, you know, the, the kind of, dynamism of the U.S startup community is preserved and that companies like ours can still download the weights and fine tune like we talked about.
27:17 But at the same time, you don't have hacker and terrorist groups being able to just download a deep Sik model, you know, make two changes to it. And now suddenly they have this like unadulterated, obliterated. But you don't have that. Do you think you do or you don't have it? I'm saying you have to. You have to create a world where, yeah, where people can't you can't have the terrorists just download. No, but. And this seems like a pretty serious issue.
27:39 Yeah. Like, these are really powerful tools, weapons. And this is part of what makes a very tricky is it's dual use. Right. Because in theory maybe it's going to be. But like these are weapons when used for offensive hacking purposes etc.. And you can obliterate them and you can get them to do whatever you want. Yep. And that's super not good. No, I, I agree with you. It seems not good. I, I just have yet to see like a really compelling proposal for how do you.
28:08 Yeah. How do you deal with it? Because, you know, I want to be able to download the way. So I actually totally understand this point about wanting to fine tune, Lagging edge, not quite frontier models and being able to then get a model that's as performant is like what the big tech companies have. Yeah. Cetera. And that also like, makes a lot of sense, but how much does the fact that you're able to do this subsidized by the fact by the distillation question.
28:40 In other words, it's like great. We take open source plus our data and then we have something good. But if the open source couldn't even exist in the first place to be like a year or six months behind the frontier, were it not for terms of service violations? Yeah. Isn't it pretty understandable because the issue is like, why didn't Lamo go anywhere? Presumably one reason is they're an American company that is not going to go around violating the terms of service of anthropic. And, OpenAI, right? Yep.
29:13 That's you know, you're like building something out and like Shenzen brilliant engineers, I'm sure. But I don't think they care about terms of service violations. I don't think anything bad is going to happen to them if they, like, find a way to do 100,000, you know, and clog queries surreptitiously. No, I, I, I completely agree with you. And I think, I think there's, there's a need to basically say how much of the gains do you think are actually coming from distillation?
29:41 I think we actually have a good barometer on this and that. There's a company that used to be, headquartered in the floor above me called reflection AI. Allegedly, it's being bought by, Nvidia, and they're probably the preeminent US open, open weight, model company right now. They released their flagship model about a month ago, and it was probably about 3 to 6 months behind the Chinese open source models, which were about, at the time, three months behind the publicly released frontier.
30:11 And same with this company. It's called reflection. I keep going, and so what that basically tells you is that, you know, that company has a bunch of backing from Nvidia. They have, thousands of black, well, GPUs that they train this model on, I believe was a cluster of 12 or 14,000. And so they have access to compute. They've got access to great talent. And then they released this model and it's about like 6 to 9 months behind the frontier.
30:33 And so what that basically tells you is the distillation is not the pure reason that the Chinese labs are where they are. They have lots of evidence of coming up with very novel techniques that are actually being used by the Frontier Labs. So how is that evidence? Because I assume that an American company owned by Nvidia is going to really play it by the book with regards to, distillation. Yeah, yeah. What I'm saying is that if the Chinese companies that played by the books, you can assume that they would probably be neck and neck with reflection, if not maybe a little bit further ahead because they have the backing of the Chinese state.
31:07 And you're saying and so what I'm saying is like distillation is real, distillation is is real. And there's lots of, of evidence for it. How much distillation is responsible for the quality of these models, I think is really up for debate. I think one thing that if you ask actual, you know, researchers in San Francisco, how do they feel about the Chinese labs? Obviously, they'll talk about distillation. They'll also talk about the like, absolute prolific, publishing nature of researchers there.
31:36 Yeah. And some of the advancements, whether that's using sparse models instead of dense models, which saves a lot on compute, from, deep sea, actually just publishing their entire inference framework a few weeks ago and how that's being used by fireworks and modal and all these other companies to to improve the efficiency of serving these chips. These guys are doing a public service on the other side. So I think the fact that the quality of their research is still being respected by the labs shows you that, hey, you know, they're still making good models.
32:05 Is distillation responsible for all of the reason why they're good models? Probably not, but it's actually helping them out for sure, especially when you see them. The similarity in performance between Kimi and Astra and there's big allegations there. But yeah, I don't I wouldn't say we're subsidized by distillation to answer your question, what in your mind would be indicative of the Chinese models, like actually catching up with frontier? Like what would be, I guess, the next deep Sook moment in your mind?
32:33 Because I assume it's not just the benchmark test, right? Because, as Jo was saying in the intro, those can be gamed and they're pretty close on those anyway. Yeah. I mean, this is such a good question because I think it's a it's actually a much broader question than just are the Chinese models catching up to the frontier models? It's really how do you know that there's actual progress that's being made relative to where we are right now?
32:55 Because if you go and you know, you spend time with ash or you spend time with fable, it's really smart. It's still makes silly mistakes. It still sounds like slop. But these models are, you know, compared to ChatGPT, you know, when it first came out, they're night and day in terms of their capabilities. And it's and for everyday use, even for me, it's like it's really hard for me even to tell the difference between a 5.1 and a five.
33:18 And therefore, how am I supposed to be able to tell the difference between a five and a Kimi 3.1? So I think where we're moving towards in a world where companies care a lot more about sovereign data, and they care a lot more about the the actual tasks that they're doing within their firm, rather than just benchmarks. It's going to come down to companies themselves, like ourselves or workers who are doing tasks actually seeing results on a day to day basis of, okay, can I just swap Kimi into an existing process that I have in a chat bot that I'm serving to customers or in my coding workflows, and, and I actually seeing similar performance because the benchmarks are starting to become meaningless, as you guys said, as as Joe said, the benchmark thing, you know, if I see that Kimi 3.2 as a higher terminal bench score than Astra 6.2, what does that actually mean to me?
34:06 Does it mean anything? Probably not. But if I can take Kimi 3.2 and start replacing it as my daily driver and how I'm coding? Okay, now I know that things have caught up and how do I know that it's replace it? Am I, you know, yelling expletives into the terminal anymore if I'm not? Probably not. Let's talk a little bit more about the economics of, from the from a from a company like yours. So first of all, it occurs to me like, one way you might be able to characterize what you're doing now is using frontier closed source for CapEx building and using cheap open source for CapEx intelligence.
34:40 Right. Is that a fair characterization? Yep. Would it is it the hope? You know, token price disinflation has been massive every year. Like isn't the sort of vision that there is something like, okay, let's take a task like, you know, AB test a bunch of different ads or something like that. It's very going to look the same is the hope that, like, you will be able to sell your service at some sort of like recurring price, while the token price collapses 90% year after year after year and like, capture that margin.
35:16 Yeah, I mean, that's exactly how I think us and other companies are thinking about it. It's first, what am I charging my customer? That's the number one question, right? I think Harvey and Lagos are two great examples of this. Lagos is another legal tech company. They also are now exploring fine tuning their models in response to Harvey. But Harvey has a per seat cost which is fixed and Lagos charges usage. And if you look, Harvey actually went down the legal lab route first.
35:47 And if you think about it for a second, it makes a lot of sense. They have fixed revenue coming in. They saw that the costs were rising because usage was going up of their product as it got better. And so naturally, when you have a fixed revenue coming in and a variable cost, you want to reduce that variable cost as much as you can. So in Harvey's case, it's let's come up with tenant. And now whether or not tenant is tenant tenants, their legal.
36:10 Oh no. Oh that's they're fine tuned okay. And so whether or not you want to serve tenant to users or you simply want to use tenant to get better leverage from the labs, it's coming from a place of I need to increase my operating margins. And that's exactly the the question that, that, that Harvey's asking. It's a question that we're asking as, as well. It's where are the, you know, highest kind of variable components of span within our product that aren't tied to usage.
36:37 And how do we then replace that with either a fine tuned or just purely we might even just say for some use cases, if it's a highly repetitive tasks, that's quite simple. We might just use a very pure open source model. And that's another question too. And so the economics really boil down to the equation of okay, I have some kind of per token costs and I'm paying to the labs times the number of tasks that's required.
37:01 And that's got to be more than a fixed costs upfront to train the model and the operating costs, of, of, of actually running that model on a, on a monthly basis has your token budget actually come down at all, or is it still, I guess, too early? You're still in the building phase? I would say our our token budget has gone up, but that's mainly because we've just gotten more money over time. When the we started the company in 2023 and we were being, you know, very, deliberate about when we were using AI back then because also it just wasn't that good.
37:36 And so we didn't all shift to agent coding until probably when opus 4.5 came out in October of last year. And so we ramped up and and as such, we're still ramping is we're finding that the models get better and we want to, you know, keep the team lean. But we've always been you know, we're not spending like ten, 15, $20,000 per engineer for no token maxing. We're not token match. So now we are output max or outcome max or something.
38:03 We're actually recording this, as everyone knows, during quite a bit of volatility in the AI trade, etc.. One thing that hasn't gone down, and I think this pretty important is for any measure of GPU rental prices. Yeah, it looks like. So I'm looking at the silicon data, each 100 rental track index. This is an old chip. It's and it's 100. It's still going up instead I don't know instead the highest in a while. Are you. What are you seeing though.
38:34 You know, this is like an index and, you know, whatever. What are you actually seeing in terms of compute scarcity, rental costs, things like that? Yeah. I mean, I think from this is more a question for, for friends. It is for me. We're we're pretty light in the amount of compute that we need to fine tune our models and we find availability to, to always be there. Actually, I think when things were really tight was two years ago, we were actually dabbling in training at, our own model ourselves and the, the H1.
39:04 It was right around the time the H1 hundreds were just unavailable. Completely to all users. So I think there is slack where you need to find it. I think there's other people that you could talk to, and so can that data to kind of explain the, the what's going on with the longer compute off takes, lock ups and why those are so cheaper than the spot market. But for us, things are pretty available. I think if you go on Twitter right now, it looks like, man, there's just no GPUs anywhere.
39:31 And well, I can't tell because I don't I feel like the narrative on this changes very quickly. Yeah. It's like you go, it's like, oh, there's is so short compute, we're going to need all this stuff. And then on the other hand, in the last several weeks, Overbuild is like back in the thing and you see this, you know, and I mean, it gets there's all kinds of complicated things. And you could theoretically have temporary overbuild, long term massive shortage.
39:59 But I don't know, like the reason I asked was like, I don't even have a sense right now. Like, what is the consensus narrative in terms of compute capacity? I think the thing that I always just like to keep in mind is, a really good first principle heuristic I have is you know... I love a first principle heuristic. -is in a fully kind of like AI native employee should have about a GPUs worth of token consumption per year.
40:23 And so the place that I always go to is how many, you know, leading edge GPUs are deployed. It's it's, you know, somewhere between 5 to 10 million right now in the US, of Blackwell's. And how many knowledge workers are there? There's, you know, by some measures, around 100 million, let's say. And so if you think that every knowledge worker needs a GPU to actually fully automate all of their tasks and have an agent running at all times, then we're massively short of that need in the future.
40:52 And so is everyone going to be as I forward as, say, one of my engineers or an engineer working at anthropic? I don't know, but roughly that's kind of the heuristic where I think things are going to net out four years from now. It won't be a surprise that we have kind of one be 300 worth of computational capacity for every, knowledge working employee in the economy. So in my mind, it's like if everyone's going to be using AI in the way that I do it, we're still massively short compute.
41:18 We're going to have these like, you know, drifts around what the baseline should be right now. And I think we feel that on Twitter because the people who are you know, most short computer, sometimes the loudest people right now. And so we hear them screaming and then sometimes you'll hear about a build out. But in the long term we're still massively short the amount of compute we need to actually be an AI forward society and labor force.
41:39 Yeah, I feel like if you have loads of compute nowadays, you keep it quiet, right? I maybe, maybe the temptation is to brag about it. I don't know, I mean, I think the most telling thing about how short compute we are is allegedly when all of the, you know, Elon and Jensen were waiting to receive their medals from the white House. Apparently, Greg Brockman asked Elon for compute there. So, if they can kind of settle their beef by trying to do a compute trade, that probably tells you that, at least the labs are still fundamentally short compute relative to their needs.
42:09 All right. Daniel Saedi, thank you so much for coming on Odd Lots. Thank you so much for having me. Joe I thought that was a really useful conversation to hear how, you know, someone in the business is actually deploying these models and thinking about closed versus open. Yeah, no, I think that was really helpful. And I think this sort of gets to show that there are a lot of dimensions that are just not that straightforward, right?
42:31 So it's like what is actually cheap? Well, if you're engineers, time is valuable. And scarce, then it may not be cheap to use a lagging edge model because, you know, like you just want like so then yeah, like pay the big anthropic premium or pay the pay. ChatGPT premium, whatever it is for something that's just like repeatable etc. like, oh, here is our harness that does like AB testing from two images on an Instagram and then like, yeah, why do you need the frontier intelligence?
43:02 And I think a lot of the big questions are going to come down to is AI usage overall going to be more like CapEx or opex? Yeah. And you know, it's probably going to almost obviously sums for both. Yeah, yeah. But you know, I would place my bets on the cheaper models winning out. But I don't know. One thing we didn't get into and when you mentioned cheaper models is also that the closed source labs do have cheaper versions.
43:29 Yeah that's true. Well, so but I agree, like at any level presumably I think the, the Chinese open ones are cheaper. Have you been asking. It's it's an interesting exercise right now if you can get, you know, someone at a company, a large like U.S company and ask them what models they're actually using. Yeah. Privately. Yeah. A lot of them we'll talk about like Quinn and. Yeah, but not in public. Yeah. It's kind of interesting. Every once in a while you do see it like but usually around like what they were able to accomplish with fine tuning etc..
44:03 Yeah. I also think it's interesting. I also think there's a kind of flip side of observe the opposite, which is I think there is this sort of like posturing to the anthropic. And OpenAI is the world, like, we don't need you, you know what I'm saying? Like, yeah, that makes sense. Yeah. We found a way to like, get, you know, for example, Brian Armstrong, the CEO of Coinbase, has talked about, I forget Kim, they're one of the ones that they use, etc..
44:31 And it's sort of like this shot across the bow. I would bet they still have. You know, I don't know for sure. You know, I think a lot of companies are paying an ant tax or an open AI tax, but, yeah. No, there's some really interesting questions in there. In the meantime, we have a new, I worry, iteration. Oh, yeah, I gotta get I've never. I got to, like, get my hands on it. Obliterated model I gotta try, I have stuff. No, no, I've always been just, like.
44:58 Sounds so evil when you say it like that. You need to be, like, stroking a cat. Yeah, yeah, well, you said why while you talk about your obliterated model. But I've thought about this, like, look, there are just certain things. I don't even want to bring them up because they're all very negative that you like, you think about like how well lock down the internet has become for certain types of information that we want it. We really like do not want to be readily available in the public's hands.
45:26 And, you know, for example, it is very difficult to use Google if you wanted to or to traditional search to find out any information on, say, like suicide or something like that, or, you know, there's certain things how to build a bomb. Yeah, yeah. Like there are things that are just become almost impossible to find on the internet, etc. in theory, that all exists in obliterated models. It almost certainly does that knowledge obliterated models aren't easy to come by.
45:56 There aren't a lot of entities hosting them, but it would be really strange and novel if that were to become a very easily accessible thing, because there has been this sort of firewall built around certain types of information that maybe, perhaps won't hold that long. Well, once we have AGI, we can get obliterated. Models developing new models. Did you see operating on themselves? Well, for real, so did you see. So today we're recording this November 3rd, there was this research report put out by someone at MIT.
46:27 And what they did is they took after six and they, they fed it quinze weights, just a big giant pool of numbers. And they said change the weights to make it better at math. And so like typically weights get to where they are through, you know, a training process back and forth. And Astor was actually able to go in and manipulate certain numbers. No, it was not quite as impressive. It was sort of like fixed a bug.
46:59 But there are certain we may be getting to a point. And this is why everyone was saying we're like, models are so good and so much like intelligence possessed in them that rather than doing the traditional way of growing a model where you like, do the reinforcement learning and all of the gradient descent and all that stuff, that the models can just go and the models can iterate on themselves. Yeah, it's pretty wild stuff. Isn't this, the entire starting premise of, if we build it, everyone will.
47:27 Yeah, yeah. Well, and the other element is so. Yeah, I mean, there's a lot of rabbit holes you can go down, but, you know, there is this question of like, you know, if you just have this really amazing model, what is the best value you can get out of it? It might just be working on AI models and that's the and then everything else is a waste of its time. Basically like if this is not being deployed towards building more AI, you should have just been using some, cheaper lagging model.
47:55 It's AI all the way. Yeah. That's right. All right. So we leave it there. Let's leave it there. This has been another episode of the Odd Lots podcast. I'm Tracy Alloway, you can follow me @tracyalloway. And I’m Joe Weisenthal, you can follow me @thestalwart. Follow our guest Dan Saedi. He's @TheRealDanSaedi. Follow our producers Carmen Rodriguez @carmenarmen, Dashiell Bennett @dashbot. Cale Brooks @calebrooks and Kevin Lozano @kevlloydlozano. And of course you can chat about all of these topics 24/7 in our discord.
48:24 Discord.gg/oddlots. And for more Odd Lots content, you should check out our daily newsletter, which you can find at bloomberg.com/oddlots. And if you enjoyed this conversation, if you like it when we talk about open versus frontier models, then please leave us a positive review on your favorite podcast platform. And if you're watching on video then please like or leave a comment. Or better yet, subscribe. And remember, if you're a Bloomberg subscriber, you can listen to all of our episodes absolutely ad free.
48:53 All you need to do is find the Bloomberg Channel on Apple Podcasts and follow the instructions there. Thanks for joining.