← The Peel with Turner Novak
The Peel with Turner Novak
The State of AI: Rise of Reasoning, Surge in Chinese Open Source, Sovereign AI , How to Invest in AI Today | Nathan Benaich, Air Street Capital
02:23:22
Loading summary
Transcript295 lines
- [00:03]
A
Welcome to the Peel. I'm your host Turner Novak, founder of Banana Capital. Today's guest is Nathan Benesh, founder of Air Street Capital and author of the State of AI Report. Nathan has been writing the State of AI for eight years. It's a year long effort on the biggest things happening in AI every year across research, industry, politics and safety. We spend the next two hours talking about the biggest takeaways from his latest report, including the rise of reasoning, the Surgeon in China's open source models where AI is working in practice, the rise of sovereign AI where he thinks value will actually accrue over the long term if we're in an AI bubble or not, and how he's investing today at Air Street. A quick thank you to Nico at Adjacent and Dan at the University of Michigan for helping brainstorm topics for Nathan. A reminder. I publish new episodes of the Peel every week, exploring the world's greatest startup stories just like this one. Check out the back catalog of over 100 episodes, including recent conversations with Barcelo Lebra, co founder of European Unicorn Remote, and Kevin Hartz, co founder of Eventbrite and seed investor in PayPal. Tune in over the next few weeks for guests like Gary Tan at yc, Chathan Putagonta at Benchmark, Jake Stoch at Servol, and Duo Security co founders Doug Song and John Oberheide. Let's talk to Nathan after a quick word from Numerl and Flex. This episode is brought to you by numerl. Numerl is the fastest, easiest way to stay compliant with US Sales tax and global vat. It's easy to set up and and they automatically handle all registrations, ongoing filings, and their API provides sales tax rates wherever you need them with all the integrations you need. Numero sports over 2,000 customers in both the US and globally and they pride themselves on White Glove High Touch customer service. Plus they guarantee their work and they'll cover the difference if they mess anything up. They're fresh off a Fundraise closing a $35 million Series B from Mayfield, which they're going to reinvest into building an even better product. If you want to put your sales tax on AutoPilot, check out numerl at their new domain numerl.com that's n u m e R-A-L.com for the end to end platform for sales tax and VAT compliance. This episode is brought to you by Flex. It's the AI native private bank for business owners. I use Flex personally and I love it because they use AI to underwrite the cash flow of your business. Keep giving you a real credit line. The best part is 60 days afloat, double the industry standard. Flex has all the features you'd expect from a modern financial platform like unlimited cards, expense management, bill pay that syncs with your credit line and their new consumer card, Flex Elite. FlexElite is a brand new ramp like experience for your personal life. A credit card with points, premium perks, concierge services, personal banking, cars and expense management for your family network, tracking across public and private assets, and a whole lot more fully integrated with your business spend. One card for your businesses, one card for your personal life, one card for everything. To skip the waitlist, head to Flex 1 and use my code turner to get an additional 100,000 points worth $1,000. After spending your first $10,000 with FlexElite, that's Flex 1 and code turner for $1,000 on your first $10,000 of spending. Thank you, Flex. And now let's jump in. Nathan, welcome to the show.
- [03:17]
B
Thanks for having me. Turf.
- [03:18]
A
Yeah, thanks for, thanks for coming on. So you put out this really interesting report. It's called the State of AI. You've been doing it for a while. I'll let you kind of explain how it all got started and I think I gave people a little bit of context on kind of what we're going to talk about, but can you just kind of talk about this report that you put together, kind of how it got started, why you do it, all that stuff for sure.
- [03:39]
B
So the State of AR report is an annual production. It's an open access document that I create in order to kind of disseminate like interest, the most interesting analysis across research, industry, politics and safety. And then we wager a couple of predictions every year in order to sort of cast things forward and see how we did the year after. The real goal is just to help people stay abreast of what's going on. There's just like so much stuff and you kind of don't know what is, what is meaningful. And most important, it also acts as like a good litmus test I think, to see insanity check to see like hey, have we over exceeded on progress estimations or under exceeded or what are, you know, just how far have we come in the last couple of years? It's been running for about eight years now. Started in 2018 with me and Ian Hogarth who produced it together for a while and then been solar production for the last couple of years. And it's a good opportunity Also to collaborate with people in the ecosystem and showcase good sort of example case studies of how businesses are using AI and how certain papers are leading through breakthroughs and what might be head fixed.
- [04:59]
A
Yeah, I feel like there's a lot of data, a lot of numbers and charts and a lot of visuals.
- [05:04]
B
Yeah, it's definitely not for the faint hearted. I do try to write it in such a way that you could consume the headlines of every slide and then get a general sense of where things are going and then you can stop where your eyes get most transfixed. And so it's a bit more of like a buffet and must read end to end.
- [05:25]
A
How much time would you estimate that you put into this thing because you do it once a year? How many hours or I don't know, total days or however you want to quantify this?
- [05:35]
B
It's a bit hard to give a clear answer on that because it is the result of just everyday consuming of research, news, talking to companies. It's very much the result of my day job. But when it comes to genuinely producing slides, it's from early August until September
- [05:57]
A
and then you usually put out beginning of October it looks like.
- [05:59]
B
Yeah, yeah, we found like a good tempo around beginning of October. So like back to school season, the
- [06:04]
A
report you put out in October, what were sort of the biggest takeaways? What we'll do is we'll throw a link in the description for people who want to actually like look at the whole thing. But if you're just like, ah, give me the, I don't know, the spark notes, like quick version. What are kind of like the biggest things that have kind of been happening that you think people should know about?
- [06:22]
B
Yeah, to me the probably four biggest things are one on research, which is clearly the move away from models consuming an input and just rapidly producing an output. But now we're going into like fairly complicated reasoning and tool calling. That wasn't the case even a year ago. It looks like, you know, you look at AI today and you're like, why wouldn't it be anything else than this? But this, this wasn't the, the tabularizer a year ago.
- [06:49]
A
For somebody who's never heard of tool calling before or reasoning like what's like what does that mean practically?
- [06:55]
B
Yeah, yeah, so we're maybe a little bit more than a year ago was, you know, a model had consumed basically the entirety of the Internet, you know, people like to say and then maybe some custom databases that has essentially like compressed and memorized all of that information and it's basically like in sort of simplified terms, it can be thought of as like an API to all of that knowledge. And so when you input a query, it's kind of looking up its knowledge and then producing an output. It did not have access to the Internet. And then like the first sort of tool call was do a web search for for example, more relevant information. Because at the time what was very important were these like knowledge knowledge cutoff dates, which is basically like the, the date at which like the download of the Internet was made. And so, you know, if something material happened after the cutoff date, it wouldn't be stored in the model because it didn't have access to real time web search.
- [07:51]
A
You'd ask like, who's the President of the United States? And it would say Joe Biden. It would be like, no, he's not exactly.
- [07:57]
B
Or who is your creator?
- [07:59]
A
What, what does it show for those Usually if you search that.
- [08:02]
B
Well, there are some interesting ones. I mean, you know, OpenAI would say, you know, OpenAI and anthropic largely the same. But the Chinese models would say weird stuff. Like they would sometimes say like, oh, my creator is OpenAI or my creator is this other company. And then, you know, that led to question marks on the Twitter sphere of, you know, are they training on outputs of American large models? You know, it's what's called like distillation. Basically There were some FT headlines about this. You know, OpenAI were saying that this was the case and that was one of the reasons, in a corner copy of other reasons why the US Administration wanted to cut off Chinese access to American AI and things like that. So it's unclear if it's actually schizophrenia or because there's just a lot of examples of AI's output that's generated by the leading labs on the Internet. And so statistically it's more common to have that memorization.
- [09:00]
A
Yeah. So essentially what this means is that it incorporates just the models with what's on the Internet so the models can learn and take in new data versus just what's been trained on and captured in there and it's set forever. So these things can actually get smarter in a way and use current relevant news and information.
- [09:19]
B
Yeah, I think it's even more than just being able to consume current news. It's like the designers of the system would ideally separate memory like storage of facts from the ability to know how to retrieve them and how to like logically reason through answering a query. And so ideally you would want to learn the latter capabilities and not have to memorize facts in the weights of a neural network because then how do you selectively update them as information changes? And so yeah, the big push towards like tool calling and you know, web search is a tool, you know, using an API could be a tool, using a software product, could be a tool, is really to combine that with step by step reasoning. Where human annotators have explained in a very, very clear way, a structured way, how they would go about answering the question of how should I implement this stock trade, given this is my goal and this is the ticker and these are my financial goals, et cetera, and anything you can really imagine. So if you repeat that enough times and you really learn the logic of reasoning, then that logic can be abstracted away into many, many other areas and it's more repeatable than like just memorizing facts and memorizing like patterns.
- [10:41]
A
So how is this kind of shown up if people are using a lot of AI products? Like where is this capability kind of shown up in what we're all using today?
- [10:49]
B
The most obvious one is just when you enter a query like in chatgpt, it'll say thinking and then it'll produce a result. And then you can click on the thinking tab and then, and then you'll be shown sort of simplified stepwise list of. Okay, I think the user is querying this. Oh, maybe they're not. Oh, maybe I should go like disambiguate what they're asking. And then this is the stepwise plan to get to the, to get to the answer. And I think it's interesting because it like gives you like a vignette into what the model is, is doing to answer your question. So you can build trust in. If the reasoning trace looks like what you would have done, you can almost
- [11:32]
A
like troubleshoot it in a sense too.
- [11:34]
B
Yeah. But then there's some other research around. Are these reasoning traces, otherwise called chain of thoughts, actually truthful? Like is the model actually doing that when it's telling you it's doing that,
- [11:47]
A
just making it up or is it
- [11:49]
B
actually just making it up? Interesting. Yeah, because, because it's, it's, it's. Yeah, because in some ways it's like trained to, to recapitulate like human behaviors of knowledge traces. And so it might be producing a result, but like showing you what the human would have done because that's what a human would rate as a good reasoning trace.
- [12:08]
A
So we're like, they're like learned how to lie maybe like that's. Or hide their hallucinations.
- [12:15]
B
Yeah, there is Some of that paper wise that's come out and I mean to some degree it really depends how you measure these, how you measure that behavior. And maybe your measurement is wrong and therefore you're like you're drawing the wrong conclusion.
- [12:32]
A
Interesting. So you said that was one of the four. Is that probably like the biggest thing that's kind of happened?
- [12:38]
B
I think so, yeah. Because that's like unlocking computer use, which is like the model can interface with your computer and do a bunch of things with it. It's unlocking more complex scientific and maths discoveries which the AI world is very excited about. Yeah. And then I say the second big thing is the rise of Chinese open source. I mean it's only 12 months ago that deep SEQ moment happened and it was probably six months before that when Deepseek, the Chinese company had released like early versions of its model, this base model that was then tuned to do reasoning and resulted in this like deep seq freak out. And since then, you know, many more models have reached, you know, leaderboard headlines, in particular in world modeling and in vision. So basically anything to do with like making pictures or videos, long form videos, Chinese systems are very good at that. And then, and then the like resurgence of, of Quand notably, which is Alibaba's system that effectively like took the mantle from Meta Islama initiatives. Or just a year ago, you know, the tech world was applauding Zuck and others for, at Meta for you know, stepping into the fold and saying we will always do open source. It's good. You know, we've done this with Pytorch and, and other of our core technologies. And then, and then of course like big vibe shift there and, and China really like stepped into the fold. And then particularly in the last 10 days there's two filings for Chinese model companies. One Minimax and one is called Knowledge Atlas Company which is which actually the regular name is like ZipU AI or Z AI. It's produced. This model is called GLM. Anyway, there's like too many acronyms at this point. But, but the point being is like these are the two like first pure play model, large model companies that have gone public and they've went public on the Hong Kong exchange for several billion dollars each and they've since like ripped. Yeah, so, so this is China has been first to market in this sense
- [14:59]
A
in the first market in getting a stock that people can buy.
- [15:03]
B
Yeah, yeah. Well Americans can't buy Minimax. If you have US Nexus, you can't buy it. I don't know, I haven't looked up the exact reason, but the other one I think you can.
- [15:14]
A
Interesting. Are there like adrs where basically they relist the shares? I think. I don't know exactly how that works.
- [15:21]
B
Technically there are no adrs yet. It's just on the Hong Kong exchange.
- [15:25]
A
So one thing I wanted to ask you about, maybe this is a good time, like open source versus closed source. What is the big deal with that? For someone who has no. Like, why is open source so important in the context of what does closed source even mean? Like maybe give us a real quick 10 second and then what the importance is.
- [15:43]
B
Yeah, there's a spectrum on this. So traditionally open source in the context of like regular non AI software meant that the, the source code, which was basically like the end to end sort of book, basically that produces the program, was available on the Internet and could be reused.
- [16:03]
A
You could copy and paste the code, publish it and use the software technically.
- [16:08]
B
Yeah, yeah. And it would work. Yeah. And you're allowed to use it sometimes for non commercial reasons, sometimes for commercial reasons, with or without paying a license. And the beauty with open source that people got excited about was that anybody on the Internet can propose a suggestion, a change or discover a bug or produce a feature and then the maintainer of the project could, could basically approve or deny or provide suggestions. And, and there was, there's like a moniker that a couple of years ago that was used to describe this. And it's like the bizarre where everybody can kind of participate and then in, in other ways. Closed source software is kind of like a cathedral where like you have one access point and no one can touch it, no one can touch the cathedral, no one can add anything to it. You can't influence how it looks or how it will be changed. It's the owner of it that decides and that's that closed source and you pay a license to get access to it.
- [17:08]
A
You, you go and you offer like an offering, like a tithe to the cathedral to like use the, the cathedral tax.
- [17:15]
B
Exactly. You pay tax in the form of a SaaS license.
- [17:18]
A
Yeah.
- [17:19]
B
And then this nomenclature just gets a little bit more complicated in AI because there are more moving parts. So there's, so, you know, for simplicity's sake, there's the training code which you know, describes how a model, how an algorithm should be run. And you know, what's the goal of the training? What, what like objective are you trying to minimize? For example, like you know, the dog cat thing, classification, you're trying to minimize the mistakes. Then you have the data set and then you have the model itself, once it's been trained, which has weights and parameters, which are like settings effectively that are produced as a result of running data through, through an algorithm. And so what is it exactly that we mean when we say open source model? Is it the training code, is it the data set, is it the model artifact once it's been trained in the sense of its weights or the code to run it? So there's a spectrum of definitions right now a lot of people use like open weights to mean. Well, it's, I'm actually just releasing the resulting parameters, tuned parameters of my training run. This would be the equivalent of saying, you know, I'm basically giving away the result of hundreds of thousands of hours of GPU training or something like this. I'm not necessarily giving you access to the data, I'm not giving you access to the training code or the pipelines to run it.
- [18:48]
A
So you couldn't reproduce it. You can just see the results of what I did.
- [18:55]
B
It's hard to reproduce the training run. Yeah, because you don't have the data set, but you're getting the end artifact. So it'd be like, I don't know, like the Formula one analogy, like the car that wins the last race at Abu Dhabi in the year. You get that. You don't get all of the learnings and recipes in order to get the car, you just drive the car. And so the community generally wants to have open everything because it leads to, you know, better reliability, more contributions from, you know, the ecosystem in terms of features and bug fixing and things, and then crucially like control over the entire system. So, you know, if a company decides to no longer publish open, open access and open source tools, then you can get rug pulled because you just no
- [19:52]
A
longer have access to it. Or updated.
- [19:55]
B
Yeah, exactly. At least you have like the last timestamp release. So if the system's good, you can still work with it. I think the second reason why folks want open source is because the training of these systems is expensive, generally for anything that's state of the art, like millions, tens of millions, hundreds of millions. Not many companies are, certainly not universities or small groups have access to those resources. And so they can't train systems from scratch. So they rely on these getting like airdrops basically for free. And, and like, I think the last one is just the community doesn't want to have, you know, one, two, three companies the entire future of AI development because it is very limited on money, on compute resources and on. And on dataset
- [20:56]
A
and then the reason you would be doing a closed source model is because it's probably easier to make money from it versus like you have to pay to use it and access it versus just giving it away and you can't charge as as much or anything.
- [21:09]
B
Yeah, I know my two takes there are like open source is almost good when it's a necessity for your customer to buy your thing because if, if your customer is a developer they probably and the tools cater towards them or some low level like database or framework or something like that. They don't want to peer under the hood and see how it works. And so if that's part of your sales motion then you have to do open source. But I feel like rarely open source has been a competitive like tool that one uses because it's like good for business. It's almost like a necessity or secondarily if the business is born from an open source project like databricks with Apache Spark. Like the thing is so complicated but very, very high value customers don't want to have to manage it themselves. So like let's pay databricks for the expertise because they wrote it and they're the best. And then the other part around like open source versus closed is considering that it's like Apple versus Android. Like at the end of the day I think customers and regular people want a tool that's like really good, that's designed well, that's updated all the time, that's reliable, simple to use and that's basically Apple. There is a period of time in like the early days of technology I think where a lot of nerds like to play with it and like mod it and change the background and like change these little things. But at some point it's like that community I think goes down over time because it's just a faff, like it's annoying. And I just think like you know when the last time was that I added a mobile phone number that was not an iPhone like this. Ooh, it's a green check mark in your imessage. It's like weird. I think this is because long term people gravitate towards convenience and quality and I feel a little bit of the same way in the open source versus closed source.
- [23:17]
A
Interesting. Yeah. Because it probably gets to a point where as the technology continues to get better you're just like I just want this thing to work so I can do my value additive thing that I'm doing on the model versus mess around with tweaking things because I just know that it works and it has what I need it to have. It kind of reminds me it's kind of interesting how Apple and I think of how Apple evolved. Like I used to think of Apple as like for old people. Right. Like back in the 90s. Right. Like your grandma has a Apple, one of those big massive, like monitor things. It's like a fancy color. Like it's like purple or something. And then it's like there's only one button on the mouse and it's super simple to use. Like, you can't get a virus because like the old people, you know, they'll just get phished and like they get a computer virus or whatever. But it's like totally evolved over time to now where most people just use a MacBook product. And you're kind of weird if you use Windows still. And I feel like people will reluctantly use them if they're working at a big corporation or something like that, where there's like a work necessity. Yeah. It's just interesting how Apple's kind of evolved the products over time.
- [24:25]
B
Yeah. And I think realistically, as, as model capabilities have gotten better. Like these, these things are so unwieldy and complicated and insane lifts with thousands of people, like, contributing in very tight control. And I just don't know how you coordinate all of that on the Internet with like randos on your laptop in cafes.
- [24:45]
A
And is it probably. I mean, and there's tons of forks, right. Where you basically take it, you make a tweak and then you republish it. So there's probably thousands, tens of hundreds of thousands of just like var variations of different models that have been slightly tweaked over time, I'm assuming.
- [25:01]
B
Well, there's that in the open source community. Yeah, but, but like to create a Gemini or like an chatgpt or something. I mean, there's. There's so many. There's like basically all these different teams that are each responsible for like a different skill or capability. And each skill or capability has a set of evaluations and sometimes like hundreds of them, which is like ways that you test whether the model has succeeded or failed or what end of that spectrum for that specific capability. And then you have to collect data specifically for that and then, and then just throw it into the soup and make sure that everybody else who's throwing in their capabilities and data and metrics and evals into the soup doesn't degrade each other's capabilities. And so like, how do you, how do you coordinate that at scale on the Internet, like Without being a centralized company. I just think it's unrealistic and I think the proof point to that is like no one has really managed to do it in the sense of no one's managed to publish like a model as good as Light Opus 4.5 or GPT 5.2 like in an open source setting. I mean people have come close but, but those are, those are basically private companies that operate. Those are like closed source model companies that just happen to open source as opposed to like true open source development which is like distributed contributions.
- [26:23]
A
So it's almost like you need a dictator or an opinionated. The person managing the whole process of training and releasing it.
- [26:33]
B
Yep, exactly.
- [26:35]
A
So then I know we've been, this has taken us a long time to get through the top four most interesting things, but what would you say is like the third or fourth most interesting thing?
- [26:46]
B
I mean the third or fourth we can probably breeze through a little bit. The third is around like real revenue scale of products. Like it was only two years ago that you know, the major labs were de minimis revenue and you know, there was that example that you know, Jasper was making way more money than OpenAI was for very similar use cases. And this debate of is the model the product or should there be lots of other like scaffolding and sticky tape and UI that around the model that would mean that it could extract more value for that targeted end user. And then that seems to have flipped as the core model is better and it's been kind of the interface for everything. And then model vendors have created their own sort of SaaS wrappers basically. So that's like tens of billions of dollars across major players. And then the last one I'd say, and there's many more, but last of my favorites at least is just like the AI sovereignty marketing agenda has just gone like on full steroids. So this is, you know, the competitive positioning idea of nation states need to have access to energy, compute data and talent and models in order to dictate their own fate. And the future of AI. And Nvidia has gone on a very aggressive marketing spree like selling this agenda to every single country that is willing to listen. And that's resulted in commitments of over a hundred billion dollars globally in different countries to build up data center capacity.
- [28:31]
A
Just pick a random country in Europe, like the fourth largest, like Italy or something in Europe. And I need to have my own sovereign AI strategy. What am I probably doing?
- [28:42]
B
Well, this is kind of my point. It's like what does that even mean?
- [28:46]
A
Yeah, what does it mean.
- [28:48]
B
I just keep invoking this Kanye west lyric of no one knows what it means, but it's provocative, it gets the people going. It's a bit of that. What they think they're getting is the ability to train and run AI models in their geographic vicinity in a way that cannot be unplugged or tampered with by any other nation state than themselves.
- [29:12]
A
So like, if you go to war with a country that a major lab is domiciled in, you don't get cut off from the ability to use AI?
- [29:22]
B
That would be the idea. Yeah.
- [29:24]
A
Yeah. And I could see how, you know, if AI continues to kind of get better, it becomes all software. Like you basically can't use software to defend yourself and maintain your sovereignty and you'd be at an extreme disadvantage if you didn't have that. So I would be extreme if it's positioned like that. To me, I'm spending a couple percentage of GDP on. Yeah, it's like I'm opening the checkbook probably.
- [29:50]
B
Yeah, yeah. But the challenge with it is that the hardware improvement cycles are very short. You still need to have like a bunch of software that runs on the chips to make them work. And maybe there's like a major update that gets shipped to them that's required to run the, to run new models on your hardware. And you know, the vendor decides not to ship it to you because, you know, we're not friendly with your country anymore. So congrats, you got the hardware, but you can't run it. A bit like the whole spiel with the F35. Yeah, foreign nations can buy the F35, but you can't fly it if you don't input the flight plan which gets communicated to the US So I didn't
- [30:35]
A
know that they worked like that. I mean, that's tricky. Yeah. How do you get around that?
- [30:41]
B
Or like it's just could be nerfed or.
- [30:44]
A
And it's interesting from Nvidia's perspective because Nvidia has a customer concentration problem where basically Google, Facebook, I don't know, Amazon and OpenAI maybe anthropic is like most of their revenue, like I don't know, 90% of their revenue, something like that. So it's like, can you figure out how do I find a couple more nine figure customers, Get a couple more people that are paying me billions of dollars a year to get a little bit more diversification.
- [31:14]
B
Yeah, I think it's a great strategy by the way. I think they're right to do it. And you can see that the next growth spurt is data centers in space. So it's like, what's always the next, the next tam? You can, you can create, right, to keep, to keep selling.
- [31:32]
A
What is sort of the end state of that? Because if you kind of go back maybe a couple years, if someone were to say, we're going to build data centers in space, it just seems like just a word salad of just random buzzwords that makes no sense. Does the entire universe just need to be like, filled with data centers? Like power AI? Where does it end? Do you know what actually happens?
- [32:02]
B
I'm not sure what happens at the limit, but I could see a pitch around as we get more assets in space of various kinds, and maybe we'll have space civilizations at some point. And so having compute there, it's just better than having it on Earth that we have to communicate to and from. Which currently is cheaper than having data centers in space, potentially even for, like, you know, defense reasons. You might want to have compute in space for all of your, your space systems. But yeah, I'm not, I'm not the space data center expert at this point.
- [32:44]
A
Yeah. Because, like, part of me is like, this just sounds like it's hundreds of years in the future, but also maybe part of me is I feel like there's that quote of less happens in one year than you think and more happens in 10 years than you think. So with self driving, my working theory with self driving is it's always going to be a couple years away, but I guess now it's actually finally here.
- [33:05]
B
Yeah. But also historically, Nvidia has been exceptional at, you know, resourcing teams, whether they're, you know, in companies or in academia who are exploring new things with their stuff. And, you know, before it was, it was like shaders and graphics, and then it became like AI with Alex Ned in 2013. And then it was, you know, these, these AI labs, nonprofits that were started 10 years ago, and then in biopharma, and then now if some, like some kids are launching, you know, your GPU in space. Like, I think Nvidia is pretty curious to know, does it work? And if it doesn't work, how can I fix it? Because who knows where the next growth frontier is? And if that's driving excitement for this stock, even better. But I think their net outlay and figuring out if this works or not is pretty low for the potential upside they could get.
- [34:06]
A
And it's not like they're, they're losing money because they're like, selling products, they generate cash on right. So it's like you can scale it up or down. Like, oh, space data centers didn't work, but we made $3 billion because everyone wanted to buy some for a year. It's like with crypto, like you go back to 2021, like the crypto bubble almost is like dwarfed by the AI surge.
- [34:36]
B
Well, this is interesting because as you, as you said, it's been dwarfed and then the, and then crypto prices, you know, year on year now are like down whatever, like 20, 30% or so. I remember this time last year was like 120 or something. And now we're like what, 95 at least in BTC.
- [34:53]
A
97. Just in, just in Bitcoin's in 97. I saw a notification. So it's going up. It's going back up.
- [35:00]
B
Yeah. Like bitcoin miners have found something far more lucrative than mining bitcoin now because I guess because of changes in the hash rate, et cetera, and energy prices and that's like basically swapping their compute workloads from mining to, to AI. And so like last year we saw this rush of companies from, you know, Irish Energy, which is this like Australian business doing energy originally and then, and then starting to do BTC mining or crypto mining and now is like transition to being a GPU company and same thing for, for Ciffer Energy and, and then like hut 8 and all these rando companies that are.
- [35:37]
A
Have you heard of any of these?
- [35:39]
B
Yeah, I mean you can look at them. There's like six of them and they've all like ripped like 100% or 50% last year. I mean, they're crazy volatile, so you got to have a stomach for it. And who knows if any of them are really going to survive because they don't have the balance sheet for it. But, but as you know, traders are looking for volatile stuff to, to just like play on Robinhood. These things have been pretty, pretty popular as everybody's trying to front run. Oh like, you know, Anthropics launched a new, you know, 1 million TPU data center thing. Who are the other vendors that can profit? And so I think for that reason, like, you know, BTC prices have also gone down.
- [36:16]
A
This kind of leads into interesting other kind of like phase of the conversation that I kind of want to ask you about is like, how do you feel about where we're at? And kind of the cycle of excitement in AI. Like maybe the more simple way of phrasing it is like, are we in an AI bubble right now? I Don't know. How do you think through that? As someone who's been in the space for a really long time, the first
- [36:37]
B
thing I think about is if I were to have seen the capabilities that we have today, like this magic box, you can ask anything and basically gets like anything right? And you showed me that 10 years ago, I would have told you that is like absolute magic. And pretty much everybody in AI would have told you that's magic. That's not possible in a decade. And this is across just general question answering, but also video, understanding, video generation, audio, like it's insane what we have today. I would have thought this would taken way longer. And the future got pulled forward so damn fast. And then we're still only in what, two or three years into this cycle of getting this like magic alien artifact with no instruction manual, no genius bar that tells you how you're supposed to use it. And the very companies that have developed these tools are figuring out how best to use it so they can educate everybody else. And then you already see like vignettes into the future where companies that have adopted this and people that have adopted this have significantly higher productivity than they had before. As reframed as if you were to take this classic question of if I were to take this thing away, how pissed off would you be? And most people would be like pretty beeping pissed off. And then also other things like cloud spend as a proportion of overall IT spend is still not ginormous. And that's been the case, you know, since like 10, 20 years. And that hasn't even factored in like AI distribution. So I think we just have so much more to run if you just consider what we have today and educating everybody on how they can use this stuff and also factoring in that the artifacts we have at least of, if you look back three years ago, were a result of like a very, very small number of like highly technical nerds who like stumbled something that really worked. And we're not product designers, we're not like behavioral psychologists, we're not, you know, like large scale systems engineers or cloud or like, you know, cost optimization experts, et cetera. All this stuff, all these tasks that are super important building like basically the Internet had not worked on AI and now everybody's going into AI. So I think there's far too much like money, resources, talent and genuine desire to use these things that progress can't get better. Now does that mean currently we're in a bubble? I don't actually think so. There might be pockets of bubble ish Dynamics. But the companies that are accelerating super fast in AI like mag7 are actually printing tons of money. They have super healthy balance sheets. Forward valuations are I think 50% of the top names back in the late 1990s. Um, they're using their own money to fund a lot of these data center buildouts. I mean there's some that are doing off balance sheet credit but I mean it's like sophisticated buyers that are buying this. It's not Joe blogs that's you know, betting their pension or something on you know, Meta's data center build out.
- [39:55]
A
I think I saw a really interesting stat too. It was that this was probably about maybe six months ago, nine months ago at this time. So this is probably even different stat now. But it's basically OpenAI and anthropic. Since the launch of ChatGPT had added more new revenue than like every other publicly traded software company. And this was you know, a while back. So it's probably an even bigger diff, like it's probably even bigger now.
- [40:19]
B
Yeah, it's astonishing.
- [40:22]
A
The valuation of these companies are just like do people pay for your product and do you make money on it? Maybe that's a whole different question here. Like you can argue about the margins of some of these things but like people are adopting the products and paying for them much more than if, if you don't have any AI capabilities in the product.
- [40:41]
B
Yeah, well the, the margin profile looks like it's actually improved quite a lot.
- [40:46]
A
So. Why? Why so? Because I feel like even within the past couple months I still see some of these headlines where it's like oh, the vibe coders have negative gross margins and they're three months from bankruptcy or something like that.
- [40:57]
B
Yeah, well I guess those are a bit different because they're like pass through to model providers. So what I'm thinking, you know, the margins are pretty good. I'm talking specifically about model makers, whether they're in like image or video or, or chat. And I'd say without disclosing like specific names of companies, like some of the best ones are like at point some 60, 70, 80% gross margin on serving their models. And yeah, they spend a lot on training but to some degree less than they did before. Because whereas pre maybe 1 year ago the sort of recipe for how to build a large scale model that worked really well was not clear or not written and now I'd say like most, most like people at the frontier would say like there is a recipe for scaling now. Like we sort of know how to do Pre training, the human mid training and then the post training and the RL and the evals and all this stuff kind of know how it works.
- [42:03]
A
So we're getting more efficient.
- [42:04]
B
A lot, a lot, a lot more efficient. And then, and then the other thing is the, the quality of the model you get today compared to what it was a year or two years ago is actually far more useful. So then you can monetize it better because it's, it's monetizable. Like people were willing to pay for the outputs and so you actually get faster payback. So you're more efficient. People pay for your thing because it's better. More people are educated on how to use it so they want it more. And then your general go to market is improving and there's some modalities that are cheaper than others. Like text obviously is probably the cheapest. Audio is actually not too expensive. Video and world models are very expensive.
- [42:48]
A
You're saying to, to make or to
- [42:50]
B
also sell to like it's expensive to do the training and also expensive particularly to do the serving.
- [42:56]
A
The inference so just the pricing to consumers then it's probably higher. Like if you want to tap into a world model API, it's probably much higher than just text based.
- [43:05]
B
Yeah, exactly. Like to make, to make like a nice video on Google's like VO3 or like the, you know, the super pimped out model, it's a couple dollars. So if you're selling to social media people and it costs a couple dollars per video, it's a bit tough.
- [43:20]
A
Yeah. But on the other hand is like if someone else like a agency or a designer in Illustrator or like literally take a camera, record an expert doing the thing or like an actor or a dancer or whatever you want. What is like the cost of literally type in make Nathan dance with an iPhone for an Apple commercial versus like actually go make it. Like is it a thousand times cheaper? And it also took two minutes instead of two months. I feel like that's part of the equation too of like as the products get better it's just like things just get so much more efficient where like you can literally have, it's like the person who's the marketing person at Apple, instead of coordinating with the agency, they're just working with Nano Banana or chatgpt or insert whatever tool they're using and just instead of sending an email to the agency going back and forth, they're just banging with chatgpt for a day. It's like cool. This is what, this is what we would have got for a thousand times cheaper. And literally, like in 10% of the time, we got it 10 times faster too.
- [44:24]
B
Yeah, exactly. So it's not particularly surprising that 11 Labs is ripping. I heard it's like. Or you tweeted it was at 330 million in revenue and got 100 million in net new ARR in five months. Bananas.
- [44:38]
A
Yeah, I saw 15 million in a day or something the they were tweeting about, which is. Yeah, I mean, that's run rate that. That's pretty high. That's what people are doing, right, is they say like, oh, we're at a $8 billion Runway because we signed whatever. Like our stripe balance hit the account today. If we Times that by 365.
- [44:58]
B
Yeah. I mean, that would be mad sketch. Yeah.
- [45:01]
A
Yeah. I mean, it's kind of happening, isn't it? A little bit.
- [45:05]
B
I think there's everything that's happening. I would say they're pretty above board, but there's certainly some odd behavior everywhere.
- [45:12]
A
Oh, yeah.
- [45:13]
B
But I think the magic here is just how much latent revenue there was in these use cases that was only made unlockable as a result of great capabilities. Right. Because I still remember, you know, two years ago when 11 started or when Synthesia started seven years ago, like, it was not mega obvious that there would be like half a billion dollars worth of revenue for a product like this, you know?
- [45:43]
A
Yeah. You'd be like, who gives a shit about why would somebody want an audio model? Like, there's no demand for that.
- [45:50]
B
Right, Right. Well, especially that there. There were audio models before. I mean, you know, Amazon had one, Google had one for TTS and speech.
- [46:01]
A
Was it that they just weren't very good.
- [46:03]
B
They were kind of shit.
- [46:04]
A
Yeah, that makes sense.
- [46:05]
B
Yeah. I mean, in London, like, where I was, like, I spent a lot of time. It was always funny to me that I'd exit King's Cross station where Google DeepMind was, and you would hear the audio voice and it was like, this is clearly a robot and it sounds really crap. But there's a company like 300ft from here that has a way better model for the last 10 years. And the tube can't even use it.
- [46:25]
A
It's like, we're still so early.
- [46:28]
B
Yeah, yeah.
- [46:30]
A
So I'm curious then, are there areas within AI that you think are, like a little bit overextended? Like when I just think about. Because if you were to open social media and you'd read the commentary from investors, a lot of people say AI is in this massive bubble and things are going to crash. Are there like, so it sounds like you maybe don't necessarily agree with that, but are there certain areas where you feel like there's a little bit too much, you know, things have gotten a little too far ahead of themselves maybe. And I think maybe another way to think about it is, you know, with kind of just, we kind of talk about across the board like the products just keep getting better. Are there areas where maybe the products aren't actually getting better at the pace of maybe like how excited people are getting about them?
- [47:13]
B
In some areas, like in science, I'm very excited about like long term progress in AI for science, which is anything from helping human scientists process research papers, formulate ideas, figure out what experiments they should run, test run those experiments and analyze the data. AI models are making substantial contributions there. That's a net positive for the world. Clearly in biotech this works. There's a good business model for it. If you develop drugs, there's ecosystem of pharma companies that'll buy them. There's contract research organizations that'll run the experiments for you at scale. And we've, the industry has transposed all this excitement over to materials. I think there's been a phase of several companies raising substantial money for materials. So much so that the venture dollars have gone to materials companies is far greater than US national funding for material science labs. So it's like, you know, whereas the US government should be funding frontier like you know, R and D, like venture dollars have stepped into there using many analogies from, from life sciences. But to me like I'm very excited about that space as well, but I haven't done any investments there because I think does it does not share a couple of the key features that is present in biopharma, which is like lots of pharma companies that have been structurally built to buy biotechs either outright or the drugs they've created. Like this is how biotech originated. This is like the tacit agreement that small biotech companies have with pharma companies. Like we take all the early risk and then if it works, you buy us. That's how it works. Material science. Not really. And then there's not, there's not like this kind of industrial base of contract research organizations which are like industries that will run experiments for you. In the case of materials, okay, my AI model has popped out a bunch of different versions of like titanium or like some composite. And I want to go make it. I can't send it to anybody to go make it. And so Most of these companies are either doing like their in house labs, expensive for all, you know, a whole bunch of other reasons, a bit inefficient, or they're sending it to academic groups who take like six months to make or make the thing and then they're making like a tiny amount of powder and it's like, you know, and, and, and then the other thing is then you have like big companies like Meta that have been pouring in billions of dollars into material science and then releasing you know, data sets and models and are really just constrained by synthesis like actually making the, the out the suggested outputs of the model. And so I think materials is like rate limited by all these things yet is seeing like irrational exuberance from venture investors as like the. That's similar to AI and biotech.
- [50:13]
A
Interesting. So what is an example of like one of these like materials or companies, periodic labs, what do they do?
- [50:22]
B
I mean they're, they're probably one of the, one of the most exciting ones out there because they have you know, very cool team, you know, two co founders, one worked at OpenAI on ChatGPT and post trading and then the other one worked at DeepMind doing material science. And then they're doing as far as public reporting as chronicled, they're doing in house testing of materials and then using AI to scour the universe of potential materials, do optimization and then using the internal apps to test, et cetera.
- [50:55]
A
Just to be like inventing a new periodic table, like inventing a new element that goes on the table or something like that.
- [51:01]
B
It wouldn't go so far as, so far as that, but it'd be like what's like a new formulation of like a superconductor or what's a new formulation of like an alloy?
- [51:13]
A
It's like stronger steel or something.
- [51:15]
B
Exactly, yeah. That could be stronger and lighter and cheaper to manufacture and therefore could be great for lots of different things. Or like a material that you could make, you know, an iPhone screen that you could then go outside and not get blinded and it would just work and you know, with direct sunlight, like
- [51:31]
A
things like that, that could be amazing. Like, like I can totally see how like you make these awesome products. Like let's use AI and make like, I don't know, like how do humans have wings and they can like fly or something? I don't know. Like I can see now like that's the tams are on. These things are insane.
- [51:50]
B
Yeah, no, I mean I think it's, it's very inspiring. The question is like would you pay Almost close to public market valuations from material companies. For a private business that's just started to do this on the basis that it could be the next frontier Lab in this space and follow the same
- [52:07]
A
fundraising dynamics, that's kind of the role that a lot of venture investors play, is the momentum identifying and trading almost, in a way. So I guess, I guess it's just like a. Is that your strategy? And if that's your strategy, then that's what you should probably do if, like, that's what you're telling your LPs is what you're doing.
- [52:30]
B
So, yeah, yeah, for sure. I think it's, it's, it's the strategy that one should do if, you know, one scaled to billions of dollars of assets. Like, I think there is probably no other way to, to move that volume of money into ideas that could be big enough, which you could, you could say on the founder side is like, amazing that, you know, entrepreneurs have the opportunity to go pursue their dreams with like insane balance sheets on ideas that should have been funded by the US Government a couple of years ago. But you know, at the same time, if, if it's your dollar that's getting spent and you look and you're a little bit more pragmatic, which is what, you know, LPs tend to be. It's, it's a bit of a stretch.
- [53:11]
A
Yeah. And it's just like so fascinating when I talk to people that were in the industry in the 90s or whatever and they're like, yeah, I had to sell 60% of the company for like a million dollars and there's like a 2x lick pref. And it was like a real product, real business. It was profitable. And then today it's just like I interned at AI Lab and I raised $8 million to go build material science thing that. Yeah, it's like I'm going to make stronger iPhone glass. It's like the, the contrast is pretty insane.
- [53:44]
B
Yeah, it's tricky because it's all about like risk reward. You know, dap, you take a lot of risk if the reward upside is big. But if the starting valuation is billions, then to reward is not necessarily capped. But I mean, there's some reality. I mean, as soon as you actually become a materials business, then you get valued as such or.
- [54:03]
A
Yeah, but the, the argument that people will use is that the outcomes are so much bigger today and that with generally with like, I mean, with a lot of AI software, it's like replacing the work. Right. So like your customer might historically only be spending 5% or 2% of their revenue on software, but they spend 50% on labor. So in theory you could maybe capture like half of their labor budget. So like your revenue potential goes from 2 to 20% of their revenue which is you know, 10 times bigger outcome. So you can afford to pay that 10 times higher entry valuation.
- [54:35]
B
Yep.
- [54:36]
A
So I don't know, you can argue it where it works.
- [54:39]
B
I think that's true on the outcome size. I would just rather not have to pay the 10x on the front so the return is 10x bigger. But you know, to each their own.
- [54:45]
A
I guess you come in at like you, you, you, you let the, the multi stage platform funds do that after you and then they, and then they help support what the, the, the capital needs to like you know, king make it towards the public markets after you've invested as an, as an investor that would be ideal.
- [55:04]
B
But I mean you know, some of the mega funds are doing this already with these mega fundraises where you know, the company announces, you know, multi billion dollar price tag. But there's been three tranches before that where you know, brand name Megafund got in first. You know what for a seat manager is a high price but for one of them is like whatever. And then there's been two markups since then in the short period of time. So it's a bit of that.
- [55:30]
A
And you can almost manufacture those in a way where you put in 2% of your fund at 200 million posts and then you put in another 0.1% of your fund in a follow on round with one of your LPs who leads the round, who for them, it's also a very small percentage of their portfolio. And then you have an extremely small round where the company sells 2% of the business at a billion post money and everyone kind of looks good on paper and nothing's changed in the two months since the first round happened.
- [56:06]
B
I mean this is exactly what's happening sometimes. And it is, it is. I don't know, it just strikes me as pretty unhealthy and sometimes the rationale is, well, employees want to know that their equity is worth a lot of money and so you know, they'll get the equity early and then there'll be this big write up. So it's almost the way that companies can offer lower percentage ownerships in their business in the form of options but because the markup is so high, like the paper value is like already in the millions. And so then they can try to compete with offers at the Big labs, but obviously dangerous because of all the liquor features that are, that are present in these, in these extended valuation companies, not least growing into them. So that's a little bit scary to me.
- [56:54]
A
Yeah. And for people who don't know, liqpref is liquidation preference. And it's basically in a lot of these rounds that companies raise any dollar that comes into the business. That if there's a liquidation preference it's usually 1 times, sometimes it might be 2 times, 3, 3 times even depending. But typically if you're in your standard venture round, whatever dollar you raise, those investors kind of have right of first refusal on any exit. So if you've raised a billion dollars, the company needs to be worth a billion dollars and all that cash goes to them first and then beyond that, the employees, the founders, et cetera, get paid. So if you work at a, if you take a job at a company that's raised a bunch of money, I mean there's a ton of nuance to this, but it's probably like whatever amount you've raised, the valuation needs to at least be that price before you get anything. Like the investors generally get kind of a rice first refusal. It's not always the case. You should probably ask when you're kind of talking to I'm taking an offer and like figuring out exactly how that works. Google it. There's tons of YouTube videos that explain it pretty well. There's like full podcasts where people explain the nuances of this stuff. But the price, like just because you work at a unicorn might not actually be worth that much at the end of the day.
- [58:08]
B
I mean the other thing is whether you're stocks get vested immediately once the acquisition happens, if you stayed for shorter period of time than is like required. So if you get acquired in your first year and you're supposed to stay for to earn all your stock, like do you earn the 4 on the date of the acquisition or not? And that's like been particularly topical and these like pseudo acquisitions of shell companies and things.
- [58:33]
A
Yeah, I did actually have that with one company that got acquired. The the founder specifically negotiated all of his employees like fully vest which I think a lot, I mean a lot of this comes down to who you work for. Like who is the founder of the company that you're joining. Like do you trust them? Not just in can they build a good product? Are they commercial? Can they sell things, can they grow? The company is also like are they going to treat you fairly? I think is like a under discussed topic that is probably pretty important to think about nowadays. So then we kind of maybe alluded to this a little bit, but in terms of. So we could maybe talk a little bit about Air street and the fund that you run and just kind of like what kind of things do you invest in? Because we talked a little bit about maybe what you not be participating in. So maybe it might be interesting. Just like how did you get into this bridge? How did you start investing? I know there's. You've been doing it for a long time.
- [59:31]
B
Yeah, I started getting interested in venture capital in college because I started in 2006 and that was the era of like Dropbox, Skype, Twitter, Soundcloud, Facebook, et cetera. All these like tools that, you know, are just used as a regular college student and get excited about them or read about the history and was this classic, you know, kid in a dorm room like thought of something and it became big. And at the time I was majoring in biology and I got the chance to work at the Whitehead Institute, MIT in Boston, and that was the place where they did the whole genome sequencing project. There were some venture funds in the area that were hunting around labs for new ideas and drugs and innovations that they would then spin out into companies. And I was just really excited about this, like working on new technology, new frontier ideas, and then try to make something out of it for the utility of real people and current companies. And after I did my PhD in the UK from 2010 to 2013, I was really convinced I wanted to somehow play a role in the startup ecosystem. Didn't really know where to start or what would be a good fit. And so I think naturally being part of a venture firm would be a good sort of aperture for everything because you get this unfair calling card that you can use to talk to enterprises of different types, meet entrepreneurs who tell you things that hedge funds pay lots of money to GLG for. You get to learn of successes and failures and then I could maybe find what I found most fulfilling long term. And honestly what I found was I like working with brilliant people or trying to invent the future. And particularly at this intersection of AI science, engineering and building real companies, I became even more interested in AI like in 2014, 15. That was around when DeepMind was acquired in London and it was not that far away from where we were and I had some friends working there and it really captured the interest of machine learners at the time that their skills could be used for practical things. At the time DeepMind had shown Atari and it was the first time that A computer could solve this video game and discover tactics that humans had never found. And I think that nugget of, hey, if you could have a learning machine that could amalgamate more experience, experience than any human on the planet and learn from all these experts, then there's certainly going to be, like, solutions to every problem in the world that we haven't yet discovered, but that the computer can help us discover.
- [62:20]
A
It's interesting, like the classic, like, you know, chatgpt or Claude code, like, do this thing for me, blah, blah, make no mistakes. It's like kind of like the meme, but it's almost like real. It's like, solve chess for me. Or like, solve world hunger. Go.
- [62:35]
B
I would challenge anybody who's interested in technology and just world progress to watch the thinking game on YouTube, the contemporary story of DeepMind and all the stuff that they've done and walk away from that and not think, this is probably the most exciting thing that could possibly happen on planet Earth and I want to be involved in it somehow.
- [62:56]
A
I've seen it, I just have not watched it yet, but it's been kind of on my. Oh, I should convince my wife on a Friday night when we're gonna watch one thing to watch it instead of like, you know, something else.
- [63:08]
B
It's well worth it. I mean, the narrative is great. I mean, I think, you know, Demis Vesavas is like one of the, you know, perfect sort of like characters that, that, that you want to win. I mean, he's like, he's intellectually curious, like, brilliant, you know, like too modest and just super driven to invent new things and do it in his lifetime and feels like this irrational sense of urgency of, especially when he's describing, like, why, why the hell did you, did you sell a DeepMind? You know, like, looking back, it'd be
- [63:40]
A
worth trillions of dollars at this point.
- [63:43]
B
Yeah, exactly. I mean, he sold it for 500 million pounds, which at the time was like insanity. Now companies are raising that right from day one. And he's in this taxi, like, describing, like, look, like if I, if, if I gave you the opportunity to have like, you know, infinite compute and infinite money and resources to like, accelerate the potential invention of AGI in my lifetime. Like, there's nothing more that I would trade than being able to use my, you know, fruitful years when my mind is still working. I have the energy, I have, like, you know, the physical capacity to go, to go do this. Like, I could have waited, you know, 5, 10 more euros and like, struggled here and there because nobody wanted to fund this stuff at the time. But like, I wouldn't trade those 5, €10 for tens of billions of more dollars.
- [64:36]
A
Yeah, that's. That's huge. It looks like it's on YouTube. Yeah, I'll throw, I'll throw a link in the comments or in the description for people. So I guess you made this transition from you were working at a venture fund. It was a 0.9.
- [64:47]
B
I worked at a firm called Playfair Capital before that, which is pretty much a family office. And then I moved to 0.9 in 2017 until late 2018 and then had the first closing of Air Street Fund 1 on January 2, 2019.
- [65:04]
A
So what was that transition then for people listening that are curious how raising a venture fund goes? Because I think a lot of people kind of think, oh, this guy has millions of dollars in just investing his own money because he's rich, because his parents gave him money. How does it typically go raising a venture fund? Maybe talk us through how that went.
- [65:25]
B
Yeah, so it was basically playing pinball with my eyes closed is how I describe it. In the sense that you have this goal and pinball get the ball in the hole and then you've got the pins and you have to position the pins to get the ball in hole but you have no idea how to position the pins because your eyes are closed. And that's like the best analogy I have for what it was like meeting prospective LPs of any kind that would be interested in 2018. Like a solo GP, first time venture fund manager, early stage focused small fund by virtue of those features. And then also focused on Europe and then a little bit of North America
- [66:05]
A
and then AI too. AI was not hot also. It's like you should be doing VR or crypto maybe.
- [66:12]
B
Yeah, exactly. So, you know, nobody wants to buy that, that thing. And I didn't come from like a brand name firm for a long time that when it was like an easy spin out with like institutional investors. But like I was honestly like convinced that I, that I had to do it. I think one of the best ways you can make these kind of career decisions is like the regret minimization idea of like, if I don't do this, will I have like lifelong regret? And that was very much the case because I had this long term conviction that AI would be important. Much in the same way that SaaS was niche 15 years ago and then became the dominant business model on the Internet. AI would be that case too. If this stuff actually works. Why would you not build your product using it or powered by it. And different industries would come online in different times. And that was motivated by just the first principles, view of progress and the science. I then spent a lot of time with various venture firms trying to see if they were like also convinced on this. And broadly speaking, it was some version of like, no. Or it's a toy or it's like going to get absorbed in different things or we don't think we need to have specialism.
- [67:18]
A
Yeah. Because when I think back of like some of those AI products, like they didn't really work very well. So you probably, you know, if maybe like your perspective just having spent a ton of time, maybe you could see the trajectory or maybe you just said, but if you're just like doing all these different things, doing some like consumer D2C brands, you're doing like, you know, SaaS, enterprise CRM management or whatever. And then there's like some AI with like, you know, email assistant, AI thing that just doesn't work properly and the company goes bankrupt. It's like, why would we waste our time on that? Like, it's just not worth it.
- [67:53]
B
Yeah. I mean on that note, we forget the original X AI, which was an email assistant.
- [67:58]
A
Yeah. Which was like pretty hyped at the time. Time like it was like 10 years ago, right?
- [68:03]
B
Yeah, it kind of worked also.
- [68:05]
A
Oh, it did. Okay. I didn't never used it. So I just remember like hearing about it now. Most people are like, yeah, it didn't work very well.
- [68:11]
B
Yeah, yeah, yeah, yeah, yeah.
- [68:13]
A
I guess compared to today. Compared to today doesn't work. Exactly.
- [68:16]
B
Exactly. Yeah.
- [68:18]
A
So were you exploring joining a big firm initially and like leading their AI investing?
- [68:25]
B
I was open to various avenues, just trying to figure out what is like the right, the right like setting basically to express my ideas. And, and then, and then I just, I've eventually found like, look like I've been following and like writing about this or like industry analyst vibes for, for a long time, like since I don't know, 2015 or something. Or 14. Yeah. I have some like essays on like why Nvidia was going to be the most epic company ever and like the different areas in AI to like watch in2016 and it was like RL generative models, like world models, custom silicon, all the stuff that eventually panned out and some things that didn't. And then was spending a lot of time with different people in the ecosystem, like researchers, engineers, startups, big companies, policy investors and doing a variety of these meetups in different Cities because I always found as a grad student, I want to learn what good looks like, but there's no place that I can go for it because the playbook was still getting, getting written. And then I had a couple companies that invested in that were looking interesting and, and yeah, it was like the regret minimization of like, I just gotta try this because if I don't, I'll regret it. And honestly, like, I'm not that scared of the risk of having to go back and eat ramen noodles. Like, they taste good. I'm fine with that.
- [69:47]
A
No, yeah, there's these good. They have good flavors. Nowadays, like, there's a lot of options. Maybe not the healthiest necessarily, but like,
- [69:55]
B
yeah, but you got, you got supplements for that. It's okay.
- [69:58]
A
Yeah, you talk about playing pinball with your eyes closed. Like, how do you do that? How do you beat pinball with your eyes closed? Like, what, what did you figure out to eventually kind of beat the game?
- [70:09]
B
I figured out that entrepreneurs who had like, some exposure to finance, fintech, like AI, data science, like, vibes with what I was doing. Like, because they'd seen, seen the value of AI at the time. You know, most of the value was in like ad targeting, recommendation systems. And then I also saw that certain individuals who worked in like high frequency trading also understood this. Again, like, because quantitative modeling is very much about machine learning. There were like a few growth equity firms that had started to like, make early stage investments in managers to sort of like prime the ecosystem. And then, and then I stumbled on a couple of family offices, but this is like completely random. And again, like the pinball. And that was mostly through referrals of like, hey, you might not like this, but do you know somebody who does?
- [71:05]
A
Did you usually do that when you have a conversation with an lp is like, you usually try to get an introduction to someone else.
- [71:11]
B
Yeah, because, you know, the, the playbook tells you ask all your GP friends to introduce you to their LPs. But if you don't have fancy GP friends who have great LPs, how are you supposed to, like, what intros are you expecting to get? And so I would just like ask, ask entrepreneurs. Ask like, yeah, these family office folks. And I go to events. I would even do things like, I don't know, look on LinkedIn when somebody announced the fund and I'd look at like, who liked it and then see on the list, oh, there's somebody who manages, like asset manager. I saw you like this. Maybe you might like this.
- [71:46]
A
Yep.
- [71:47]
B
Yeah. Or like, you know, in, in the uk, another trick is like UK Incorporated Venture funds or comp. Every company in the uk, frankly has to list its shareholders on Companies House, which is sort of like an SEC register, if you will, and it lists all the names of the entities. And so you can like go there and look at your pure funds and try to see who's an lp and then maybe, you know, casually mention, oh, I heard that xyz, maybe he's investing and like, you know, they're an LP in their fund.
- [72:15]
A
Yeah, I definitely. Yeah, that's definitely the trick of like, you know, you know exactly what you want to get going into a conversation and you just kind of float this and just see how they respond. Like, see if the, if there's like a chance that something might happen.
- [72:30]
B
Yeah. So, yeah, short answer is like I tried, tried every, every trick really. And then it was just to some extent just like brute, brute force and then, and then some luck along the way. And you know, I started at, in January 2, 2019 with like 9.862 million. I know because like that, like Delta with 10 million, like pulled out three weeks earlier. So I'd had a bit less than 10 million.
- [72:58]
A
Nice.
- [72:59]
B
And then I did seven closes over like two years, which is pretty, pretty intense. And then hit Covid also towards the tail end of fund one when you couldn't squeeze 50k out of anybody. And then, then after that summer, it sort of assuaged a little bit and then managed to get a bit more into the fund after that. And then I ended up about, about 26 and a half. It's a fun one.
- [73:25]
A
Nice. And so, I mean, it sounds like it's basically like take a lot of, takes a lot of patience to kind of get through the process. The way I describe it to a lot of friends who maybe are founders, it's kind of like raising a pre seed round. But you don't stop when you get to a million or 2 million bucks. It's just kind of like you just get 150k checks or 100, 100k checks. They just kind of keep stacking. And even when you get a lead quote, unquote, typically when you're doing a seed round, they do most of the round, but most of the bigger checks in most people's funds are 10% of the funds. You almost need 10 leads, basically. So it's like, think of that process like it took you two months to get a lead. It almost takes you two years to get 10 leads. Lead checks, basically. So it just Takes a long time.
- [74:14]
B
The challenge is you have no leverage. At least with like a pre seed company or a startup. You're, if you do well, you're at that specific stage once in the entire life of your company. And so there are investors that specialize in that stage that either they're in or they're out and that's done. Whereas by definition good venture funds are here in perpetuity. And then the game is like, can you become access constrained as fast as possible? So then you earn some element of leverage, assuming that you don't grow beyond the capacity of the partners you have.
- [74:50]
A
Yeah, and then there's the element of people can kind of wait too. Like they can just be like, you know, your fund two seems interesting but like, I don't know, I just kind of want to see how well you do and maybe Fund 3 is more for me. And in theory if you're good, you'll, your returns will be just as good in the next fund. If you're, if, if you're a really good investor, which is probably why they'd be investing anyways for that reason. And you can't just be like, oh, we went out and signed a new customer and our revenue doubled. Like you can kind of go, oh we got a markup from Sequoia or something. But also a lot of LPs were like, that's cool, I wonder how it goes in 5 years check back and did you return cash? And maybe what you get to is you're on your fund three or your fund four and your fund one, the second investment you ever made, it got acquired or it went public and this is 10 years later. It literally takes decades actually. Show the tangible proof sometimes.
- [75:53]
B
Yeah. And then you don't know, maybe, you know, out of five funds it's like fund one, three and five that are great and like two and four are less great. And if you missed one or you pulled out, like you don't get, you don't get the benefits of, of this asset class where you really do have to, you know, invest through multiple cycles and then consider it as a multi fund commitment and you know, look at overall venture dollars like deployed across those vintages.
- [76:20]
A
So then what's your strategy with the fund? I don't know. What fund are you on today? I think I remember seeing a $100 million number.
- [76:27]
B
Yeah, I've just finished fund two. I'm on fund three now. So yeah, from fund two I went up to 121, which was a big step up from fund one. Really just enabling me to offer all the money that entrepreneurs wanted when they're raising either pre seed or seed, you know, anywhere from you know, 1 to 5 million, maybe $6 million. I've just generally been of the belief that you know, not, not that many companies matter. I only, I want to be concentrated in, in the fund. So I do like 20, 20 companies per fund. I don't get excited about things that often but when I do I want to be able to move with like conviction and have the money to offer to the entrepreneur because I think the experience kind of sucks when you meet somebody you absolutely love and you love the idea and they're raising X and you can only do 0.2x because that's that, that's the amount of money you have in your fund and then you might be kind of forced to sort of massage the fundraise to meet your fund model. And I feel like that's just net bad for everybody. And then I want to do like Europe and North America. I'm flexible on the geo and generally like pre seed to seed and then a couple areas I like doing are like vertical software where like AI is the product. So I've done things like you know, V7 which does kind of process automation and spreadsheets or like synthesia and 11 labs a bit in dev tools and infra like poolside, you know, coding model and then defense and security. Because I'm like I think this is really important. Freedom does come, doesn't come for free and it's non negotiable. And so I have investments like Delian in the UK which builds like hardware and software for like perimeter security, anti drone and then the fourth bucket is tech and bio. And so there I've had some early exits in Fon1 around you know, generative models to design new chemistry. It's a business called Valence that I led there was sold to Recursion and then another one called all site where it was kind of the opposite where like testing cancer drugs on samples of patient tissue from cancer patients and then kind of running a clinical trial in a dish, then using like computer vision with microscopy to take pictures and analysis of these cells and figure out which ones are working and which ones are not working with respect to the drug response. Then we sold that to XTNTO and then XTNTO went public. Those are the four buckets. And the things I've done less of are these large capital raises for model companies. In part because of what we discussed earlier. It just feels like the economics are a little bit tough for early stage investors. I'd almost prefer to invest in a slightly later stage around for those businesses once the economics are much clearer and they have customers and they're scaling versus like a tabula raza, you know, 50 million on 200 and something to train a model and maybe it'll work, maybe it's not. Whereas you can invest in like a Series BCD company that's like printing 100, 200 million revenue, growing 2x year on year, like I don't know, a billion or 2 billion. So it's like the cost benefit like seems completely off to me.
- [79:49]
A
Had a company got acquired by Anthropic and I have like some Anthropic stock and I, I didn't know how to feel about it at first because I was like, ah, I don't know about this. I'm like, it's just like, it's like growing pretty quick I guess like of all the assets to own, maybe it's like a good one to just kind of have and like, I don't know, it'll be like a driver in the fund. It's just like, I don't know how big it will ultimately get. Like is there 100x upside from here? I guess it depends who you ask.
- [80:14]
B
Depends what heroes.
- [80:16]
A
Yeah. So it's like I don't know, I guess it's like better to own that than something else. I don't know.
- [80:22]
B
It's like yeah, I mean I think in a, in a portfolio like different assets play different roles. You know, there's some early exits that can provide recycling that are good. There's, there's others that maybe at a later stage to provide good irr. And then so I think it's just about thinking what's, what's like the right, the right mix for these like later stage opportunities. I do think that to your point earlier, like these companies can become way bigger than we thought. You know, Anthropic. I think an OpenAI of like hundreds of thousands of customers. Like same thing with some of these model vendors and different modalities and whereas 10 years ago it was like pulling teeth or maybe worse to try to get any enterprise to even try your AI widget. Now it's like, please can you help me? And like what can you help me with?
- [81:13]
A
Yeah, how much money can we give you?
- [81:15]
B
Yeah. And so you know, if we start with like some use case one and you succeed on that, likelihood is other departments are going to see that use case one and think, oh well, I have Something that looks similar to this, can you build that? And then when you can you know, code these things way faster than you could before and everybody gets like mass customized software, I think you can start eating into customer budgets way more than you could in the past and act as like that one sort of like lighthouse guide into this next generation of software. And so you become multi product and then add revenue lines and I think you scaled to really big sizes. So I think some of these later stage bets still make a lot of sense.
- [81:59]
A
So you're thinking basically the at least your opinion. It sounds like you want to invest when somebody is starting a company and there's appropriate risk for like this is probably going to fail and not work or you invest in this clearly is working and this is a company that is going to exist. Like there's no going concern and it's growing really fast. So it's basically you invest like company creation or like this is a mature AI company and there's almost like a dead zone in between of like it's still unclear if it's going to cross the chasm but you're almost paying a price that it's appropriate that's related to like a publicly traded software business or something like that. And like that's just not a place you want to be in.
- [82:42]
B
Yeah, yeah, exactly. And I do at the moment, you know like 90% in the first bucket and like 10% in the latter. And then from the 90 I also do you know, reserves for, for the core bets. But yeah, when I have like you know, 20 core bets where I'm buying you know, 10, 20% of the company for like serious money and then you know, following on in a couple of those like through A and then maybe like 2 through B in special opportunities. So for example in fund two, I'm really excited about this company Profluent which is you know, training large models to do protein understanding so they can engineer proteins with specific functionalities and they've basically focused on genome editors like these proteins that go into your DNA and like extract certain pieces of DNA and insert another or fix a genetic mistake which can be responsible for disease and therefore like that solution is curative. And you know, in that example like I've invested in seed A and B. You know I have like 15% of fontu in the company which I think is probably larger than like every other shareholder. So I'm like definitely risk on on like companies I like because I think you know, if, if these businesses really become generational and work out like it really moves the needle and across several funds. I think this is the better strategy than personally than you know, large number of portfolio companies, smaller checks and then, you know, higher likelihood you don't lose money, but also lower likelihood you have a blowout fund.
- [84:10]
A
Yeah, because if you, if you look at a lot of like the data and research, the, the data driven approach to this is a super diversified portfolio. Like that's, that seems to kind of be more of the consensus right now in early, early stage. And I think 20 companies in a fund is below the threshold of what people would say is good diversification. So you maybe explained it a little bit. So why did you not say maybe 40 or something like that based on the research? Because you're a research driven guy. You obviously read a lot of research.
- [84:45]
B
I think part of it is it's hard to make the math work at 40. Like if, if most companies are, are raising 3 to 5 million dollars if they're not one of these like model training shops to be able to buy up like good ownership and do most of the round it and then you don't want to have a fund that's like half a billion or something which has its own, you know, exit value capture assumptions. It's hard. So I prefer to skew with fewer companies. And by the way, like I invest over three years, so it's quite slow. So I get some time diversification. I could sort of see like, okay, is you know, the vintage from year one and year two panning out good or, or less good than I hope. Should I add more names at the expense of pro rata in, in year three or not? So I had to think some more flex in the system to decide should I expand the number of names or keep it small because of the longer investment period just based on my own pacing and lack of excitement every single day about. Ooh, Squirrel. Ooh, Squirrel.
- [85:56]
A
Yeah. Well, I think the other thing too to bring up is we are in the hottest fundraising market ever for AI. You should be deploying in a year and raise a new fund at twice the size. Why don't you do that? Because you can make a bunch of money.
- [86:10]
B
Personally, I think the best answer to that is, and this is a founder who brought it up when we were discussing like fun strategies. He's like, eventually used to be about the you're in the 2 and 20 business or whatever your math is, but now you're in the two or 20 business. And I think obviously big firms are basically in the like in the two, most of them. And I just, I want to be in the 20. Like, I'm, it's just like in my, in my gut, like, I, I'm performance driven. I want to be able to, you know, get a line item, you know, in an endowment saying, like, you know, we gave a couple tens of million dollars to this guy that we, like, found from, you know, from Europe, whatever, like, doing AI before people thought it was cool. And like, he printed billions of dollars for us over, like, our relationship and that bought us like a bunch of buildings.
- [87:04]
A
And yeah, he returned the endowment.
- [87:07]
B
Yeah. And like, this has happened before. Like, for example, if you read some of the old Yale reporting, like Swenson wrote about Hill House, which was founded by some analysts in the investment office covering China, and they wanted to do Chinese equities before people thought it was cool. They were like in their mid-20s. And the line is something like, we gave them tens of millions of dollars and they printed billions for the endowment. Like, I don't know if this is going to be repeatable. I mean, obviously high mark, but philosophically, I'm much more aligned with, with that than asset gathering.
- [87:37]
A
Yeah, I think like a whole other vector of this is just like, how big is the team needed to execute the strategy and like, what are the cash inflow needs of the firm? So if you, if you are literally one person and you can live off 200 grand a year, live a nice fine life, maybe you have like, kids you want to, like, they do dance classes. Maybe you need 500 grand. Maybe you, like, live in New York, downtown Manhattan, whatever. But there's an upper limit of what you need to survive and live a comfortable life. To execute the strategy versus do you need 50 people on the team, 100 people on the team? You can't do that with a couple hundred thousand dollars. You need millions of dollars to pay and compensate these people. So you do actually need a significant amount of management fees. And the funds do stack over time. We'll say 10 years, you're paying 2% a year, which is maybe like the average. And you're 10 years into the strategy, you have five funds stacked. You can afford to pay some people, but depending on what the strategy is, you do need a budget to work with and you need to generate the management fee. So I feel like that's also too is like, depending on the team size, depends on how big the fund needs to be. Really. Like, that dictates a lot.
- [88:51]
B
Yeah. But then there's this other narrative on Twitter recently of these spin out gps spin out because they say Our partnerships are too bloated. There's too much process, too many meetings, too many companies. Decision making is inefficient and we're fresh. We have no companies. Smaller partnership, whatever, smaller funds. And then you just kind of wind the clock five or five or so years on most of those teams and then they've become exactly the firm that they left. Like they've hired more people, they've raised a bajillion more dollars, like they have more portfolio companies, the experience degrades, et cetera. So I feel like, you know, I got into this in like a non traditional, like not super popular route. I mean most LPs would say like go find a partner instead of here's some money. And so I do, I do want to stick to what I think makes like this product like different and feel different. And it would just feel like a bit disingenuous if I'd go and hire five partners and be like, oh well we're like everybody else now because we want to scale like Aum or because the opportunity sets bigger. Like I do like just only having to care for portfolio companies and like what I do every day versus having to care about and spend a lot of time on, you know, nurturing someone else's career and, and, and keeping them happy and teaching them, which I think you have to if you're going to hire people. Like you have to make that commitment. I'm just not ready for it.
- [90:25]
A
Yeah, and I think too is when you think about like, like if you were to go work at the. We'll just say like the largest multi stage fund. I think there's one. They just announced a new set of fund. It was like, is like $15 billion or something like that. Your personal compensation, that what you get personally from any check that you write. There's this whole fund from angel, small fund, medium fund, big fund. So if an angel writes a check, they invest $100,000 or $10,000 of their own money. That is the equivalent of a solo GP writing $100,000 check or $200,000 check. And the more bigger you go up the stack, like literally a $10,000 check from an angel is like the equivalent of like a, a mega fund writing a $18 million check or like $50 million check of like how much they are personally compensated for that investment. You know, it kind of depends on how a lot of different things, fund size, compensation structure. But like so like you're almost like disconnected from like the actual outcomes. The bigger the fund is on like a Like a personal decision making level. So I feel like it's just like this whole other. And then there's this other function too of like as a solo GP. It's like you get I don't know, 100% of the carry maybe like depending on, you might have an associate but like you are personally compensated about like the result of this thing. And like do you, when do you need access that like if you're just like you're 35 years old. I don't actually don't know how old you are, but like.
- [91:55]
B
Yeah, 37.
- [91:56]
A
Yeah, yeah. It's like, okay. Like you have, you have like hundreds of millions of dollars of just like illiquid net worth. It's like whatever, it's not that big of a deal. Like I'll get it in 20 years. Like it's, I don't need it today. It's like I'll just let it continue to compound, I guess, versus if you have a bunch of people like a bunch of cooks in the kitchen of like I need it today. Right. Like you might actually need the management fees to like pay bonuses or like it influences like when you take liquidity which impacts how much money your LPs make at the end of the day. So I think there's that whole element of it too.
- [92:30]
B
Yeah, yeah, yeah. I think you just make different long term decisions because you're less like, you know, markup driven potentially and therefore less seeking momentum because that's like the only way you can show to your manager that you're sourcing good things. And then maybe like, you know, everybody wants to do a spin out fund at this point because like why wouldn't you? For the same reasons you discussed of like your look through ownership and the companies you invest into and your big fund are so small. And I think for in a market that's like flush with capital, like the closest that entrepreneurs will get to a venture product that is entrepreneurial is somebody who started it themselves because they've been through this shit of like no one believing in them, of like figuring out the strategy of like tuning the strategy, meeting all these people, like managing like the organization, even if it's just one person, there's still like the operational stuff you need to get right, particularly if you want to work with the serious institutions. And then there's like the brand, the marketing, the sales because we have to do all these things as well, like to create a differentiated product. And then if it goes wrong, there's nobody else to blame. If it goes right, then all of your success to you. So I think for an entrepreneur who wants to find somebody who's most aligned with them winning, it's going to be somebody who's like, new, who started it themselves, who has their, like, career and reputation on the line and money on the line on the success of the entrepreneurs they work with. And I think for, you know, founders who are like, who vibe with that, it's amazing product, and for founders who, like, want the, you know, nice brand, like, I think in that case, it's hard to. It's hard to convince somebody otherwise. A bit like consumer preferences.
- [94:13]
A
Yeah, it's like a market. Just like the market, people want different products and it all coexists and you can pick. That's the beauty of, of capitalism. It's like everyone will come up with a different product that does different things and you get whatever you want. So it's all good, actually. So one thing I wanted to ask you about, maybe going back into maybe some of the stuff in the report, maybe we actually hit on this. I can't remember, but there's all these benchmarks. Every time a new model comes out, there's all these benchmarks. Each new model is like the best at something, right? Like, how should I interpret that? As just like a observer of reading all this stuff. Like, do they matter? Are they super important?
- [94:59]
B
I think the way to look at it is almost like the Olympics, if you will. And like, each model is a country and then each, like, sport is a task. And so model vendors are trying to do the best they can across the board in different tasks and different sports. And some companies will care more about certain sports than others. So far, Anthropic cares a ton about code and basically nothing about audio. Multimedia doesn't even expose those capabilities, as far as I can tell. Whereas OpenAI cares about all of them. And those, like leaderboards and competitions are useful as systematic ways that one can compare the pros and cons of various systems. But we do have to have benchmarks because otherwise there's no fair way to test different options. The problem is maybe at least twofold. And so one is, is the benchmark really addressing the task? So, for example, if I want to, if I want to do, like, is this model good at biology is just asking questions about what does a mitochondria do? Like, does. Does like a eukaryotic cell have, like a cell wall or like these kinds of facts? Is, is that like, the right way of mastering of understanding if a model understands biology? Or is it. It has to be able to design an experiment that can achieve a certain goal, or it has to be able to like write a cohesive research paper. Sort of like unclear to some extent. And then the second order issue is to what extent are the very evaluations, the very benchmarks that we are using to test our models available on the Internet or in otherwise in other forms in the training data? Because almost by definition we're like learning from textbooks, we're learning from quizzes. And even if the benchmark quiz that you're using to evaluate is not present in the training data, maybe something that looks like it is, because the same human wrote it, to use a trivial example. And so then you have too much similarity between your training set and the test set. So you're effectively like what nerds call like benchmark. Maxing.
- [97:34]
A
Yep. Just like manipulating and making sure you just beat the thing specifically that you're going to get tested on. Well, it's like studying for a test. It's like the whole, are you actually smart or did you just memorize what was needed for the test? And you got all A's, but like you, you know, you're book smart but not street smart in a sense.
- [97:49]
B
Yeah, exactly, exactly. And then there's there's things like more recently in code, you know, there's a popular benchmark like, called SUI bench standing for software engineering bench. And then verified to make sure the actual evaluations are human verified. And it's mostly like bug fixing in Python and unit tests and some models perform super well on it. But software engineering is far more than unit tests and bug fixes in Python. It's like a lot of other things. And just that, you know, we probably haven't gotten around to writing all the evaluations for the litany of tasks that we want models to be good at. And so I think they're an important litmus test to look at. But know with a grain of salt that companies are doing as much as they can to do well on these tables. Because everybody else looks at these tables.
- [98:46]
A
Yeah, it's like the Olympics. Your point where like, when I think of like Canada Extreme or like Norway, like extremely overachieves in the Winter Olympics, like Canada usually is like top three. I feel like Norway is usually in like the top five. But like, it's Norway. There's like, you know, a couple million people that live there. Like, how are they beating? Like they're beating China in the Winter
- [99:07]
B
Olympics or random skills like the, the Turkish guy who like won the pistol competition.
- [99:13]
A
Oh yeah, yeah. It's like that's incredible marketing for Turkey. Right? It's like this insane sharpshooter.
- [99:18]
B
Yeah. Slightly scarier. They have a pretty crazy defense industrial base these days that may or may not be selling to shoddy nations.
- [99:25]
A
But yeah, yeah, fair, fair enough. What up? So one thing you, you kind of mentioned I want to talk to you about. Do you need to train your own AI model to build an AI company? Like, is that a necessary thing to do or can you get away without doing that?
- [99:42]
B
I don't think it's necessary. The way I look at it is like, what's the problem you're solving? And if that might sound super basic, but I think a lot of truisms are kind of basic. And if what you are solving is sufficiently workable with an existing model, then I think honestly, like, congratulations, like you can now build a SaaS company. You can now build an AI company at the speed of a SaaS company because you don't have to faff around with all the nuances of making an AI model work. And you know, for so long as the unit economics are fine, then I would just run with it and, and, and yeah, use all that budget that you're no longer spending on R and D on product market fit and growth. But if what you're doing for, if the problem you're solving is not workable with current systems and you're going to have to do your own stuff. So to state like an obvious example in the sciences, you know, if you're designing crisprs and genome editors, like you can't really ask ChatGPT for that yet. Like it can probably blag its way, but it's not going to be that great because it hasn't been trained for it and there's some like architectural nuances that maybe are less suited to the task. And then relatedly, if you're doing like genometers, you kind of don't care about its ability to make pictures or jointly learn audio or jointly generate video. It's just like not relevant. But, but like if you want to compete in the, in the model race, then yeah, join, join the club. You gotta start from scratch. And there are, there are some that have started like you know, Deep Seek in China or Poolside or you know, or xai.
- [101:27]
A
Well, so, but I feel like another element to that question is like, okay, you just used the OpenAI model to build this thing and like, oh Nathan, that's a cool company. Like you just use the OpenAI model. Like I'm going to do the same thing. Like I'm also going to use their model and build the same thing like is there any defensibility if you don't
- [101:44]
B
have your own model, product experience, taste and then user data which people would call nowadays like preference data like is this good or is this not good? I think that's what it comes down to in some ways like in some ways competition is you have to do whatever it takes to like such that your user, your customer either doesn't fire you because like who gets fired for buying McKinsey basically or IBM kind of thing in the past or B in a consumer space that you have equivalence with like a Coca Cola. And I think you know OpenAI to me is like Coca Cola. You know 90% of the time you're going to use that and if it's not available on the flight Malik maybe you'll buy a Pepsi but you'll think about it. I don't really know what model is Pepsi yet. And, and then for, for enterprise yeah it's like the don't get, don't get fired for buying this product and then do whatever it takes to get there. I think once you are there then then there's you know different ways of, of of like having moats long term but in the short term yeah it's product quality I'd say.
- [103:06]
A
So then where do you think most of the value ultimately accrues? Like should you just buy Nvidia and just like that's your AI exposure Like
- [103:13]
B
what's I think all the things honestly like I worry a little bit that like this, this point has been like propagated or contrived a little bit by VC blog posts that like to like pontificate about the future of an industry. Not, not dissimilar to how they pontificated about like the modern data stack and all these tools that were needed and I think all those investments in basically gone to zero. So I, I, I do wonder what the I do, I do worry a bit with this like where does the value accrue Because I think it's kind of everywhere to be honest and clearly like Nvidia long term but also its suppliers also like in, in energy and and like some neo clouds I think are compelling model companies themselves but also like product companies that wrap on model
- [104:02]
A
like you're doing a unique thing for a customer and building a deep relationship with them where you get, you have like a proprietary relationship with the customer serving them in some way like no matter what you're doing like is that maybe just like a pretty simple way to Think about it.
- [104:18]
B
I think in the early days, yeah, for sure. Because at that point the customer has to trust in the journey and the relationship they're going to have with you, that they're going to invest some time and you know, use a slightly janky product today because it's going to get better in the future. And then you'll listen more to the feedback that they have so that you can just, they can, you know, purchase a product that's fitter for their, for their needs. I think, you know, I think many of the, many of like the best, like enterprise software companies that I've invested in, when I did like, customer diligence on them in the early days, it was something along the lines of they really intimately understand our problems and what it takes to build a solution for us. They listen to our feedback, they're super fast on fixing any problems and getting back to us, and they care. And then on the flip side, enterprises I've been part of, where the customers churned and then the founders fly there and then they're like, hey, trying to rescue the relationship. Usually the relationship owner is like, thank you for coming. This means a lot to us because you're here. We will resign.
- [105:31]
A
So it's really just like caring about the customers, like give like a good customer experience.
- [105:36]
B
Yeah, yeah. If your thing is like a little bit hard to use and then, or maybe it's like super easy to use and then you never have to talk to a human. But I think anytime your customer account value like goes up above like 50 or 100k, you're probably going to have to talk to a human.
- [105:51]
A
I think one thing you kind of alluded to earlier that I wanted to ask you about was you think there's a big opportunity in defense and specifically in Europe. What's kind of the thesis behind that,
- [106:06]
B
the general thesis there is that Europe hasn't been investing in its industrial base for a very, very long time, basically since the end of World War II. I mean, a bit similar to how the US has been a broad consolidation of primes since, since then in the Cold War, sort of like termed the peace dividend. And yeah, in general, you know, Europe has like forgotten that, you know, war can happen on, on its doorstep ever. And the other problem is that in the Munich Security conference last year, which is February, 12 months ago, that was when J.D. vance and others basically dropped the bomb that, you know, U.S. security guarantees might not be what they were up till today. And that means basically like the US is not going to subsidize Europe's defense. This was like a big shock to basically everybody on the continent. You know, this is several years after the start of the Ukraine war when you know, the US has been sending, you know, more aid and more military equipment than pretty much anybody. And, and then you know, ensuing that, you know, Trump pushed really hard on European nations to beef up their percent of GDP spent on defense. Many countries were below 2%. The highest was probably Poland. It was close to three and a half or four. Some countries don't even feel the need to spend more on defense. Like Spain doesn't want to spend more than two. Apparently doesn't want to.
- [107:47]
A
You're the furthest away from the threat.
- [107:50]
B
I think that's part, part, part, part of the reason. Yeah, they just feel far away. But that's the danger, dangerous concept of this problem is far away. It doesn't affect me. Like it sure as hell affects you in like energy prices and migration and integration and social issues and, and then potentially, you know, if you're, if you're part of the European Union and NATO, then you do have to contribute forces to potential like troop deployments. So it's a bit, it's a bit short sighted, I would say. And then there was, you know, post that of like holy shit, like no one's going to come to save us if we have big problems. Like certainly daddy, America is not necessarily around anymore. Then European Commission spun up a big initiative around mobilizing additional funding for European defense where they said each nation is allowed to increase its defense spend up to a certain amount to mobilize in total about 800 billion euros. Now this is not contractually required. It's a bit like school teacher telling children you're allowed to spend your own money on this new candy if you so wish. I suggest that you do, but you don't have to. Hence like the ability of Spain to sort of opt out. And then NATO setting targets of, you know, 5% of GDP spent on defense. And then a new instrument called the safe, not the YC safe, but security and something, something for Europe which is 150 billion of money backed by the European balance sheet to do advanced procurement of military equipment when at least two European nations have sponsored like the desire to buy said product and that product has to be made in. And then, and then certain like fracturing of, of, of European procurement for US equipment. Like there's, you know, certain countries like I think it was Portugal saying, hey, we're not going to buy F35s. Some other nations saying, we're not going to buy F35. We're not going to buy the Patriot missile defense system or Switzerland, you know, has, has come out saying, you know, we have 6 billion and a half of, of loans that were approved to buy F35s and we're going to buy as many as we possibly can. This is a country that's been neutral for basically forever. Sweden was a recent joiner in NATO, has been neutral forever and now is like radically remilitarizing Germany. I mean the, you know, who would have imagined, you know, decades ago that the Germany would ever like remilitarize inconceivable and now is like, you know, probably the biggest defense spender in, in Europe after you know, the Chancellor basically enabled the, the country to take on significant more debt than what it was allowed to in the past after Merkel's government which basically crippled infrastructure and defense spend in the country. So some pretty massive like macro tailwinds. If like VCs are looking for massive tailwinds to motivate investments like these are pretty fucking huge. And then there's the qualitative aspect of things where entrepreneurs are no longer afraid to build in defense. It was before deemed to be pretty taboo. Now there's money to be had and investors have suddenly woken up to okay, we're happy funding this stuff. It wasn't even just a year ago that there's a lot of chatter of Europeans investing in defense, but not that many that were actually doing it. So, so yeah, it's really like there's, there's like no time to waste. The biggest enemy of everybody is basically time to like rearm and to have capabilities sufficiently large that they deter your opponent.
- [111:48]
A
And there's no, and the idea is that it's, there's no existing like European domiciled providers or there's just like very few. So there's just like there's an opportunity to build more.
- [112:01]
B
Yeah, I mean there, there are you know like Ryan Natal is probably one of the biggest. It's in Germany. You know, it provides a lot of different equipment where it's you know, anti drone, you know, tanks, other things. That company's worth more than Volkswagen at this point. So it's like close to a hundred billion market cap, you know, has ripped a bit like Nvidia has. There is, you know one of the biggest like ammunition makers is a Czech family owned business for many, many years. It's allegedly today expressed interest in filing for an IPO in Amsterdam. It's going to be worth around 30 billion euros, obviously. Like the French have a lot of primes like Telus, like naval. So aviation, you know, Rolls Royce in the uk, BAE systems, you know, Leonardo that helps make some of these like missiles and aircraft. Saab, that's famous for making the Gripen, which is like a sort of pseudo alternative to the F35. There's definitely an industrial base. It's just been like kind of sleepy.
- [113:10]
A
Interesting. And you think that there's still opportunities for startups to kind of emerge. Do you need to like, what's like the approach you would take? Do you have to come up with like a new product that doesn't. The existing guys don't have, or are you just, you know, quote unquote AI native or something like.
- [113:24]
B
Yeah, it's pretty much the same. The same narrative that has played out in the U.S. which is, you know, like lots and lots of products that are autonomous, that are each cheap and potentially disposable. So, you know, in, in Ukraine there's like hundreds of different drone makers for different applications. Whether it's like surveying called ISR or it's for like strike drones, where it's basically a drone is a missile effectively. And then you have different sort of, you know, environments that matter. So these are aerial systems and you have land autonomy, like reconnaissance land drones or demining drones or. Or like logistics land systems. And you have on the water, so like sort of strike boats or boats that can transport material or boats from which you can launch drones. Same thing with underwater. And all these areas, particularly in autonomy, have been just like underinvested or just not been a focus for a long time.
- [114:34]
A
Yeah, kind of. It seems like if you just think about the evolution of a war over time, eventually we created metal and you could have a sword instead of a stick. And then horses. We introduced the horse and that changed. And then we introduced guns and planes. And now autonomous is almost a new era of it. And everything needs to be able to defend against that new plane that's kind of been created.
- [115:05]
B
Yep. Yeah. And you need to protect almost against like the lowest, lowest common denominator of like stupidity. Like what is the dumbest person that has access to these weapons gonna do?
- [115:17]
A
One of the. One of prior guests of the show, his name is Rahul Sidhu, he has a company called Aerodome that's well to flock safety and he's. It was basically like 911 response. Drones, like you call 911 drone immediately goes up and just within 60 seconds it's there. And it like basically tells the police what's going on even before the officers get there. But he's. He's kind of. He's like, I'm super surprised that we have not had any cases yet with just, like, lowest common denominator of, like, the worst case you can imagine that could happen with a drone. And, you know, you can just imagine of, like, using a drone to cause chaos, basically, in. In a country.
- [115:56]
B
Well, we sort of seen that in the last quarter in Denmark where, yeah, around, like, the Copenhagen airport. Some unnamed or undisclosed, like, organization people, whatever, were flying drones in. In the airspace. People couldn't figure out who it was. It shut down the airport for several hours and just showed like, oh, well, there's holes in our air defense systems against these rogue drones. And people obviously suspect it's like, Russian and. And then, like, even more serious incursions of, you know, like, Russian jets into Polish aerospace. That was, like, pretty egregious a couple of months ago. And, you know, at the time, there was no. No intervention because it's like this mix of. I mean, the politics is, like, very difficult, but there's a bit of, like, the talking game of, like, you know, we will. We will defend ourselves. We will defend, like, all our NATO allies in case of incursions. But, like, when a jet flies through and, like, threatens you, you don't shoot it. So at some point, I don't know, people are gonna have to decide, should we actually shoot it? Because otherwise the enemy's not gonna think that you're serious. Or do you. Or do you, like, keep going and sort of, yeah, do the policy talk, but not the action. And then there's. And then there's like, you know, like, voter approval ratings for military engagement. For example, a couple months ago in France, there was a general that said, this might be the first time, some version of this might be the first time in recent history when we would have had to have our children die on the front lines of a combat. And while there's potentially good support for, yeah, we should be arming Ukraine against an aggressor, it's different when it's like, send your kids there. It's a country that you've never been to, you've maybe never met somebody for. You don't understand, like, how this directly affects you beyond, like, yeah, we should be, you know, defending democratic nations that get attacked. And like, in a European continent where, like, yes, every country is part of the same, like, economic union, are they really, like, are they really the same? Like, does somebody from Portugal really have affinity for Someone in Finland, because they're both part of the same continent.
- [118:11]
A
Yeah, it's super fascinating with, like, the United States of America and Europe. Like, the US is like 50 different states and someone in. I live in Michigan, someone in Oregon. Like, are we really that different or that similar? I mean, we kind of actually are. Like, it's interesting how the US Is just. The way, like, we've evolved is we basically all evolved as one. Like, we've always been one nation. Like, I couldn't imagine if Oregon was a different country. And suddenly, like 20 years ago, people were just telling me, like, you should care about Oregon or something.
- [118:50]
B
Yeah, exactly. I mean, like, the US has like a singular. I mean, you know, historically, like, has a singular dream of everybody's there for the same reasons. And there's some principles that, generally speaking, most Americans, like, abide by. There's the same language, the same cultural heritage. Like, everybody knows what Independence Day is. Everybody knows what the slave trade was. Every knows, you know, the. These other events. But in Europe, it's like a hodgepodge of. Of history that is like, incredibly complicated. That. I mean, we. We all get educated about, like, first and Second World War, but, like, you don't know too much about, like, nuances of crises that, you know, have shaped generations and like, you know, Czech Republic or. Or other, like, countries that are maybe too far afield for you. So it's hard to feel that kind of unison, I think, when actually the most important test comes to bear. And that's like, would you put your life on the line?
- [119:46]
A
Yeah, well, and it's interesting the episode's releasing tomorrow, but by the time this comes, when we're recording, by the time this comes out, this will have been about a week or two with Marcelo Labre, the founder of Remote.com, he had a pretty interesting description of. The European Union is basically a regulatory body, really, at the end of the day. So the purpose of it is just to regulate. That's kind of what it does. It's a pretty good kind of ten minute, almost monologue. I'm going to post it on Twitter. Just here's 10 minutes on Europe's regulation problem, basically. But yeah, it's like, it just kind
- [120:23]
B
of
- [120:25]
A
is kind of like, we need to have a thing that unites us. Let's make this thing that gives us a consistent purpose of being. And it's just. It's not a nation. There's kind of shared beliefs and cultural values, but it's really just like regulations that we kind of do to Subsidize things and make people follow rules that maybe one of us wants but a different nation doesn't want. It's kind of fascinating.
- [120:53]
B
Yeah, I mean there's some benefits, like making trade a lot easier, making travel a lot easier. You don't have to cross like immigration for every country you cross and then like a single currency system which was beneficial for a lot of like tier 2 or tier 3 European countries, you know, but, but yeah, like uniting all these different interests. Very, very tough.
- [121:20]
A
Yeah. And so maybe slightly different topic, but I'm curious, your kind of personal AI stack, like what do you use on a daily basis? Which is kind of interesting. I don't ask this question all the time, but I feel like you might have a pretty interesting. Maybe you have the most researched most like maybe you don't. Maybe it's just chatgpt. Like what does your personal AI stack look like?
- [121:44]
B
Yeah, yeah, I'm like heavily invested in ChatGPT and, and, and mostly that over like Claude because I mean I don't, my day job's not coding and web search and research is really important and, and OpenAI was like earlier to building out that capability. I'm very, very, I could document pretty much everything. So also for, for calls I try to use ChatGPT or like an alternative to it to generate transcript and then pipe that into ChatGPT. So then, then I have like my.
- [122:19]
A
Your second brain almost.
- [122:21]
B
Yeah, like just company history in there. Who did I meet? What did we talk about? What are the things I need to follow up on?
- [122:27]
A
What do you use to record and transcribe?
- [122:29]
B
Most of the time I just use a native chatgpt recorder.
- [122:32]
A
Do you join a zoom call or it's like an in person meeting and you just press a button?
- [122:37]
B
Yeah, it bugs out a decent amount and it's incredibly slow to do the transcription. So if you have back to back meetings, it's not really possible. You'll like, it'll take 10 minutes or something. So then I'll use like a, a local version like MacWhisperer or sometimes I'll use granola and then I'll just like copy paste the entire transcript into ChatGPT versus like consume the notes. And this is like useful for, for like startup meetings because you know, if you have a bunch of meetings on the same topic with different companies then you can like compare and contrast, build a richer picture. It's also useful for like investment memo writing because then you have the entire corpus of like all the interactions you've had with the company either through like, you know, recorded meetings or then like a Q and a doc that I do with entrepreneurs a bunch of times on, on their, you know, plan and then the assets that they produce. And then I have like my template memo. And then it's like pretty trivial to be like, here's the memo, here's all the material. Can you help me? I can do 80% of the work. And, and then like the more advanced like Excel functions it's been really good at. So you know, doing cohort analysis, like finding outliers in like financial data. Very useful.
- [123:52]
A
What do you use for the spreadsheet
- [123:54]
B
stuff most of the times. So there's like two use cases, I guess. Like for, for financial, like analysis, customer core analysis, things like that. I'll use ChatGPT. But for other kinds of tasks which involve tabular data, I'll use like our portfolio company V7 because they're like. I can have like a table, for example, like a list of leads of like companies or founder profiles.
- [124:19]
A
Is this V7 Labs?
- [124:22]
B
Yeah. Dot com. Okay.
- [124:24]
A
I'll throw a link for people in the description.
- [124:26]
B
Yeah. And then I just upload the data into, into the V7 product which looks like a table, except every column I can do different things. So I can call a certain model with a certain prompt. I can do a web search. I can have like a little Python function. I can have a categorizer. So I can basically like implement all the like formal reasoning logic I would do when I'm looking at like a sheet with Turner. Started a new company and used to work at Google on this thing and was educated there, like, and is working in Timbuktu like does that pass my filter of wanting to talk to him? But if I have thousands of these people like every year, I can just smash it through V7 with reasoning models and then it'll produce really good outputs that are basically the same decisions that I would have made.
- [125:10]
A
Interesting.
- [125:11]
B
So very useful. And then what else? Yeah, for audio generation I use 11 labs for sure.
- [125:22]
A
This is for Air Street Press.
- [125:24]
B
Yeah. Like if you want to use. I used to like actually sit there and read it out. But it's pretty time consuming especially if like building works in your, in your apartment or like dog barking or zan outside or whatever. Yeah. So I use that. It's great. And most people like don't care slash can't tell the difference. What else? I'm looking at my application tray. I feel like that, that's like most of the use cases, to be honest. I've done some like, some like OpenAI Atlas, like Web Automations and I would use more. The problem is I think it's been too safety guard railed. So it'll like ask for approval every single time. It'll like do a task. It's like, you know, I want to send this message to somebody. It's like, yeah, is this okay? Like, yes, stop asking me and it'll ask you again. And then just like jarring. I think it, it, it also has evolved very, very well, it being like ChatGPT for understanding research papers and like digging into adjacencies. I think honestly, because anything that AI research folks like, like and want to do the thing will be good at and so understanding AI research, it's going to be good at.
- [126:47]
A
That's fair. Yeah. I mean they're building the product for themselves probably pretty much.
- [126:52]
B
So you're sort of immune. If you're building something that like AI research people think is like a tier 2 or tier 3 problem, you're safe,
- [127:01]
A
interesting, like plumbers, like AI for plumbing, like great, great categories for the company. Like no AI researcher will ever like want to build those capabilities.
- [127:09]
B
No, but it's like, it's like the workday example of, you know, the CEO getting on that earnings call not too long ago and some analysts asking him like, are you at a, at a thread of OpenAI or anthropic or something, you know, rebuilding your product. And I think his answer was like, they're actually our biggest customers. They don't want to, they don't want to rebuild workday. Like you don't, you don't go to OpenAI to go build AI Workday.
- [127:32]
A
Yeah, it's like the most uninspiring thing in the world.
- [127:36]
B
Yeah, yeah. It's like anything, it's very stretched. But I think like anything that's not generality is like a low class problem. Hmm.
- [127:46]
A
So in the sense it's like the more vertical, specific, like the more niche the thing you're doing is, the safer you are from the big platforms, maybe the better a startup category it is in a sense, even though you're niching yourself down. But that's kind of like the classic, I mean that's been true for decades. Just find a really specific, unique problem. Like 11 labs, you could say audio models, like that's not a thing. And then now it's like, oh, like we were talking about earlier, like hundreds of millions of revenue getting added every year, 15 million in a day, like,
- [128:18]
B
but then there's also arbitrage around, like, how much risk are you willing to take? And I mean these are my thoughts. Not, not the company, but you know, there at the time, you know, OpenAI was under like immense like regulatory pressure. And you know, there's a lot of touring, you know, with nation states and there was big issues around copyright, like, where is this data coming from? And so I think like the, and then there was also the like, did they copy Scarlett Johansson's voice or not? So I think the biggest PR disaster could be like, I don't know, OpenAI launches like audio model and somebody like uses it for some like weird task or wreaks havoc or something. And so if, if you're like an independent company that's willing to take some of those risks and be really thoughtful around doing it well, then you can arb the fact that your competitor is too scared or is distracted with other things or can't afford it.
- [129:12]
A
That's like a. As a whole, Google should have won this all really at the end of the day. And there's too much risk not only in the product, but also the business model. And then they're still slow. I mean they've been in slowly adding more of like, you know, the AI mode over search results. But still it's like, how do you monetize it? That's a big question. So it's like product risk, business model risk. I don't know, whatever. Policy risk maybe is like a big one.
- [129:45]
B
Yeah, it keeps, it keeps asking me to beautify my slides. It's like, stop telling me beautify my slides.
- [129:52]
A
Well, and it's interesting too. Like there are times where like I don't want AI honestly like having like an email provider like, or like, like a text message. Like if I'm texting you and I'm just like, hey, are you jumping on if there's like a pop up that's like, hey, it looks like you're sending Nathan a message to join the podcast. Would you like to make this like a more comforting and like, you know, like friendly interaction instead of just like, you're very direct. It's like, are you getting on? Like, I don't want AI there. So there's almost like a risk of, you know, adding AI features that people don't actually want that make the product harder to use in a sense.
- [130:27]
B
Yeah, but if it said like, you know, here's the link handy, that's true.
- [130:32]
A
Yeah, that would be helpful. But then it's like building the scaffolding around the product of like knowing when to introduce that kind of capability.
- [130:40]
B
Yeah. And then also taste like, how does the model know when it's the right time and what is genuinely useful? That's. That's hard. I had this interesting experience that I've been telling some AI friends about where I was using like the ChatGPT, sorry, OpenAI Atlas browser to do this automation task on substack because I wanted to rename the naming convection or various articles I've written. And substack doesn't allow you to like mass rename. Like you can do it on your desktop finder should have to go link
- [131:12]
A
by link, page by page.
- [131:14]
B
Brutal. Including like changing the SEO and the URL slug and the title and, you know, clicking. Okay, so I wrote down the instructions to Atlas so I could go do that. And like the first batch, like, it did it really well. It like worked autonomously for 45 minutes, corrected a bunch. Then my credits ran out because I haven't bought the 200 bucks version yet, which some friends of mine think I'm really stupid for not doing, but is what it is. And then the next month I opened the exact same chat and said, hey, can you continue? Then it continued for like 20 minutes and then it exhausted my credits. And then the third month I tried it again and it's like some version of like, hey, like, this task is like really manual and requires like a lot of, a lot of clicks and it's going to take a significant amount of time. And it was basically some version of like, I don't want to do it. Like, it, it didn't, it didn't do the task. Yeah, I was telling some friends, like, why is it like refusing to do this task? And, and some, some people would say like, oh, because you're restarting the same session. Like it has like all this like crazy amount of click data and screenshots and whatever that it's like overloaded its context and has basically navigated outside of the in training distribute. I know there might be another example of some human annotators have said this kind of task is really manual and it's regurgitating and refusing. This is crazy. The very automation I want to get this thing to do, it doesn't want to do.
- [132:47]
A
Yeah, well, I had an interesting example where I can't remember what it was. I think I was trying to get it to clean up a transcript from the podcast or something like that, where I was just like, hey, can you just do this thing? And it took like, it was thinking for a long time, like an hour. And I kind of kind of came back and I was like, you know, how is this going? Yeah, it was like when you're texting an intern, like, hey, that thing you're working on, like, how's it going? And I was like, I'm still thinking. And it was like the next day I came back and I was just like, did you finish this thing yet? And it said no. Like, I stopped working on it. I'm like, what the heck? It's like, wait, would you have like an intern that just like doesn't do the work? And you're like, hey, like the thing I gave you to do because you're like, you work for me. Like, did you do it yet? Like, no, I didn't feel like doing it. It was too hard. Like, what? Like, let's. What am I? Like, you're, you're AI. Like you're software. Like, you just, just go, just do it.
- [133:37]
B
Yeah, I mean, man, like this is like the mystery of this high dimensional box that we're poking with a stick to go do certain things. And sometimes you just. The stick poking is not good enough.
- [133:47]
A
Yeah, well, and then sometimes it just like does this insane. Like I think probably my chatgpt usage like doubled when it came up with image generation. Like, and it just did it right in there and you just kind of like say, make this thing for me. And it's like pretty good. I think this was probably like a year ago when they first came out with it and I was like, damn, this is like really, really impressive. Like this is like super helpful and useful. So there's like two ends of the spectrum where it's like it refuses to do something that just is pretty simple. And then the other end is just like gives you this amazing product that you would have otherwise spent a ton of time on or like didn't even think was possible. So it's just kind of like amazing, the technology.
- [134:26]
B
Yeah, yeah. I think the TLDR is just force yourself to use it and develop the behavior because this is not going away. And so there's like arb to figuring out how to extract the most use out of it.
- [134:39]
A
Yeah, it's just going to keep getting better.
- [134:41]
B
Yeah. And it is really mad that, you know, the vast majority of the population is like using ChatGPT as if it is like a thing. Like it's a person with, with like one opinion. But, but you know, it's like the amalgamation of like every opinion possible on planet Earth. And so like really you should be telling it like, hey, you are like an expert Car dealer. What should I think about when I'm buying a new car? Not, not just naked asking it, what should I think about when buying a new car? I don't know, like, based on whose opinion?
- [135:12]
A
Yeah, because you can tell it. You are an expert car buyer. You've spent millions of hours researching every possible scenario. You are literally the highest regarded. You buy cars for the president of nations and just the most extreme example you can think of. And it just improves the results because the guidance you gave is crazy.
- [135:37]
B
Yeah, yeah. So maybe at some point we'll, we'll get to the point where the model can infer like what you're, what you had in mind. And you know, most of the time you're asking about a car, like, you should respond as like somebody who knows about car dealerships and stuff, not like some noob.
- [135:54]
A
Yeah. It's like, hey, you're my crazy uncle. Like, give me like a bad recommendation for a car. It's like, okay, here you go.
- [136:01]
B
Yeah, exactly. You might want that if you want to, like, message your goals.
- [136:05]
A
I had one last question for you. It's from our mutual friend Dan Feder. He was curious. I know you're really big into tennis. Who's better, Federer or Nadal?
- [136:16]
B
It's a possible question. I mean, I think there's nobody who's played prettier, more classy, tennis, effortless way than Roger Federer for sure. But I would say that like ruff and adult are like. And the AI analogy expanded the frontiers of the style of the sport. Tennis has often been taught in a very textbook way of your swing has to look this way, you have to put your front foot first. You can't hit open stance. And Nadal just threw all that shit out the window and was like, you know what, I'm just going to play how I feel naturally. So things like open stance and swinging with your racket over your head and in a full circle, a lasso. And those two things have like, meaningfully, I think, changed, changed the game. And you can see like the top players, like Carlos Alcaraz, which is sort of his Mini Me, plays in a very, very similar style. A lot of like top players are playing open stance. So I think he kind of, yeah, he definitely pushed the technique and style game to another level. And then Roger Federer just like executed the textbook in the most beautiful, elegant way as possible. And as a fun fact, neither of them have ever smashed a tennis racket in their entire professional career, which is pretty epic.
- [137:40]
A
Like you're saying, when you get mad and you slam it and break it.
- [137:42]
B
Yeah. They've never thrown their racket. Yeah.
- [137:46]
A
So this is just like a self control thing, like an emotional control that they have or.
- [137:50]
B
Yeah. In particular, Federer was like, very, very known for this. A bit like Bjorn Borg, who is this, like a Nordic. It was kind of nicknamed Iceman because he would show no emotion whatsoever. Roger Federer was always very, very little emotion. Always very focused. A bit like Yannick Sinner today where like the peak of his derangement is like tipping over his water bottle when he's sitting down. You know, it's so funny. Like Instagram beeps about that.
- [138:17]
A
Oh, I didn't, I didn't say that. Have you seen that one kid who can use forehand with both hands? He like switches his racket. Is that real? Is that like a sustainable thing?
- [138:28]
B
It's not sustainable at a high level. I think the game moves way, way too fast to have the time to do that. There was at my time as a teenager playing tennis on the women's circuit, there's a lady called Monica Selles who was playing with two hands on both sides.
- [138:45]
A
I think I remember that name.
- [138:46]
B
I think it was her. There was definitely one player who was doing that. But yeah, it's kind of Frankensteinian.
- [138:54]
A
Makes no sense because to the point one of my friends who played in high school, he said it's sort of like playing speed chess where you're doing this tactical, strategical things, but you have no time to make any decisions and you probably even the function of switching the tennis racket in your hands, you lose half a second doing that and you need to be able to move even quicker than even.
- [139:17]
B
Yeah, it's one of those things where you need to like, develop this insane muscle memory that your body just reacts naturally in certain scenarios because you've, you've tested it so many different times and. And then you have this like, natural intuition of like, what you should do. A bit like taste, I think. But it is, it is. I mean, I'm sure like a lot of professional sports are like this, but like the moment you get distracted and think about something else, I'm not sure professional tennis players have this distraction at all. But like, as a casual, like, pseudo, like, protest player in the past, like, if I think about like this deal or something, like, I'm so toast.
- [139:56]
A
Yeah, like, you got to be constantly focused on, like positioning even. Like, I think about a lot with, like, with hockey, like I play hockey, and it's just like a big piece is just like, where are you standing? And like, how open are you and are you like, you know, if, if something were to happen, like, how are you positioned to react to it? It's probably, that's probably even more common in team sports because, like tennis, you're. You're always the one that's doing the action. But if you're like playing soccer, you. You could theoretically play a game of soccer without touching the ball, but you impact the game based on how you, how you move around the field.
- [140:31]
B
Yeah. At least in tennis, we're always looking in the same direction. Generally the problems in fighter, even most of the times. But yeah, like, I think nowadays the speed of the game is insane. The reaction times you need to have the, like, the power you get out of the rackets and the athleticism the athletes have. It's. It's very scary. I mean, the stretches that they're doing and the fact that they're sliding on hard courts, it's like these kids are gonna have all sorts of like, bodily damage by the time they reach the old age of 30. It's. And it's kind of scary.
- [141:04]
A
Maybe we'll have AI generated training routines or like, you know, physical therapy to keep them healthy or surgery to fix the damage.
- [141:15]
B
Yeah, yeah. I mean, that's going to be like regenerative medicine and stem cell biology, the sort of industry I came from many years ago. It'll hopefully help.
- [141:23]
A
Yeah. Well, this is a lot of fun. I know we were kind of talking for a while, but hopefully people learned a lot. Where can they kind of like find you, like, what's like your. I think Twitter, you're pretty active on. You write the new the Airstreak Press, which is, I think pretty frequent. You send things out. What can people look up?
- [141:42]
B
Yeah, Twitter just Nathan Benesh, I'm pretty active on. And then. And then Airstream Press, which is just press.sg.com. yeah, those are my two favorite outlets at the moment.
- [141:53]
A
And then do you send the State of AI report every like from the Airstream Press or is it like a separate URL?
- [141:59]
B
Yep, yep, it's just that URL. But you can see all the prior editions on State of AI or stateofai.com as well. The like eight years that came before that as a Google Slides and then I do like monthly newsletter also State of AI newsletter, which is the renaming convention as hacking away with my AI. And that's available on a. On our shoe press as well.
- [142:21]
A
Okay, cool. We'll throw all those links too in the description for people to check out. Cool. But yeah, thanks. For doing this. This is a lot of fun. Thanks Turner and I hope you had fun too. Thanks again to Numero and Flex for supporting this episode. Put your sales tax on autopilot@numeral.com and upgrade to Flex Elite to get $1,000 on your first card using code Turner at the Waitlist link in the description. If you like this conversation, please like comment. Subscribe and name your next industry report after me. If you missed it, make sure to check out last week's episode with Marcelo Lebra at Remote on Building the Unicorn in Europe and tune in over the next few weeks for guests include Gary Tan at YC Chatham, Putagunta at Benchmark, Jake Stoch at Serval, and Duo Security co founders Doug Song and John Overhyde. If you don't want to miss any of these episodes, subscribe to my newsletter the Split linked in the description to get each one plus the transcript emailed directly to your inbox every week. Thanks Sam for listening. See you next time.