Tomasz Tunguz of Theory Ventures: The Impact of Generative AI on Software

the seed, preempting the A, and then participating meaningfully in the B. And so each company has like 20 to 35 in it.

Got it. And how has— now that, I mean, obviously you've been investing for how long, a decade or so, maybe a little bit longer?

Yeah, about 15 years.

So now it's a very interesting time to debut your own fund in 2023. I know you've recently written about the declines in later stage funding and everything, but if you had to summarize it for the audience, now that you've got your own $230 million to deploy, What's— throw away 2021 for a minute. Let's pretend it never happened. What's, what's different about 2023 compared to your history of investing? Like, what's the flavor? What's the tenor? Capital efficiency, everything. What's, what's really the pulse of venture right now with your brand new fund?

Yeah, well, so we're looking for companies at a high level, take technology innovations and turn them into go-to-market advantages. That's really what we're looking for. And the main difference in the market between, say, 2021 and today is that efficiency matters more than growth. So the way that the private market is pricing companies, it's as if it were 2017. And that's as true for round sizes as it is for valuations. And it's true for exits. So we're basically back in 2017. And what that means is like, you know, top decile growth is probably looking like a 3 to a 4x instead of like a 5 to a 7x, which is what it looked like Maybe 9 months ago. And, you know, burn multiple, sales efficiency, all those things matter significant. I mean, much more even in the public markets. The highest correlate to forward multiple right now is still sales efficiency. And so that's reflected in the privates too.

And let me ask you, because you've done a lot of great investments in data and others, and there are segments of data that are capital intensive, right? There are companies that have higher burn rate than your typical B2B workflow app, right, that we, you and I could put together. With a couple of great engineers. What— how do you, how do you deal with a company today that needs a couple million dollars a month burn to get past that next level? Are they stuck or how are you thinking about those types of companies, OpenAI aside, that are just more capital intensive?

Yeah, well, one of the reasons I structured the firm the way that I did was to be able to support companies that look like that. Because, I mean, take a look at Samsara, right? Samsara was a high burn company that's now worth about $10 billion. With a completely different go-to-market model. And so there's, there's the Salesforce way, kind of the classic Salesforce way of building a software company. And there are other ways that might be more capital intensive to get more growth. And ideally, you know, you have a founder would have capital partners who are able to support both strategies.

Yep.

Okay. Now one layup question for you because you have a new fund and then we'll get to the presentation. If I'm, if I'm raising money in 2023 for my A or seed, Is it easier to get attention from someone like you with a new fund than someone that's got 20 struggling investments that they're dealing with? Like, who's— it's all over the place in terms of who's really investing right now, isn't it?

It is. Yeah, it's uneven. I mean, you look at the BC market, they're down 80, 90%. The CD market is down 60 to 70%. So funds that don't have portfolios to manage, much like mine, are much more on the hunt for building a portfolio.

Yeah, it is. And so just to founders that are seeing this or watch it later, it's hard to judge from the sidelines who's really in market and who has how much appetite to deploy right now. People are back, I think. No one's scared like last year, right? Everyone wants to fundraise today, but a lot of folks are still sitting on a bunch of headaches, right? And if you've got someone like Tomasz that's got a fresh fund without too much baggage, it's Hunt those ones as a founder at the margin, right? It's hard to tell from the outside what stress VCs are under today, isn't it?

It is. Yeah. There's much more portfolio management than there used to be.

Yeah.

And so the question is, how much time is that taking? And it's unclear. You're absolutely right. It's unclear.

Okay. So I will— I don't want to take up all the time, although we could spend the whole time talking about Theory Ventures, $230 million, data, AI, what's the rest of the focus? Then we'll transition to the portfolio.

What are the core focus? Data, machine learning, and blockchains as databases. So thinking about blockchains as new infrastructure for building apps.

Pretty cool. All right, I'm very excited. I'm very excited to watch you thrive on your own. And it's exciting, but also you got to build some infrastructure too.

Yeah, there's work to do. Great.

Cool.

Thank you.

Thanks so much, Jason. We really appreciate it. And thank you, Tomasz, for diving into all things Theory Ventures. I know we do have another topic to discuss though, the The Impact of Generative AI on Software: Where Opportunity Will Exist for Software Startups. So we can hand it over to your presentation. As a reminder, everyone, I think there are going to be a lot of questions again in the chat. Get your questions in early because I do do my best to ask all of them. I will be asking in the chat if you want to unmute yourself to ask those questions live at the end of the session.

Great.

All right, take it away.

Awesome. Well, my name is Tamash Tengus. I'm founder, general partner at Theory. And I'm here to talk to you about the opportunities for startups with machine learning. So I've worked commercializing machine learning at Google. I was part of a product— I was on the product team there, invested in 3 or 4 different companies, one called Chorus, which was doing speech recognition, another one called Lilt, which is using machine translation. And then the last is a computer vision company called Spot. And we're at a place now where, you know, machine learning has 4 different parts to it, or there are 4 different kinds of machine learning. There's classic classification. So you've got, is this a dog or is this a cat? That's what runs Google's ad network. You have prediction. What is the weather going to be tomorrow? You have natural language processing. So if you're talking to Siri, you're using that technology. And then the last is generative, which is asking a computer to produce an image like the one you see before you. And during the last decade, each of those classes of machine learning have seen meaningful advances. Right now, generative is having its moment. And the big question on venture capitalists' and startups' minds is, how do startups take advantage of what's going on? So that's what I'll be walking through today. So, but to start, one of the most important questions is, does it matter? And the Goldman produced— so I was really curious, like, what is the impact to GDP of generative or machine learning? What could it be? And how would it actually compare to the personal computer, the last major innovation in technology? So Goldman Sachs produced a paper where they said that machine learning will have a 1.5% to 2.9% improvement in GDP. And the personal computer, according to research from the National Bureau of Economic Research, was 0.006%. So just to kind of put it into perspective, it's really the potential here is about 1,000x the improvement to GDP. And it would basically double the growth rate of, of the country. That's how significant of this, this is. And that's why everybody is so excited. And, you know, just to kind of put it into perspective, part of the Goldman Sachs research was looking at the fraction of jobs that could be automated. University of Pennsylvania did a very similar study, but on average, both of these studies find that something like 15% to 25% of all tasks for both blue-collar and white-collar work can be automated. So if you're a startup, you have 4 questions to answer. And they are, if I'm going to be building a startup within this ecosystem, Which layer of the stack am I going to build in? Which market will I pursue? What moat will I create, which is different than in other waves? And then the last is, how deep in the technology do I want to go? So let's talk about the layers. When we think about the software world, classically there are 3 different layers. There's the infrastructure, which is the databases, the networking, the compute. There are the platforms, which are developer tools to simplify building applications. And then there are the applications themselves, the Salesforces, the ServiceNow. And there's opportunity each of these three layers. One of the questions though, if we can take a look at like Web2 or cloud, is which layer actually accrues the most value? And there, there's not enough publicly traded platform companies, but there are enough publicly traded application companies and publicly traded infrastructure companies. So if you take the market cap of Amazon Web Services and Azure and GCP and you put 'em together, you get about $2.1 trillion. And if you take the 100 publicly traded cloud companies, both B2B and B2C, and you put their market caps together, you get roughly the same. So that's Uber, Netflix, Airbnb, Snowflake, ServiceNow, CrowdStrike. And so the observation is, well, the, at the infrastructure layer, you can generate as much value as you can at the application layer. But the major difference is the number of companies. So at the infrastructure layer, you really have significantly more concentration because it takes a lot more money in order to build these models. I mean, classic, classic machine— classic infrastructure takes a lot of money in order to buy servers, rent data centers. The analogy holds with machine learning where The cost to train some of these models is measured in the tens or hundreds of millions of dollars, particularly as they get bigger. Great. So, so let's go through each layer. So at the application layer, this is what's happening today. You've got— my mental model is you've got 3 different categories of companies. You have companies that are helping users create content. So Jasper is helping you write sales emails. Gamma is making presentations. In fact, I get, I typed in like 4 par— 4 sentences in order to, to put together this presentation and it spit out, Gamma spit out a 12-page deck for this presentation that was like 80% of the way there. Synthesia is a company that will listen to people's voices and actually create learning material. So you can learn how to, how to use Adobe Photoshop at your company.

Yeah.

Their designs, there's a company called AI Dungeon, which actually creates text-based games. That are endless. And so the— there's a whole category of content creation. The second category are copilots. So these are app— these are pieces of software that watch you work and then help you. So GitHub, if I want to plot the, the stars— DuckDB, GitHub stars over time, it literally took me maybe 5 seconds in order to do that because I just typed what I needed to do inside of my editor and then GitHub completed it and I ran the code and it worked. Harvey is doing the same thing for legal. If you're writing an NDA, those terms are relatively constant. And so it can autocomplete contracts. Lilt is doing this for translation. And the third category are agents. And this is the newest category. The technology that probably encapsulates it most right now is ChatGPT. But there are many companies that— and I've met some of them— that are building next-generation RPA. There's, there are companies that are trying to be basically the bluebird on your shoulder, remembering everything that you do during a day, being able to interject into a conversation a key data point. And so at the application layer, this is really kind of the state of the art. The, the, the sort of the rub in, in all this is we're basically training our replacements. So a lot of these technologies aim to, to to reduce the amount of rote-level work so that we can focus on a higher level of, of work. So that, that's what's going on at the application layer. At the platform layer, so these are developer tools. The platforms are technologies that help existing companies integrate or use generative AI. And there's basically two different techniques. The first is developer tools and APIs, and then the second are managed enterprise services. So if you are a big company, you, like a Global 2000, you will want to use a managed enterprise service. You have a problem, you wanna use generative AI, you're gonna go to a B2B business who is gonna solve the problem end to end for you. And then in 2 or 3 years, you might use some of the individual developer tools to optimize that stack. But in the short term, what I've learned with these markets is The big companies just want an end-to-end solution. And so there's a handful of companies that are building— like Stable Diffusion— that are building bespoke models for companies to their spec. That's a smaller number. There are a handful of companies now that are building developer tools and APIs to allow you to integrate to help you with things like prompt tuning or getting the most out of the model. In the short term, I think that more of the money will actually aggregate towards the managed enterprise services because you'll be able to— charge a lot more for that value. But in the long term, you might see some very, very big developer tools and API companies. And that parallels what happened in Web2. If you think about the Stripes and the Twilios, the major API businesses, it took about 15 to 20 years in, for internet to develop to get to a place where those companies had enough of a developer base to be able to be big. At the infrastructure layer, there's 3 big opportunities. So One is like the chips and the infrastructure. Google just released Tensor V4, and they have innovated on like optical networking to reduce the cost of training a model by 3x to 4x. So, you know, there's a lot that's happening at the chip level. Cerebras is building bigger and bigger chips in order to train models better. The next layer, the infrastructure, are the models themselves. So there's GPT-3, GPT-3.5, GPT-4. Llama. There's about 50 different models. There's the amount of research that's happening in this category is growing geometrically. So new papers are being published every day. There was this awesome paper. I mean, just to kind of share a bit of what's happening. I'm a swimmer. And when I was learning to swim, I could choose to learn in two different ways. I can choose to read a book. And but if I read a book and then Kaitlyn threw me in the pool, I would drown because I can't just read and learn how to swim. The other way of doing it is I just go into the pool and, and I learn by doing. And what we started to discover with these foundation models is we knew that they could do the first. We knew that they could learn how to take the LSAT by seeing enough LSAT questions and then scoring better and better. But there's this emergent property. So there's a property that happens only as a result of it doing its thing that it's learning. So it's actually learning by doing. So the computer models are learning by doing just the way that a human would learn to swim by swimming. And there are all kinds of downstream implications of that, but it's just to give you a sense of how new and how raw and how much innovation will continue to happen at the foundation model. And then the last part is the training cost reduction. So there are companies that are building application-specific integrated circuits, so dedicated chips. There are companies that are using machine learning models in order to trim these big models. And then, you know, Llama, which is a model that was produced by Facebook, actually is a significantly smaller model with similar performance to ChatGPT-3.5. So at the infrastructure layer, there's a lot happening here. The question for a startup is, can I raise enough money? Can I raise $100, $200, $300, $500 million to be able to compete with some of the big guns? Okay. So the second question for a startup is, what's my market? And the question that everybody, the question that everybody's asking is, is generative AI enough to unseat an incumbent, right? Can you compete with PowerPoint if you've got a generative AI PowerPoint? Or can you go directly against Salesforce if you have a tool that automatically fills in the notes or Word or email or Adobe? And, you know, my experience is I keep all my notes from meeting startups in, in files, and I have about 10,000 of these files. And over the weekend, I used a, a library. In about 3 hours, I had a chatbot running where I could ask questions of my notes. And what that tells me is in order to get to like a 60 to 70% solution, it's, it's actually faster to, to build a generative AI feature into a product than it is to build a mobile app. If you go back to 2008 when Jobs launched the Apple Store, most of the startups were significantly ahead of the incumbents because they figured out how do I write apps in, in Objective-C, how do I distribute the app through the Apple App Store? They were very, very quick and the incumbents were slow. During this phase, startups actually have, you know, what I would call negative time to launch. They're actually behind because if you look at the pace of innovation of Microsoft injecting OpenAI technology into Word, into their design tools, and Salesforce launching a similar technology, Adobe launching a Midjourney competitor. The big companies are moving just as quickly, if not more quickly than the startups. And so if you're a startup, you— I don't believe you can go head to head. You can't because an 80% solution, by Microsoft or one of the incumbents may not be perfect, but it's probably good enough to keep the existing customers on that product knowing that underlying models will improve. And so you really have to go and pursue a market where nobody is really paying attention. And that's why there's a lot of interest in vertical software. And so this raises the question of like moats. Okay, if you're a startup, you decided which layer you're probably gonna build at the application layer, you've decided you probably don't want to compete with an incumbent, So then the question is, OK, there are probably going to be a handful of startups that I will compete with. And the question is, how do I establish a moat? And I think there is, at least for right now, my view is that there's 3 moats. The first moat is this idea, my model is better than yours. So I'm going to develop a more sophisticated machine learning model. It will have better accuracy, lower latency, lower cost, better results. It's possible to do and compete this way, but the performance difference really has to be noticeable. I mean, I remember using Google for the first time compared to Ask Jeeves or any of the other search engines. And that was a case where the machine learning model at Google was significantly better for a bunch of different reasons. And, and that's the reason why they grew so fast. So it is possible to do this. It just requires a lot of technical expertise, an enormous amount of data, and then also a pretty significant amount of dollars to be able to train, train these models. So I think this is a possible moat, but it's not the entire solution. The second mode is— where are we? The second mode is usage. So one of the things, if you study machine learning, one of the things that they teach you is in order to improve the model, you need more data and more usage. When people query Google or when they create content using one of these tools, there's feedback loops. There's ways of telling, Grammarly, I don't think the suggestion is right, or telling Google this is not the right link for this particular search parameter. And the more queries you have, the, the more feedback you have, the better the model will become. This creates and spins a flywheel. And the flywheel is what, you know, the co— will lead you to a, a better model. The, and so it's not just enough to have a better technology. You really need to have the traffic to be able to tune and optimize the model with time. And then the last moat is we're going to see companies that start with public data, which is commodity, build and train models using public data, and then use that as a way of getting access to private data. So, yeah, a medical company could start as a B2C business doing base-level medical scanning and then use that in order to drive a significant user base, and then they'll go to hospital systems and start accessing HIPAA-controlled data. And then that, the combination of better access to data, more usage, and then access to private data, the combination of those three things will probably create the strongest moat in the ecosystem. Okay, and then the, the, the last moat is enterprise readiness. So there are a lot of concerns around these models. I'm not going to talk about the philosophical concerns. But if you're a business, if you're a globally strategic important bank or you're one of the major hospital holding companies and you're looking at this technology, you have a lot of concerns. The first is, if my engineers use Copilot to produce code that infringes on copyright or patented ideas elsewhere, is that illegal? Is that net new or is it actually a violation of copyright? All that case law is being created. The second is privacy. GDPR and CCPA regulate and, you know, HIPAA regulate whether or not you can use personally identifiable information in generative models. What happens if one of your employees accidentally introduces some PII into a big model? Nobody knows how to handle that. This is the same idea with compliance and then security. Imagine there's two different vectors for these models to be insecure. The first is an employee starts to upload a bunch of information, which is intellectual property core to the business, into the model. Now it benefits everybody else using the model. That's an issue. The other challenge with some of these models is DDoS attacks. So feeding the model lots of negative contents. Imagine you're on Netflix's— chatbot, and as a result of some massive DDoS, chatbot now becomes hostile. How do you manage that? So the enterprise readiness, there's a big opportunity in taking some of these base technologies and making sure, you know, that, that they're ready to be consumed by some of these buyers who really will have pretty significant concerns around a bunch of legal and compliance and risk issues. And then the last one is execution. So this is really, startup's advantage has always been execution and speed and focus. The way I think about this is I had this marketing professor in grad school who first day, he was a Belgian guy, he was awesome. First day of school of his class, he wrote this equation which has stuck with me ever since, which is innovation equals invention plus distribution. And what that means is in order to move the world forward, you first need to create something that's new and better And then you need to tell people about it and convince them that they need to use it. Well, with generative AI, we know that the cost of invention is going to zero. It is easier and easier to build a website. It's easier and easier to build a product. It's easier and easier to make a PowerPoint presentation. And so if the cost of invention goes to zero, the idea is that the value of distribution will go up. And this ties back to the idea of more traffic improves the model. And so I think the great part about this moment in time is that there's so much interest in generative AI that if you're a startup and you launch and you attach yourself, you latch onto that wave, lots of people will try the product. And if you can do it the right way, you can take all of that traffic to just reinforce your moat. Great. Okay. So the last of the 4 questions that a startup needs to answer is, is the level of technical depth. So how far within these models do you go? There's 3 different levels. So the first is a plugin. And I— anybody can write a plugin for ChatGPT. I can say, I'm Expedia. Here's a way for users to ask how to book flights. And they've done that. And that's the definition of a single file, maybe 100 lines long. The next level of sophistication is this idea of prompt tuning. So if you remember, I mean, we've all used Google. You can ask Google different questions in order to get different results. You know, like if I construct a particular phrase, in other words, like if I'm looking for economic data, I could type in like site:National Bureau of Economic Research, filetype:CSV. That's going to dump me a whole bunch of files, which is what I'm looking for in a particular domain. That's prompt tuning. And instead of just doing this by hand, a company can actually build a machine learning model on top of the generative model to decide if a user puts in this input, it's actually better formulated this way in order to get the right output. So that's the mid-level of sophistication, building a machine learning model on top of the generative model. And then the last is a bespoke model. I want to build my own large language model. I want to train all the data, put it into a database, tune it, tweak it. I want to own it. And the further down you go in the stack, the more defensibility there is. But again, it costs you a lot more. You need PhDs, mathematicians, statisticians. You also need significantly more infrastructure. And so at the end, you know, I think a lot of companies will probably move down the stack. So use a plugin, use prompt tuning, get a bunch of traffic. And then once they hit a particular level of scale and are able to raise the capital, then reinforce their moat by building a bespoke model. So this is one of the most exciting times in venture that I've seen. I've been doing it for about 15 years. You really have this incredible platform opportunity, and the impact to GDP can be absolutely enormous. And so just to summarize, these are the 4— if you're thinking about joining or starting a company in the space, These are the 4 questions you really have to answer. What layer am I going to innovate in? Which market will I pursue? And if there are incumbents, how will I compete? The third, how am I going to establish a moat? And then the last is, at what level of technical sophistication am I comfortable building? So, you know, the race is on. Good luck. I'm always here, happy to help however I can. My contact information is here. And, uh, I had a lot of fun with this presentation. I spent a bunch of time in Midjourney. All the images you see here are machine generated with one of these systems, and these are the prompts that I used in order to produce them. So they'll be in the slides.

Wow, thank you so much. That was the most like on-brand presentation for what the topic was we've ever seen with all of these like images using AI to create it, etc. So very cool presentation in total. Thank you so much. If you want to stop sharing your slides, we are going to shift to the Q&A session. So we have questions kind of in two departments, AI and then also in Theory Ventures and some of the new work you're doing there. Do you want to start with Theory Ventures?

Sounds great. Yeah, let's do that.

Great. Jason M. Lemkin, you have the first question. Not Lemkin, another one in the line. Thanks.

I'll turn on my— video here, I think. Maybe it's not working. Okay, thanks. So yeah, I just had a question. I get— as a CTO, I get asked to become a technical co-founder of a lot of startups, and I try to talk with them about their valuation, what they're offering. A lot of them, especially if they have an AI and fintech component, their valuation is just so high, even if they just have an idea and they haven't executed on anything. So I'm just curious how you navigate those kind of and advice you have about those situations. [Speaker:JASON M. LEMKIN] Yeah. I mean, so the valuation environment is changing quite a bit, and I'll put this in a historical context using public numbers. So in the last 20 years, the average multiple to value a company has been about 5x. You take the revenue that's projected for next year and you multiply by 5x. That's for public companies that typically grow about 30%. And The, in 2021, that 5x, the high watermark, the top company was at 59. And the private market followed the public market. And we've been investing at companies with like 100x, 200x, 500x AR multiples. And now the market is really quiet. So there are very few transactions. No one knows how to price a company. If you ask, I have growth investors at the Series B and the Series C and the Series D, no one knows. And they, they look at the public markets and they want, they want a price at those levels. But the companies are sitting at such high post-money valuations on the last round that it's really difficult to approach a founder. And the reality is, so the public market's down 70% right now. I'd say the private market's probably down 50%. 40 to 50%. And if you're an employee looking at one of these companies, the question is, what is the value of my equity? And so what do I have to assume for the value of the company? What does that mean on a price per share basis? And how does this translate to me? I think that's probably the only way that you can look at an opportunity today and convince yourself that there's meaningful upside.

Unsaturated.

Great. Any follow-up on that, Jason? Okay, great.

Thank you.

Okay, perfect. Moving on to our next question from Matt Wong.

Hey, sorry. Hey, Tomas, um, thank you. Um, this is a real treat for me. I'm a big fan and follower of your, of your blogs and, and your work. Um, I had a really quick question. It's probably cut and dry. It's generic about new funds, but I thought maybe there's a gray area that depends on your personal experience. Like your own experience as a professional in your past work. And the question is, with a new fund, does that mean you're 100% conflict-free from any prior investments that you were involved in legally, morally, practically, you know, in all senses? Yeah, so I'm still on some of my Redpoint boards. So for my particular case, I could be conflicted out there. The way that it would, you know, it's sort of easy to resolve is just a conversation with Redpoint. But you should think at a high level, yes, you should think of me as being conflict-free.

Got it.

Thank you.

Great. Okay. Moving on to our next question from Milan, which actually I believe quotes some of your work. So another big fan.

Hey, Milan.

Hey, Tomas. Thank you. Yeah, you wrote in PLG and Profitability that PLG isn't necessarily the efficient go-to-market motion that it first appears to be. And as you know, Carrie Liu, who I just happened to see in the audience, which sidebar is pretty cool, described this as the biggest problem in tech right now, which is a quote I've been passing around a lot internally here at Genesis. So will theory have a preference for PLG PLS, sales-led, or will a sequence matter from founder-led and sales-led and then PLG?

Yeah, great question, Milan. So I wrote that post, and you know, the post that Milan is talking about looked at publicly traded companies, categorized them whether they're product-led or sales-led, and then looked at their sales efficiency through time. And the net result is there's no meaningful difference, or at least no statistically significant difference between product and sales. So the, you know, the way I think about it is a company has basically 4 different ways of generating leads. One is products. Somebody uses the product, then they convert to paid. One is marketing. They hop on a webinar and then they convert. Another is sales. So there's an outbound, somebody calling, or, and then the last is a referral or channel, right? Those are the 4 different ways. And historically, like maybe in the first 10 years of software, it was mostly sales-led. And then we went to marketing and then we got to product. And channel is now starting to come to be a little bit, but still quite small. I don't, I think for theory, there's no preference. The most important thing is you have to align the product with the go-to-market strategy. So if you're going after an enterprise buyer, it's tough to get there with PLG because an enterprise buyer is typically buying to minimize risk. As opposed— and the buyer and the user are separate. So the short answer is no. The thing that we're looking for is basically alignment between the product and the go-to-market boards.

Great, thank you.

Great, and thanks for quoting some of the articles. That's so cool to see on these and that you guys know each other. Love the community here. Um, our next question comes from Ajay. D.

Thanks, Thomas. Thank you for that excellent presentation. I love that 4-point framework that you, that you laid out. My question, I've been talking to a lot of deep tech people who've been considering starting to, you know, start something in the space. And one of the concerns that I keep hearing is, will the next drop of ChatGPT obsolete what I'm about to do? You know, and a case in point, like Harvey is a great case in point, right? A lot of companies wanted to come up with AI for legal and, you know, NDAs are public and you can, you can get a lot of help on NDAs through ChatGPT. Who's to say that in the next version MSAs won't also just show up over there and things like that? I would love to hear your point of view on that. And then since I have the microphone, I'll sneak in just one more, which is a lot of confusion on whether to take some of the open source models, you know, that OpenAI and ChatGPT have put out, or, you know, should you invest in something a little bit more capital intensive and create your own proprietary models and fork out of that open source to get some competitive advantage. Would love to hear your point of view on that.

Yeah, on the first one, I think I would just start building now because yes, I mean, look, all infrastructure improves with time. This infrastructure is changing very quickly, but the more experience you have with the infrastructure, I think the better of a product that you will build. And it's unlikely, what I haven't seen is sort of, breaking compatibility changes. In other words, like, if you built on 3.5 and then 4 and then 5, it's unlikely— in fact, the job probably actually gets easier than rather than harder. And so I don't, I don't see that as sort of a, a reason to pause. Um, what may happen— so we talked a little bit about prompt tuning. If you're building a machine learning system on top of ChatGPT-3 and then all of a sudden 4 comes out and 5 comes out, you may no longer need that prompt tuning system, or you may need to tweak it again. So you may have to retrain it. That, I think, is where the marginal expense happens, but it's not a breaking change. On the second question, I think a lot of companies will start with an off-the-shelf model, like a GPT or Ollama or BERT, whatever, and then use it in order to get distribution. And then at the point that they have distribution, they're able to raise enough money, $50,000, $75,000, $100,000, and then go and build a bespoke model as a way of creating a moat. I think it's tough to do it the other way around because it's hard to raise that amount of money, really.

Thanks.

Thanks, Aljay. We have another ChatGPT question in the chat. So if Catherine, you want to unmute yourself, feel free to.

Just a second. Can find my video. OK. Right here. Sure. So the question is in relation to the latest open letter to stop the ChatGPT to develop for about 6 months. And the question is, how safe do you think the use of ChatGPT in real operations for startups and other companies? And what are the potential outcomes of the regulatory reviews? If the regulators catch up to the technology development, what could that be?

What could be the impact on the companies?

Yeah, great question. So I disagree with the letter. I think you have to let technologies grow, and technologies can be used for good or bad, right? An airplane can be a commercial passenger aircraft. It could also be a fighter jet. You can use splitting atoms in order to make atomic bombs. You can also use it to produce power. So it's really a question of how we use the technology that I think is the most important. I also think there's an imperative, like a US national security imperative, to actually keep it going because falling behind, I think, has material implications for GDP, which is sort of why I started the presentation that way. On the regulatory side, I think in crypto it took regulators maybe 10 or 12 years to catch up to where the market was. I think it's very likely that regulation probably takes a similar amount of time. I think it's unlikely you see a total block of this just because the, the harms aren't yet well understood. So when we start to see, I mean, it's like when I was reading a history of aircraft and after the Second World War, there are all these new startups that were building aircraft and they designed their windows in different ways. Some of them had round windows and some of them had square windows and the ones with square windows, they would crash because square windows are, are, create weaknesses in the airframe. And then after that, you know, the FAA stepped in and said you can only make round windows. I, I think you're gonna kind of see a similar dynamic here where, where the technologists are just gonna keep pushing the technology and every, every, sometimes something will break and then the regulators can kind of step in immediately in order to shore that up. But the, I don't think you're going to see any kind of comprehensive generative AI legislation because even the people who are building this, like reading about it every day, we sort of, we're so early in our comprehension that to saddle it with a bunch of regulation out of the gate, I would just be a huge shame. Thank you. That sounds like we can continue leverage it safely. Yes, exactly.

And not only am I learning a bunch about generative AI, but I also just learned a really cool fact about planes. So thank you for that. Our next question comes from Armando.

Hey, Tom.

Good to see you.

Hey, Armando.

Good to see you. How are you, man? Like, Kry is always a pleasure. Curious, so you talked about check size for seed and Series A, but curious to know a little bit more about what gets you excited with companies that match the thesis. What gets you going? What are the things that you're looking for? At both those stages? Yeah, I think the, I mean, okay, so what are the things that really matter? I think team chemistry matters a ton, interpersonal chemistry. The second is whether the startup is solving a top 3 need for their customer. We're in an environment where the budget is a lot tighter than it used to be. And so your buyer has to be able to advocate to the CFO why they need to spend this money. That's really important. And then the last is a vision that's big enough to sustain a public company. 10 billion plus. Yeah, I mean, that's a billion in revenue, right? So today's multiples, but something like that. That makes sense. Thank you. Good to see you, Armando.

Thanks.

Always likewise.

Great. We have a lot of 'I like that' questions and a lot of thank yous in the chat saying— some people saying this is the greatest session we've ever had. So thank you so much. Do want to open it up for any last questions? Questions or conversation topics. I know a lot of conversations are also happening in the chat.

Tomas, could I steal one of the last questions?

Oh yes, be my pleasure.

What do you— I know you touched on it, but let's talk about all the legacy vendors, including ones that are newer, like we've invested in previously. How are they all going to be full of AI? Are there going to be multiple APIs? Like, like, will there be multiple just like we have 4 leading clouds, will there be 4 leading APIs and will they be integrated into every product? Like, what does it look like for the existing vendors?

Yeah, I think you're right. I think there'll be 2, 3, 4 vendors of these major universal models and everyone, everyone will start by— all the incumbents will start by just integrating those and seeing how far they can get. Can they increase the ACVs? Can they drive more share of wallet? And I think For those existing products, they'll be really successful in doing that because now that AI, I mean, if you're, you're not that deep into technology, I'm sure at this point you want some AI in your day. It's just so cool. It's, it's part of the new wave. And so now the, the thing I'll sort of caveat that is, is like there's all of machine learning and we talked about the 4 different categories and the most recent one is generative. I think that one of the byproducts of generative AI is that everyone will want AI, but it can be classification, it can be natural language processing, it can be prediction. It doesn't have to be generative. But there is, I mean, we're gonna see this marketing wave for the next 24 months of this is AI enabled, this is AI enabled.

Yeah. The one that I, that I think you touched on in the beginning, you and I have both invested, I think repeatedly in contact center broadly defined, right?

Yeah.

And that's an area where, you know, old AI was early, right? Everyone's tried to have AI and it's actually accel— in my portfolio, it's accelerating like massive change. Um, but I don't know that that favors a brand new startup necessarily, right? I don't know that it— I don't know that, that someone with a brand new help AI chatbot is going to benefit over a customer or Gorgeous or anybody else out there. Um, I'm just not sure. I'm trying to learn.

Yeah, I agree. Look, I, I think distribution matters more in this ecosystem than technology innovation because like I said, you can get to a 50 to 80% solution with a couple hours worth of work.

Yeah, that's the big question, right? That's what's fascinating is how far you can go with the API, right? It's epic.

Yeah, it's incredible. And so there might be a handful of customers that in the short term value, you know, 95% accuracy instead of 80% accuracy, but most customers won't want to switch because the cost of switching is not worth the benefit. It's not worth 10x the benefit of what you get from the generative.

Yeah.

Cool. Just curious your thoughts.

All right. I'll leave it open to anyone else for the last question.

Great. I think Carolu has like a conversation topic to bring up with the group. Hi, Tomas.

Hi, Carolu.

Great, great presentation. I was just talking about as companies scale, when you're talking about the depth levels from going from tuning the model to bespoke, it kind of reminds me of other infrastructure changes we've made over time where, you know, you have a startup that creates an infrastructure and then you have to go back and like re-architect for SOA back in that phase, or, you know, from on-premise to cloud. It just seems like there's going to be these, like, fast growth and then re-architecture problems as you readjust to a different, deeper model. And as a marketer, it's terrifying because in some companies I've worked for, it took like 5 years to get to cloud for real, or, you know, to re-architect the core for APIs or something, right? So I guess it's comment, not as much a question, but I think it's like a real growth challenge when so much is changing on the fly.

I think you're absolutely right. I mean, and the cloud analogy, the on-prem to cloud analogy is dead on. Most of the companies that I've seen make that transition, they think it's 12 to 18 months and ends up being 3 to 4 years in order to really get there because you need to hire a new team and re-architect the product and you may not get it right the first time. And it's very likely that this movement to bespoke models will be just the same. You need a new kind of engineer inside of the company. It's all new technologies. So that's brilliant insight, brilliant insight.

Great. We have one more question in the chat, and then I think we are going to try to wrap up. Do you think having not reached $1 million ARR is impossible to get funded? They do have an AI off— their context, their AI is off the charts. Love that promo for themselves and no one knows about us. I've been doing some scouting, but CEO is stuck on the $1 million.

Yeah, so I was saying that the market basically looks like it was in 2017 and $1 million in ARR, like 50 to 60 posts was the, was kind of the classic A. So there's some wisdom there. It's possible the generative AI fundraising market is partying like it's 2021. So that may sort of help you, but I think better to be prudent because this environment is just so uneven.

OK, I love that, partying like it's 2021. There is one more question in the chat I'm actually going to hit as well. So I lied, one last question. Are there any verticals that you think are ripe for disruption?

Yeah, this is what we spend our time thinking about. There's one vertical that I'm really excited about is marketing. In the last 10 years, there have only been, by my count, 4 publicly traded startups in marketing. And that's because of the dominance of Google and Facebook. And now that's changing because the third-party cookie is going away. It's being regulated out of business. And so I think for the first time in maybe 10 or 15 years, you're going to see a lot of opportunity in the marketing stack. So there's a regulatory discontinuity for that reason. And then there's also this generative stuff, right? Imagine if you had an ability to create creatives on the fly with like Midjourney. You could create 5 different creatives and then tie it to a machine learning system in order to do optimization of ad spend across different channels. That's suddenly possible. And so a new marketo. And so that's one space I'm keenly looking. So if you're building that space, please hit me up. I'm really, really interested there.

What an interesting take. Love that answer. Um, definitely not what I was expecting, but excited to see what comes next from there. Um, I think we're going to wrap now. Thank you so much, everyone, and especially thank you, Tomáš. We really appreciate you coming in, and thank you. I know it's been a really busy week with everything Theory Ventures, so thank you for taking the time to come to SaaStr this week.

Thank you so much.

Of course. Speaking of marketing, next week we are discussing 10 Things That Always Work in SaaS Marketing with our Saastr CEO and founder Jason M. Lemkin, who is on the call right now. All right, we will see you all next week. So many thanks and great topics and etc. in the chat. So thank you again and see you again Wednesday after