Thanks so much for joining me on the podcast here today. How are you?
Great, Jared. Thanks for having me.
Excited to have you. Let's dive right in. If you don't mind, give us a little bit about your background, and then Theory, for those that maybe are just meeting you for the first time.
Sure. My name is Tomasz. I've been a venture capitalist for 17 years, worked with eight unicorns, and started a firm called Theory. We raised about $700 million to invest in early-stage data and AI, and we have three businesses.
And for the companies that you invest in, like what is your — I know it can change, right? But what is your core criteria? What do you really look for? Specifically, let's just focus on the AI piece for this episode. But for those AI companies, what are you really looking for that stands out?
Yeah, we're typically looking for companies that have technical founders that can sell. That's really important. The AI wave is really about selling the change in the way that we work, which is really different than the last wave. It's almost category creation at every step. And so a technical founder who can sell is really important. And the other component that we're looking for are companies that exist at the intersection of three different circles: one is a lot of rote work, labor really unattractive work, a shortage in a labor market for whatever reason — there aren't enough graduates or people aren't attracted to the job — and then presence of many hiring managers who need to fill that seat now. And at the confluence of those three, you have a hiring manager who's willing to accept a solution that's probably 75% accurate, which is probably where the state of the art is in AI, and they're willing to accept the inaccuracies and work with vendors to improve that accuracy.
Do you think we're still like — obviously a big fan of how quickly the technology has progressed, but I see this a lot, and I'm sure maybe you do too — I won't put words in your mouth — but like, are we also overstating in some cases where the technology is at today, right? Because like I use it when you look at for video editing and for image generation, I think we kind of make it seem like it's seamless. It still messes up the images. So how do people looking at these technologies — like what's your advice for them when deciding is this technology actually ready to go, or maybe there's some technologies, right, that they're going to have to use and then it'll continue to get better, but they have to be an early adopter? So I guess what is your advice for those early adopters at companies exploring these enterprise solutions?
Yeah, I think you really need to be an early adopter because you need to figure out exactly where they fall down. They all fall down. I mean, putting together a PowerPoint presentation or image generation — you want an image with a particular word on it, good luck, right? Anyway, there are particular vendors who will do that. But or if you're trying to create an outline or a podcast, I'll use two different AI systems just to make sure it's comprehensive, some like Claude or Gemini as an example. So they're kind of Swiss cheese, right? Where sometimes they work and sometimes you'll put in a query and you'll get a big dud. But in order to sort of take advantage of this or understand where it's at, you need to use it because you need to understand exactly where those holes are in that Swiss cheese. So I think set your expectations today that the products will not be phenomenal, but set your expectations that in two or three years they will absolutely exceed our expectations. And so you can't — it's really hard to kind of step on an exponential curve halfway through and really know how to use the software. Ideally, if you can grow with it and have the patience, you'll be in a much better position to lead your whatever segment, part, totality of the business that you're responsible for.
I will say this too: for a lot of the companies that are still early but people should adopt, right, and start using it — they're not pricing like they are kind of discounting it right now for a lot of people as they're still making it better. So that's the good part, right? For these enterprises out there, in many cases they're not paying the price they'll pay when it is fully figured out, which is partly a benefit, right?
Totally. I mean, you hear Sam Altman say there's a $20 price point for ChatGPT, there's a $200 price point, there's a $2,000 price point. And I think across all of those different price points, you're seeing people using it, customers using it much more than you thought, and we're not making money, I think at least on two of those price points. So that just tells you, you're right, the market is underpricing it today to drive that software improvement.
Yeah, so like even more benefit to use it because you will more than likely get way more than what you're paying for. Yeah. From a usage standpoint, when you look at, you know, which models really stand out to you at the current moment? It seemed like Google was so quiet for so long and then had a bunch of announcements with Gemini. Like what really excites you about the models that exist today? And, you know, do you think we'll have a clear-cut winner? Like, and when might that happen?
Yeah, I think Google's transformation in the last 12 months in terms of ability to ship a product is — I don't know if I've seen, maybe Satya Nadella around Microsoft is on par, but it's a legendary accomplishment. The small models like the Gemma, the really small ones, or the Gemini Flash, I think they call it, and even the Deep Research on that — playing around with that one, that's pretty amazing. I told it the other day I was researching advertising, so I found a trade magazine. I said, summarize all the articles from the last 12 months about startups and give me a list of all the startups with a description and URL, and it did it. And so I think Google's pace of innovation here is pretty remarkable. Anthropic, my view, has carved out the API part of the business, at least a big chunk there. So if you're not the very, very largest software companies, but you're sort of mid-market up the mid-market, they're there. And then OpenAI, biggest brand out there, 300 million monthly users. So there's a lot of opportunity for innovation. Big push now is making the model smaller, and you see DeepSeek launched V3, trained it for six million bucks. I think we're going to continue to see those kinds of efficiencies because efficiency hasn't really been important — do bigger and bigger models. And now these small models with comparable performance, orders of magnitude better cost and pretty similar performance at least on some benchmarks. So I think you continue to see pretty significant deflation in that core part of the market. And then the big pushes, two areas: multimodal — I talk to a model and I wanted to produce an image or a video, it skips the whole text part. I think that'll be important because speech is the way that you and I, Jared, talk to each other, it's the way that you and I will talk to our computers. And then the other part about it is reasoning. So current models with the Transformer architecture are not great at planning because planning is a search problem. So we need a different kind of technology, reinforcement learning, to solve those search problems and create plans for long-term projects, the way an employee at a company would. And so I think we see a lot of big there.
What are your thoughts on the state of agents today too? Because like I think this kind of goes back to what you said earlier — many of these agents, it'll still be a work in progress depending on which ones you decide to work with, right? What are your thoughts on that? Is this an area that you're actively looking to invest in, or are you pausing on this space? Like, would love to hear.
Yeah, so agents is the GDP unlock that I think will be totally transformational for the US economy. I think we measure the impact in whole percent of GDP. And so just to kind of put that in perspective, the PC was a few basis points, 100x productivity. So I think that's where it's going. We talked, you know, if you have a model that's like 75% accurate and you ask it to do three things one right after the other, you know, with an output that's less than 50% accurate. So that's the problem that we're trying to solve. So design a slide, find, write the text, going to put something that's 50% accurate. So at the output of each step you need something that reduces that somehow. There are lots of different techniques. People are saying let's use chain of thought, maybe we can get the computer to produce a plan that's more accurate. Some people are running three models in parallel, just like Mechanical Turk, saying whatever the top two vote, let's take that. There's other architectures where you have one model that's proposing and you train another model that's a critic, just the way that a journalist would work with an editor. And we're testing all these different techniques to determine which is the right architecture. It'll probably be some huge complex system that depending on the kind of work that you're doing will have a different kind of architecture. And I suspect in 2025 we see some pretty significant innovations there.
