Navigating the AI Investment Landscape

Summary

A 32-minute conversation with Lauren Hawker Zafer on Redefining AI, the Squirro Academy’s podcast, recorded in the summer of 2023 and released September 5, 2023 — two months after the Contrary Research report The Openness of AI (mirrored as the essay The Openness of AI), which Kyle co-wrote with the founders of Nomic. The episode is essentially that report’s argument, delivered conversationally, plus a compact version of Kyle’s framework for evaluating AI investments.

Where the money went in 2023. The most fervor and capital went to foundation-model companies — OpenAI, Anthropic, Cohere. The second, quieter area is MLOps and orchestration: excitement about what the technology can do has raced far ahead of the ability to deploy it, and startups are filling that gap. There is not enough talent to go around (a Netflix ML product-manager posting at $900K is the example), which is itself an opportunity: a product that lets one person do what would otherwise take six.

Progress vs. distribution. Borrowing from a conversation with Aidan Gomez of Cohere — one of the authors of the transformer paper — Kyle argues there has been no sudden step-function in the technology. Researchers have been standing in the same spot and the world caught up. The current phase of the cycle is therefore less about innovation and more about distribution: how incumbents like Microsoft or Canva push the capability into products, and how startups get it into the hands of the people who can extract the most value.

Value creation and value capture. The framework that runs through the whole episode. Innovation creates value (Stable Diffusion created a lot of it); the question is whether there is a business model to capture it. Nomic is the worked example: GPT4All, a pared-down model that runs locally, is open source and free — pure value creation — while Atlas, the data-visualization and fine-tuning tool used to build it, is a traditional software product people pay for. Replit is the second: Ghostwriter, its AI coding assistant, is valuable but replaceable (GitHub Copilot competes); what people pay for is having every developer tool in one place.

The Openness of AI report. The report unpacks how the largest players — Microsoft, Google, OpenAI, Anthropic — are increasingly focused on centralizing power, echoing Clem Delangue’s line that concentration of power is the greatest danger in AI. Kyle thinks centralization is avoidable for two reasons, one market-driven and one philosophical, and the dollar is more powerful than the dogma. On the market side: nobody wants Platform Risk. The same instinct that made companies multi-cloud will make them refuse to route everything through one model provider’s servers. Philosophically: a model that formulates answers can shape how people think even more powerfully than Facebook’s feed shaped what they read, and people will not want one monolithic brain deciding how everyone gets information.

Show me the incentive. Applying Charlie Munger’s rule, Kyle splits the industry by whether a company owns a cloud business. Microsoft put $10B+ into OpenAI so the compute runs on Azure; Google backed Anthropic so it runs on GCP. Meta has no cloud business, so it open-sourced Llama — it just wants the best possible product, and an open model that the world improves makes its product better. Since the vast majority of companies on earth do not own a cloud, most of the world’s incentives point toward openness, and “monopolies never lead to the best products.”

AI safety, reframed. Many of the discussed risks are dramatically overblown — an LLM is autocomplete on steroids, and extinction-level talk is a far cry from what is happening. Citing the Cerebral Valley conversation between Clem Delangue and Amjad Masad, Kyle argues the extinction narrative is partly marketing that draws attention away from the real problems (disinformation, content moderation), and that OpenAI releasing less context around GPT-4 makes it easier to hide inherent biases. Openness makes safety better, not worse. On the bad-actor objection: every technology eventually becomes widely distributed, and the counterbalance is how well the long tail can defend itself. A world where only a few big companies hold the capability leaves every smaller company and nation more exposed, not less.

Who has to get smarter. Asked how non-technologists should educate themselves against disinformation, Kyle declines the premise: a course everyone on earth must take is not realistic across so diverse a population. The responsibility sits with platforms and governments. Laws that once shielded platforms like Facebook from liability for content (the Section 230 style of provision) are changing, and platforms need to be capable of understanding and governing what happens on them.

How to evaluate an AI investment. Two buckets. First, value creation unpacked into three things: a strong community of smart people congregating around the thing; meaningful top-of-funnel traction that translates into activity (Taylor Swift’s Instagram following is a community, but the ticket sales are the traction — likewise, GitHub stars need to convert into use); and a genuine product that can be monetized like traditional software. Heavy open-source traction with no clear business model is the pattern that makes him hesitant. Second, price is relative to outcome: $1M for a possible 100x is cheap, $1M for a 0.8x is expensive. People get distracted by OpenAI’s valuation, but OpenAI has massive corporate ties to Microsoft and is building a foundation model for every purpose — a ChatGPT wrapper or an orchestration tool does not have the same potential, and pricing it as if it did is actively damaging.

Transcript

Lauren (cold open): The number of people who work in a venture can vary wildly depending on the size, type, and stage of the venture. A venture, we know, typically refers to a startup company or a new business initiative that is seeking to grow and achieve significant returns. The number of employees in a venture can range from just a few founders or early team members to hundreds or even thousands of employees as the venture scales. Kyle, what’s it like working in venture? What excites you about the field?

Kyle: The thing I love about venture capital is the opportunity to be a resource for people that are building at the cutting edge. That’s my number one goal — to not present myself as the person in the seat driving the ship, trying to build the plane as we fly it, but the person who can bring the resources to bear to help those people accomplish what they’re trying to achieve. That’s my favorite part about venture capital.

Lauren: And when you say be a resource, what can we envision by that? What sort of resource are you? Are you providing emotional support? Financial support, obviously. What are your resources?

Kyle: I’d love to say all of the above. Certainly every VC thinks of themselves as spiking in very different areas. At Contrary, our focus is on being a resource for people to be able to access talent. A lot of what we do is, as companies are building their early team, they want to hire really exceptional engineers, product designers, product managers, things like that. Contrary has a network of 500-plus folks that we stay very close to, and we help companies hire from that talent network. So in addition to trying to be an emotional support and all these other things, and providing capital, we also want to let people access our talent network to build out their teams.

Lauren: It’s such an interesting word though, isn’t it — talent. What is talent? How are you defining talent? How are you measuring it?

Kyle: We think about the folks that we have in our talent community as specifically people who have demonstrated slope in their career. They’re trying to accomplish a lot of different things and they’ve been successful in a lot of different areas. But beyond that, it’s really trying to match the best people with the best companies. So it can mean a lot of very different things for different companies, depending on what a company is looking for. For example, as we invest in AI and ML companies, they’re often looking for people with both academic rigor and exposure to the latest thinking and frameworks and architecture, but also people with practical capabilities in building models and fine-tuning these systems. So it’s finding the people that have the skills that match the company. Our job is just to find as many exceptional people as we can and then match those to companies that are a good fit for their background.

Lauren: Hi everyone, it’s Lauren Hawker Zafer. Welcome back to Redefining AI, the tech podcast. Today I’ve been joined by Kyle Harrison. He’s a general partner at Contrary, where he leads Series A and growth-stage investing. Welcome, Kyle.

Kyle: Thanks for having me.

Lauren: So we’ve discussed and established at the start that you’re an experienced investor and that you’re a GP at Contrary Capital. Kyle, if we look at maybe the evolution, or what you have witnessed in the landscape of AI over the last year from an investment perspective — what types of investments are being made, and where are they being made, and, in your opinion, why?

Kyle: Definitely the most fervor in the investment landscape has gone to companies building foundation models. That’s very much the OpenAI, Anthropic, Cohere companies like that. That’s probably sucked up the most capital and most excitement and most attention. But another area that has been pretty active, and I think is a pretty important part of the ecosystem, is around things like MLOps and orchestration — different things to allow people to put this technology into production. That is definitely a big area of investment and excitement, and I think a big part of that is driven by the fact that you have all this excitement about what the technology is capable of, but that has raced far ahead of the capabilities to deploy that technology. And so what’s filling the gap are companies coming along and saying, hey, we can help you bridge that gap. You may not have the expertise internally to be able to do these things. We can help you put these models into production.

Lauren: And do you think that there’s enough talent to be able to feed those requirements — in the sense of, there’s enough talent for companies to be able to build these foundational models, to develop MLOps cycles, to twin together all the orchestration capabilities to ensure that it can be offered?

Kyle: Candidly, as of today, the answer is no. There’s not enough talent to be able to effectively leverage these systems. I was just reading an article today about how Netflix has a job posting out for, I think it was a product manager for some ML products, and the salary was $900,000. Netflix, which is not, maybe, a traditional AI company per se — but they have a use case, and there’s a bunch of other companies that are very similar, paying top dollar to try and find these folks, because the reality is there’s just a massive talent shortage. But I think that creates an opportunity for startups who are trying to enable companies to leverage these systems. They can say, hey, you’re just not going to be able to compete with the Googles and the Microsofts and the OpenAIs for talent, but what you might need six people to do, with our product you can do with one. That is a big opportunity for companies.

Lauren: Is that what you’re encouraging people to do at the moment — in product development and the implementation of products that can glue that gap together?

Kyle: I’d say most of the companies we’ve invested in are early enough in their life that they’re pretty scrappy, and so just naturally they’re already going to do that. Most startups are not throwing people at problems. They’re trying to find the most efficient way to do things. So if they can find an orchestration tool that helps them put a model into production, that’s great, and they will do that. I think it’s the larger companies that are having a harder time adjusting to this stuff, that would probably love to throw people at the problem and can’t because there’s just a shortage of talent. And so they have to start going and looking for solutions. But again, when I look at investing in a company that’s building in this space, the thing that I’m focused on is understanding: do they have the technical capabilities that they need internally, or if not, do they have the systems in place that can allow them to take advantage of the latest technology?

Lauren: If you look at the technical capabilities, do you now think that the evolution of where AI stands — especially with the explosion, or the mainstream attraction, around generative AI — is now in a space where the technical capabilities are rocketing the innovation or the advancement, and where startups are as well?

Kyle: What’s interesting is I did a podcast a few weeks ago with Aidan Gomez, who’s the CEO of Cohere, a company that competes with OpenAI and Anthropic. He was one of the original authors of the Google paper that introduced the transformer architecture, so a pretty important piece of this boom that we’ve experienced over the last few years. And he made a really interesting comment where he said, you know, most of us feel like we’ve kind of been standing in this same spot and the world suddenly caught up. And so I don’t even know that I look at it as there has been this rapid rocket ship of progress per se. There’s certainly been a ton of progress, but it hasn’t been this in-a-moment, everything-technologically-leapt-forward. It was that we have made these incremental improvements to the ways that we’re doing things, but suddenly it came to everyone’s attention. So I think where we’re at in the cycle right now actually has a little bit less to do with innovation per se. Certainly there’s still progress being made and people are still making breakthroughs. But it’s less about, oh, there’s going to be this step-function change that occurs. We kind of already have that. Now the challenge is: how do you effectively distribute that, whether it’s into incumbent platforms like Microsoft, or even things like Canva or whatever, that are leveraging this technology internally? How do those incumbents distribute it? Or, as startups, how do you take advantage of what’s out there and then get it into the hands of people who could get the most value out of it?

Lauren: How important is innovation, though, when you’re looking at it from an investment perspective? Because obviously if we’re talking about the cycle, and it’s been a slow development in maybe incremental stages, there is also that part of the cycle — if you look from another angle — of the mimicry that goes on. You see a lot of startups that are possibly copying each other to a certain extent. From an investment perspective, how do you recognize the potential that a startup might have, outwith the three things that you may have listed at the start?

Kyle: The equation that most people use, or the framework that people use to think about where good investments exist, is basically a function of value creation and value capture. Value creation is a lot of where the innovation happens, where you’ve created something that can create value for a lot of different people. And there are a lot of models that have been built — Stable Diffusion being one of them — that a ton of folks have gotten a lot of value out of. Then the question is, is there a mechanism to capture that value? Which is effectively just your business model. Do you have a good business model to capture value? There are some of these products that have not demonstrated a really exceptional business model, and so it’s not as easy for them to capture the value that they’re creating. So what I look for is not necessarily just “has value been created,” because there’s a lot of innovation that creates value, but also, is there a business model that can capture that value? You’re trying to find a combination of those things, and that typically leads to much better investments.

Lauren: And what would be an example of “is there a business model that can capture this value”? Can you give us an example?

Kyle: One of the companies that we’ve invested in — and I’m sure we can talk a little bit more about it, I co-wrote this deep dive on the openness of AI with them — is a company called Nomic. The reason we made the investment is because I think it demonstrates this very well. Value creation happens because what Nomic has done is created a tool called Atlas, which is basically a data visualization tool to understand and fine-tune large language models. You can visualize the entire dataset that a model is trained on. And by doing that they were able to create things like GPT4All, which is a pared-down language model that is as performant as GPT-3.5 but at a fraction of the compute resources, so you can run it on a machine locally. They were able to do that with their tool. So they created a lot of value — hey, we have this model, it’s much easier to use, people in healthcare and financial services and government feel more comfortable using it because they can use it locally, it doesn’t rely on somebody else’s server. All of that is great. But Nomic’s not charging for GPT4All. It’s open source. So that’s an example of value creation. And then the question is, how do you capture that value? People look at that and they say, wow, Nomic was able to create GPT4All using Atlas, this data visualization tool that they charge for, that is a traditional software product. And people say, man, I’d love to be able to do that with other models, or to be able to build my own models that I can use locally. And so they want to go and pay for Atlas as a product. That’s a business model that captures value, because it attracts that top of funnel — attracts people to a software product that they can then charge for and monetize.

Lauren: Do you give advice on that as well? If a venture was struggling to identify a business model that could capture value that they’ve created, do you provide services that would support the development of a coherent and hopefully profitable one?

Kyle: As VCs, I think we typically think of ourselves less as the consultant or service provider, pointing people in a particular direction. What we’re trying to do is identify those things that already exist. Is that mechanism already there? If it is, or if there is the opportunity for that, then we want to invest. So I think it’s more a function of, I look to identify that. But I see that in a lot of the companies that we invest in, where the value creation is often being good at something else. It’s often not necessarily an AI capability exclusively. Another one of our investments is a company called Replit, which is an IDE — a development, a coding environment. What they have done really well is they take all the things you could possibly need as a developer and put it in one environment to make it very easy for you to build software. One of the things that they have is Ghostwriter, which is their code assistant AI. They’re able to leverage AI to create value for people, and that’s awesome. But what people are paying for is not this AI tool, because they can get that tool from a lot of different places, not the least of which is GitHub’s Copilot, which is a competitor. They create value by putting everything into one place, and that allows them to charge for that, because people say, man, I’d love to have everything in one place rather than having to go pay for 20-plus different tools. So a lot of times what we’re trying to do as investors is identify where that value-creation-and-value-capture relationship already exists, or very soon could exist. Hopefully we get there maybe a little early, and then we’re able to invest and help them create that.

Lauren: It definitely seems that we’re on a trajectory of this hyper-personalization, this accessibility to hyper-personalization — of not going outwith, but staying within some sort of containerized opportunity that’s supported by AI. You mentioned before that you were part of the author team who wrote the report The Openness of AI. Can you provide us with a brief overview of the key findings and insights from that report, and how it sheds light in particular on the accessibility and democratization of AI technology?

Kyle: We produced a report as part of Contrary Research, which is the research arm of our firm. One of the things Contrary Research does a pretty good job of is creating this engine to take expertise and engage with the expertise of other people. Huge credit to Brandon and Andriy, who are at Nomic, because basically what we did was work with them to understand their perspective on the space and then capture that into something that’s consumable for everyone to read. It was a lot of fun to dig in with them on what they understood of the space. In the report, what we focused on was unpacking a lot of the biggest players in AI right now. So you have, whether it’s Microsoft or Google on the large end, OpenAI, Anthropic, what have you — and specifically how increasingly these folks are focused on becoming quite powerful. They might not say it that way, but there’s a real emphasis on trying to centralize as much power and influence over this space as possible. Clem Delangue, who’s the CEO of Hugging Face, has said that the greatest danger in AI right now is this concentration of power. So in the report we really focused on what are the implications of that centralization, and how there can be a better way — rather than having one winner-take-all environment where it’s a Microsoft–OpenAI universe and you’re dependent on however they want to do things, there’s really this hope that there are lots of ways to push the space forward and make progress, and lots of different products that can be specialized in different areas, and how that actually leads to better results as opposed to being dependent on one monopoly player.

Lauren: Do you think it’s possible, though, to avoid the centralization?

Kyle: I do think it’s possible. I think there are two buckets of that equation. One is a market driver and one is a philosophical driver. From a market perspective — which I think is honestly much more powerful; the dollar is much more powerful than the dogma — the driver will be that people don’t like to have platform risk. They don’t like to be completely dependent on one provider. Which is why, even in cloud computing, which is a big component of this whole equation, people don’t like to be single-cloud. They want to be multi-cloud, because they want to have some sort of control over what they’re dependent on and when and why, and have some failsafe. So just from a market perspective, people are not going to want to have one system. In addition to that, they’re also not going to want to be forced to send all of their data to OpenAI’s server, for example. People want that semblance of control, and they’re willing to pay for that. So a lot of companies will emerge and say, hey, we can help you, use us, but we’ll play nicely with others. I think the market will force some of that behavior.

And then philosophically, people are also concerned, especially with something like a large language model, because it has pretty deep ramifications. Once you start relying pretty heavily on a model, say ChatGPT, that can actually have significant influence — even more powerfully than, for example, Facebook. Facebook had a really significant influence on how people thought about the world around them and how they consumed news, and that was just algorithmically deciding what they read — showing them what they read. ChatGPT can actually formulate the way that people think, because of the answers that they’re getting and how dependent they are on it. So philosophically, people don’t want to have one monolithic brain deciding how everybody gets information.

Lauren: Very insightful input, and I don’t disagree — I certainly align in agreement with one particular component of what you said. I think it is interesting that you’ve displayed two different sides of the equation: the market perspective and the philosophical perspective. One question would be, which is more conscious as well? If you’ve got the market perspective, is that leading people’s decisions more than the philosophical perspective? And what plays into that? Is it one industry that’s contributing to the market perspective, or the philosophical perspective? And what you’ve interestingly brought into the discussion is the whole desire to go multi-cloud. If you look at geographical components that align with that, is everyone at the same stage of that journey, or are they misaligned in terms of market perspective, philosophical perspective, and also their journey in their own cloud — which, as you mentioned, is a huge component of this discussion, the intersection of AI and the cloud? That was quite a lot of questions.

Kyle: Big questions. Well, I think the way that I unpack it — it’s certainly multifaceted. But I think about a quote from Charlie Munger where he says, effectively, show me the incentive and I’ll show you the outcome. Help me understand who’s incentivized toward what, and that will dictate what they ultimately end up doing. And so I certainly don’t think it’s uniform — there’s not one industry that drives all the changing forces. Part of that is because AI, while we talk about it like it is an industry, is as ubiquitous as SaaS, software as a service. It’s a mechanism that is going to impact everything. It’s the same way people used to say, hey, I invest in internet companies. And it’s like, what does that mean? Every company is on the internet. Every company is leveraging the internet somehow. So it doesn’t really mean anything anymore. Internet, SaaS, AI — these things become very ubiquitous. So one of the things that will make everyone a participant in this discussion is the ubiquity of AI. Everybody is impacted by it. That’s thing number one.

Thing number two — when you start talking about financial incentives, I think you have two dueling incentives. On the one hand, you have the companies who are very highly incentivized to make sure that as much of the compute in AI, and in AI use cases, runs through their servers as possible. Which is why Microsoft invested ten-plus billion dollars in OpenAI — because they want as much of this to happen on Azure as possible. It’s the same thing with Google investing in Anthropic; they want as much of this to happen on GCP as possible. So there’s this one side of the universe that’s very incentivized to cobble together as much compute into their business as possible, because that’s more revenue for them. And then on the other side, you have folks who are much more focused on making sure that they have the most performant products possible. I would put Meta, or Facebook, in that bucket. You see that demonstrated in the way that Meta released Llama, which is their foundation model, which is open source. They just recently released Llama 2, which is available for commercial use. The reason that those companies exist on different ends of the spectrum is because Meta does not have a cloud computing business. So they’re not incentivized to make sure as much of this is in one place as possible. They just want their products to be as good as possible. If they release a performant foundation model to the world and people build on top of that and improve it and create use cases that are more performant, that Meta can bring into their own product, their product gets better. And that’s better for everyone. It’s better for them, it’s better for their users, it’s better across the board.

And if you think about it, the vast majority of businesses on earth do not have a cloud computing business. So the vast majority of companies just want their products to be as good as they possibly can be, and monopolies never lead to the best products. So the vast majority of companies, of countries, of industries, are going to be pushing for what allows me to have the best product. Being dependent on Microsoft, where they can do whatever they want because they’re the only game in town, does not lead to me having better products.

Lauren: You mentioned the open source component of AI, and there is that debate over open-source AI and AI safety. It’s one of the central themes that also appears in your report. Can you elaborate a little bit more on the different perspectives within the industry — we’ve touched upon it there — but also on how much they impact the development and deployment of AI technologies?

Kyle: I’m always careful, because I don’t want to downplay that there are certainly risks of the technology. But I think that many of the risks — and we touched on this in the report — many of the risks around AI that people discuss are dramatically overblown. People kind of joke that a large language model at its core is effectively autocomplete on steroids. It’s a massive thing that’s guessing what the most logical next word is. So to say that this is going to take over the world or destroy humanity is a pretty far cry from what’s actually going on. But a lot of people like to spend time thinking about that. There’s a really good conversation that Clem at Hugging Face and Amjad Masad, who’s the CEO of Replit, had at a conference called Cerebral Valley, where they unpacked all of this. They talked about how Microsoft is really pumping up the marketing around these risks because it just draws more attention to the conversation, when in reality a lot of the most poignant AI issues that are facing the world right now have very little to do with extinction-level events and much more to do with things like disinformation and content moderation. And those issues get swept under the rug when you focus on some bigger risk. Effectively, when companies like OpenAI decide to release dramatically less context around models like GPT-4, that actually makes it easier to hide things like inherent biases that exist in a model, because you can’t understand what the training set is, or what those weights are, or anything like that. So those actual issues get exacerbated while people focus on these really big problems.

So in my mind, the thing that I focus on and pay quite a bit of attention to in thinking about AI safety — I actually think it gets better when you are more open, when more people are trying to understand the biases and potential risks. One of the counterarguments that people make is, well, what about bad actors? If we make these models open, bad actors can get access to those, and different countries can use them to influence elections and do that even more effectively. I think that’s fair to say, but I think that’s also somewhat inevitable. For the most part, every piece of technology has eventually become widely distributed. And the thing that acts as a counterbalance to that distributed technology, for good or for evil, is how well prepared the long tail is to defend themselves. The more distributed that technology is, you basically get into almost a Cold War standoff. If everybody’s capable, then it’s less of a risk. If a centralized group of folks are protected — these big companies — and a bad actor gets access to it, every other smaller company or smaller nation is more at risk, because they don’t have the technology to defend themselves. Because the big company over here is saying, hey, we want to make sure that those folks are having to pay us to protect themselves. And that actually makes the world less safe.

Lauren: Because it goes back to what you’re talking about — the emphasis on the importance of making AI accessible to everyone, regardless of factors like cost, geography, or political affiliations. If we look as well at the two of the most poignant risks that you mentioned, that are a lot of the time swept under the rug — the disinformation, the necessity for stricter content moderation — who is it that needs to educate themselves better, and how can people educate themselves better? If you’re not a technologist, if you’re not immersed in the scene and you don’t understand the consequences of the inherent biases that are produced by the models, how can you educate yourself? What would you advise listeners to do?

Kyle: I think the unfortunate reality is that technology is getting better and better, and so I think it is more so a function of platforms’ and governments’ responsibility to be better prepared to manage their own platforms. It’s really tough to say we should have a course that everyone on earth has to take, and by taking that course they will recognize things that have been created by AI, or disinformation, or whatever. It’s really tough, because there’s such a massive, diverse population of people with different backgrounds and experiences and levels of education, and it’s really difficult to say, if you just do this, you will be more prepared. I think that increasingly — and this is one of the things that, for a long time, at least in the US, and I think this is true in other countries as well that have similar laws — platforms like Facebook were protected by certain provisions where they weren’t able to be held liable for the content on their platform. Those laws are increasingly changing, because they’re basically saying, listen, Facebook has a responsibility to understand and control and moderate the content on their platform. So Facebook, rather than being able to cede responsibility and say, hey, people are doing stuff and we’re just trying to make sure it falls within these parameters and then everything else is fair game — now they are more responsible for the claims that are made on that platform. And if that means that they have to pare back what people can and can’t do, those levels of controls need to be in place for those platforms to more effectively govern themselves. I’m more focused on those things needing to happen — more platforms need to be capable of understanding what’s going on and managing it more effectively — as opposed to saying, oh, each person is responsible for educating themselves, because it’s just really difficult to do with how effective technology is.

Lauren: So, to circle back and close our conversation to investment — and investment from the perspective of an individual listener, because obviously people are curious. You’ve mentioned that technology is offering a lot of opportunity, it’s changing, there’s a lot of players, there’s a lot of interest in startups. If someone was curious about investing in that market, and also a potential investment that would provide a return in the future, what would be your advice from that angle?

Kyle: I’d say there are two buckets of thinking that I revisit most often. One is a bit of an unpacked version of the framework we talked about earlier, where there’s this value creation and value capture. The way that I think about that specifically in this market, in AI, is I look for three key things. The first one is a really strong community element, because I think communities are a very powerful indicator of successful technology. You’re basically looking for where smart people are congregating and contributing and paying attention. Where is that community of thinkers? The second is trying to understand meaningful top-of-funnel traction. It’s not always open source, but often it is — where you’re looking for things like, something as simple as GitHub stars, where there’s a lot of people paying attention to this specific repo or contributing to it. But what you’re trying to see is not just, hey, have you been able to attract attention? Because when you think about it, what is a community? Taylor Swift’s Instagram following — that’s kind of a community. OK, but is there actual traction to that? What are they doing? Well, that translates into ticket sales for Taylor Swift concerts. It’s the same thing with open source. You’re not just looking for where people are congregating, but whether the congregating translates into specific activities. Those are two elements — if I were to unpack value creation, if you’re creating value, people will gather around you and often want to be involved. That’s value creation unpacked. Then the third piece, value capture, is really looking for a genuine product — something that can be monetized and sold like traditional software. Where I get really hesitant is when I look at something that has a lot of open-source traction but no clear business model. There’s no clear way to capture value. So that three-pronged bucket is one thing I think about a lot when I’m evaluating what are interesting investments.

The second thing I would pay very close attention to is that any investment — how good an investment is — is relative to the potential outcome. If I say, hey, you can invest in this thing and it might increase by 100x, and you have to pay a million dollars for this investment, a million dollars could be cheap for 100x. But if you say it’s a million-dollar investment and your return is 0.8x — you’re actually going to lose a little money — that’s a terrible investment; a million dollars is too expensive. It’s all dependent on the size of the outcome. A lot of people get very distracted because they look at things like OpenAI and they say, oh, OpenAI can be a 30-plus-billion-dollar company, so an AI company can be a $30 billion company. In reality, OpenAI is playing a fundamentally different game. That makes it effectively an irrelevant comparison, because number one, they have massive corporate ties to Microsoft, and number two, they are trying to build a foundation model that everyone will use for every purpose. Not every company is like that. If you’re looking at a wrapper around ChatGPT that does a very specific thing, or if you’re looking at an orchestration tool that can do something, it doesn’t mean that it can have the same potential as OpenAI. It’s actually damaging to think, oh, this could be a $30 billion company. So with every investment, you really have to consider what could the eventual outcome of this company be, and what does that mean for my investment?

Lauren: Interesting, fascinating insights, Kyle, and I really enjoyed the conversation today. So I’d like to thank you hugely for contributing to what we’re exploring and discovering today on Redefining AI.

Kyle: Thanks for having me. I had a lot of fun.

Lauren: Thank you. I’d like to thank everyone else that’s listening today. And if you’d like to find out more about machine learning, search, and insight engines, then go to the Squirro Academy at learn.squirro.com. Thank you.

Connections

• The report this episode is built on: The Openness of AI (Contrary Research, July 2023), co-written with the Nomic founders Brandon Duderstadt and Andriy Mulyar and mirrored as the essay The Openness of AI. The centralization-of-power framing, the Clem Delangue “concentration of power” line, and the incentive split between cloud owners and everyone else all come from it.
• Companion appearance: The Centralization of Power in AI (MLOps Community Podcast, October 2023) — the same report argued at length a month later, with Aidan Gomez, Clem Delangue and Amjad Masad cited again. The conversation with Aidan Gomez that Kyle references here (“standing in the same spot and the world caught up”) predates both.
• Frameworks that recur across Kyle’s writing:
  • Value Creation vs. value capture — the spine of the episode; see also The Value Cycle and Markets, Markets, and Markets.
  • Incentives — the Charlie Munger “show me the incentive” rule, applied here to who owns a cloud business; the same quote anchors Eat What You Kill and Building an Actual Unicorn.
  • Platform Risk — the multi-cloud analogy for why no one will route everything through a single model provider.
  • Price relative to outcome — the “$1M for 100x is cheap, $1M for 0.8x is expensive” test, and why OpenAI is an irrelevant comparable for a ChatGPT wrapper.
  • Talent — Contrary’s talent network as the resource a VC brings, and “slope” as the selection criterion.
• Concepts: Open Source, AI Safety, Artificial Intelligence, Open-Source Knowledge, Llama, Section 230.
• People: Lauren Hawker Zafer (host), Aidan Gomez, Clem Delangue, Amjad Masad, Charlie Munger, Brandon Duderstadt, Andriy Mulyar.
• Companies: Contrary, Nomic, GPT4All, Replit, Cohere, OpenAI, Anthropic, Microsoft, Google, Meta, Facebook, Hugging Face, Stable Diffusion, GitHub, Netflix, Canva.