A Conversation with Lex Sokolin: Signals to Watch in an Emerging Machine Economy

As AI systems move from experimentation into daily use, investors and researchers face a practical question: what will real adoption look like?

Lex Sokolin is the Managing Partner and co-Founder of Generative Ventures, a venture capital fund investing in the machine economy powered by fintech, accelerated by AI, and settled on web3.

Lex recently joined FCAT’s Head of Research, Megan Kelley, to discuss the signals he watches for and explain why the capacity for sustained abstract thinking continues to be a differentiator for humans as this economy evolves.

1. What signals are you looking for to show that the machine economy is happening?

   Lex Sokolin: There have been two groups of mid to large scale companies that built technology and agentic protocols for others to integrate. That assumes, first, AI agents that do things—and second, agents that do things and then get paid in a novel way. We need to see commercial efforts around those themes. And that has to show up in human behavior: people need to interact with systems differently. Will people actually use native integrations with a chatbot as a purchasing flow? If they do, that has a whole set of implications. If I’m shopping using an LLM, that means the LLM has my wallet and is operating like my bank. That’s a winner take all situation on the payment side.

2. Does that same dynamic apply on the capital markets side?

   Lex Sokolin: I have quite a bit of skepticism there. Distribution is very small: most of the current approaches are pulling users from crypto wallets. Crypto wallets are smaller than crypto exchanges, and crypto exchanges are smaller than banks and brokerages. So you start with a very small market of people with access to a limited selection of DeFi products. We’re watching transaction volume and user counts around these products. That data is visible—and the numbers aren’t there yet. On the hardware side, we’re asking similar questions. How many devices are actually attached? And are they doing anything meaningful with one another, or just idling? It feels quite early, but it has started

3. You’ve talked about “novelty search” as part of how you think and work. How do you see research evolving as machines take on more of that capability?

   Lex Sokolin: “Novelty search” comes from machine learning. I picked it up from the book Why Greatness Cannot Be Planned, which contrasts novelty exploration with linear optimization. The observation is that no company gets from zero to scale by following a clean, linear five year growth plan. Large machine learning systems behave the same way – the fastest way to optimize towards some task was the novelty path. Machines are just as good, if not better, than humans at novelty search. I don’t think humans have an advantage there. I also don’t think we’re uniquely good at recombination anymore. People used to say machines couldn’t be emotional or creative. That’s clearly no longer true. Image, video, and music generation make that obvious. Innovation is recombination and experimentation. Machines can do that.

4. So where do humans still have an advantage?

   Lex Sokolin: What humans still do better is maintain a consistent mental model over time. We build frameworks informed by lived experience. Machines don’t really do that yet. I’ve built my own mental model of how markets and technology evolve. When I encounter information that doesn’t fit what I already believe, that’s where I focus. Either I was wrong, or something changed, and I need to update my model. That’s why I write publicly—to update my thinking in real time. Today’s AI systems don’t really do that. They don’t remember what they believed before. That may change, but it’s a real limitation right now. So research, to me, is still about identifying what is genuinely different in the world. And that’s harder than it sounds.

5. As more work happens around spatial computing and world models, could machines end up with something closer to lived experience?

   Lex Sokolin: Imagine it takes a $500 billion data center to index all human knowledge. Now imagine that fits on a phone. The entire internet is indexed and updated continuously. But instead of tracking raw facts, you track abstractions—what changed and why. We don’t say “commodities went up.” We say “commodities went up because of these underlying forces.” Those are abstractions. Inference models already operate at far more layers of abstraction than humans. What’s going to be interesting is how those abstractions evolve over time.