Our Approach
Two things AI labs are getting wrong
- The real world is multiplayer, adversarial, and always changing. No amount of training on problems that sit still (code, chat, math) will create superintelligent agents for the real world.
- Monetizing intelligence by selling an API or SaaS product is inefficient and leads to commoditization.
Protogon Research is a VC-backed AGI lab that solves both these issues by taking a radically simple approach: training and monetizing AI agents directly in financial markets through proprietary trading.
Markets are (1) the most multiplayer, adversarial, and fast-changing environment on Earth, with verifiable rewards at scale, and (2) the most direct way to turn AI intelligence into cash without having to worry about sales, marketing, customers, or product.
Problems that sit still
A benchmark does not change when your model gets better at it. The test questions come from the same distribution as the training data, and there is no opponent adapting to your model.
Real-world problems change when you solve them. When a security model blocks an exploit, attackers write a new one. When a content filter blocks a prompt, users rephrase it. A model trained on last year’s data is solving last year’s version of the problem.
Selling intelligence by the token
Selling tokens leads to commoditization. Most AI labs invest enormous capital to build similar models, test them on the same benchmarks, and sell them through the same interface. Whatever one lab releases, the others copy, so the labs compete on price. The price of a token is falling toward the cost of the electricity used to generate it, and the labs keep only a small share of the value their models create.
As a result, most AI labs are burning money and can only fund their research by raising ever larger VC rounds. Their AI development does not pay for itself.
Why markets
Chess and Go are adversarial too, but a game has one opponent and fixed rules. Markets are a unique environment for developing and applying AI agents because they combine four things:
Multiplayer and adversarial. A market has millions of participants with different goals, time horizons, and information. Each of them adapts to what the others do. Our models have to keep working as the other participants change their behavior.
Verifiable rewards. A trading model has one objective: maximize risk-adjusted return. Every decision is scored by the money it makes or loses. No human rater, rubric, or preference model decides what counts as good. This is reinforcement learning with verifiable rewards, where the market is the verifier.
Direct revenue. Protogon has no customers. We trade our own capital at our own risk, so the profit our models make is both their training signal and our revenue.
Broad scope. Prices respond to anything important that happens in the world, from elections and earnings to supply chains. So anything true a model learns about the world can improve its trading.
Our goal
We are building AI agents that understand the world well enough to act in it on their own. We believe training them in markets is the most promising path to superintelligent AI.
