Velocity Matters Most

Products only improve when real users engage deeply, testing everything and surfacing feedback that actually matters. This has always been the core truth behind finding product-market fit, but it takes on new weight for intelligence-native products like foundation models.

What's different now is that models improve specifically through the quality of data they receive, and humans-in-the-loop are the primary source of that quality. When people use a model, inspect its output, and correct it where it falls short, they generate exactly the signal these systems need to get better.

Frontier labs proved this with coding. Large language models became exceptional at code generation because millions of developers ran these tools through real build cycles and produced rich, high-volume feedback, via reinforcement learning from human feedback, whether dense, binary, or continuous in form. That scale of usage, not just architecture or compute, is what pushed coding models toward their current ceiling.

Having crossed that threshold for coding, the industry's next move is clear: build specialized models for other domains and subject them to the same feedback-rich testing. The pattern that will define this next wave is simple: whoever can generate rich, continuous feedback at scale wins, because that requires getting real users to adopt the product and tolerate its imperfections long enough to improve it.

Biology makes this dynamic unusually visible. AI-designed drugs only get better once humans actually try them, and the velocity of that feedback loop is amplified by patients desperate enough to bypass traditional clinical trial phases for a chance at treatment. Robotics follows a similar logic: mass deployment lets robots collect real-world failure data, which is precisely what sharpens their ability to predict future states.

This same feedback-driven transformation is unfolding across climate prediction, voice prediction, world modeling, human biology, robotic action, and code generation. The winners in each category this decade won't just be defined by model architecture; they'll be defined by how fast they can get people testing, generating rich feedback across every scenario, and feeding that back into the system.

Speed of execution, brand trust, and real-world penetration now matter as much as the underlying model, because a model is only as good as the data it can earn. Velocity and need are what matter most for a startup executing today.