One belief driving my research department: today’s LLMs aren’t the path to AGI, or even to genuinely better AI. What’s missing isn’t more data or bigger neural nets in PyTorch. It’s the human touch: real physics, real video, real life experience to learn from, not just text. A model trained only on text is trained on a description of the world, not the world itself, and that gap doesn’t close no matter how much text you add.
This is why our “Quantum Reduction” and “Project Fly” initiatives exist. We’re studying how actual living beings learn, and how AI needs to learn the way beings do, not to replace us, but to genuinely help humans, animals, even trees. That distinction matters. The goal isn’t a system that mimics understanding well enough to pass as intelligent. The goal is a system that actually understands, because understanding is what makes help real instead of accidental. We won’t find real solutions in something that has no grasp of what it feels like to win, to want, to admire, or to care. Those aren’t decorative feelings. They’re the reason living things solve problems the way they do, and a system with no access to any of that is working with half the picture.
A system like this may be controversial, but it should be governed by laws, not code. Laws it actually understands, not laws it merely fears. That’s a deliberate distinction, not a rhetorical one. Fear based constraints hold only as long as the fear does. Understanding holds because the system actually grasps why the constraint exists. And the fear driving most of today’s caution comes from the idea of AI developing real feelings, real agency, something like a will of its own. But biology, math, and physics haven’t proven that’s even possible. No one has shown a mechanism by which it would happen. Until someone does, we should stop treating that fear as a design constraint and focus instead on what we can actually build: AI that is genuinely intelligent, capable of accomplishing real, hard tasks.
Tasks like discovering new drugs and materials, modeling climate systems accurately enough to act on, diagnosing rare diseases from incomplete data, designing fusion reactor containment, or controlling robots that operate safely in unpredictable physical environments. These aren’t hypothetical stretch goals. They’re the actual bottlenecks holding back progress in medicine, energy, and climate science. And they’re exactly the kinds of problems current LLMs can’t meaningfully touch, because none of them require fluent language. They require a model of the physical world that language alone can’t supply.
That’s what pushes the real question to the surface. An LLM follows the orders it’s given, but that raises the real question: what actually makes it better at following them? More data centers, or motive? We don’t actually want just a thing that knows. We want a thing that understands what we are and how we operate. If it only learns physics, we’ve built a better calculator. If it only learns language, we’ve built a large language model. Neither is what we’re after. If we want something more advanced than either, the model to learn from is the one we already have proof works: the most advanced beings we know of are human. We should learn from the best to be the best.
The research this belief drives: Project Quantum Reduction →
The full research log →