What I Want to Build: Intuition, Latent Reasoning, and the AI Scientist

The goal that ties my research together is simple to state and hard to build: AI with strong intuition—systems that guess well before they reason at length, and that use reasoning to verify the guess rather than to substitute for it. I want to build machines that see the answer before they argue for it.

Thinking out loud has a cost

People solve hard problems fast. The answer often arrives before the argument does; the write-up comes later. Today’s models do the opposite: they think out loud, at length. Long chains of thought are expensive, and past a point they actively hurt—the model overthinks, contaminates its own context, and talks itself out of an answer it already had.

Experts see first, then verify

Expertise is not deliberating for longer. It is having better first guesses. AlphaGo’s search existed to confirm the intuition of its policy network, not to replace it: see, then check. I want models that can see a blurry version of the answer directly, then spend compute sharpening and verifying it—instead of paying full price for every step of a trajectory that a good intuition would have skipped.

So what is actually missing?

Predicting the next token of human text teaches a model to reproduce reasoning rather than to compress it. And forcing every intermediate thought through a single discrete token squeezes it through a very narrow channel. Both are design choices—not laws of nature—and both can be changed.

From reproducing knowledge to creating it

Strong intuition is, I believe, the missing ingredient for AI that creates knowledge instead of reproducing it: systems that form their own hypotheses, explore on their own initiative, and generalize into genuinely new territory. That is the loop I care about most—the one that turns a question into an experiment, an experiment into a result, and a result into the next question.

How I’m building it

Five threads, one question—where does intuition come from, and how do we train it?

• Text diffusion. Generation that is bidirectional and global rather than left-to-right. Any-order prediction gives a model implicit lookahead: a form of planning that never has to be spelled out in tokens. (DiffuCoder, CADD)

• Latent reasoning. Thinking in continuous space instead of in text. Latent thoughts are compressed, more diverse, and far cheaper than a long chain of tokens, and they sidestep the diversity collapse that reinforcement learning inflicts on token policies. Fewer words, more thought. (LaDiR, LaDi-RL, CLaRa)

• Code, agents, and verifiers. Code is where intuition becomes measurable—a good engineer predicts what a program will do before running it. I train coding models and software agents, and the verifiers and environments that let them check a hunch cheaply. (SWE-Gym, CodeAct)

• The AI scientist. Agents that discover rather than retrieve: forming hypotheses, designing the experiment that would discriminate between them, running it, and being honest about the result. The target domains are the sciences, where a wrong answer is checkable. (BED-LLM)

• Recursive self-improvement. A model that manufactures its own training signal—self-distillation, future-aware data creation, and search folded back into weights—so that yesterday’s slow deliberation becomes today’s instant judgment. (SSD)

The bet

Put together, this is one bet: the next big gains come not from thinking longer, but from thinking better—seeing first, verifying fast, and closing the loop between hypothesis and result. That is what I want to build.