Research Scientist - Simplex

Senior Research Scientist

At Simplex, we're building a science of intelligence. Our aim is to develop and apply a rigorous theory of latent internal structure in neural networks, and how that structure relates to computation and behavior. We believe that when dealing with intelligence, understanding is safety. Without genuine understanding, we can't reliably monitor, control, or even reason clearly about what these systems are doing.

But these same systems also present us with a new opportunity. For the first time, we have AI complex enough to serve as testbeds for theories of intelligence, including biological. We aim to build a theory applicable to intelligence, both artificial and biological.

We have the beginnings of such a theory, grounded in the physics of information and experimentally verified in transformers. Now, we are scaling our team. Some of the near-term goals we have are building unsupervised methods that recover belief geometries in real LLMs, extending the theory to more complex cognitive tasks, and pushing toward tools reliable enough to matter for safety.

We're looking for people who can do rigorous mathematics and get their hands dirty with real models and data; ideally someone who moves naturally between theory and experiment, and feels deeply driven to understand intelligence.

• You learn across fields. Our work draws on many fields, like dynamical systems, probability, deep learning, physics, information theory, and neuroscience. You don't need to know all of it coming in, but you're the kind of person who picks things up, and follows your curiosity—and surprising experimental results—wherever they lead.

• You have taste. You know the difference between a problem that matters and a problem that's merely publishable. You have opinions about research directions, not just techniques. You care about the craft of how experiments are designed, analyzed, and presented.

• You're self-directed. We're a small team. You'll have real ownership over your work, which means figuring out what to do, not just how to do it. You'll work closely with the research team, but we expect you to develop and pursue your own ideas within the broader research program.

• You communicate. You can explain your ideas clearly to collaborators, in writing, and on a whiteboard. Science is a team activity for us, and that requires being able to think together.

• You build. You're at home in front of a whiteboard and in a terminal. We are building new theory, new code, and new experiments. You think big, but you're serious about it, and you actually try to make things happen rather than just ideating. You use AI tools, you tinker, you're excited about what's becoming possible.

• You have depth in at least one quantitative field such as physics, mathematics, neuroscience, machine learning, etc. A PhD is typical but not required if you've found another way to go deep.

Our foundational result showed that transformers trained on next-token prediction spontaneously organize their activations into geometries predicted by Bayesian belief updating over hidden states of a world model. Even when trained on simple token sequences from hidden Markov models, complex fractals emerge in the residual stream, structures far removed from the surface statistics of the training data. We think of this work as providing the first steps into an understanding of what fundamentally we are training AI systems to do, and what representations we are implicitly training them to have.

Since then, we've pushed in several directions. In Constrained Belief Updating Explains Transformer Representations, we asked how attention implements belief updating when Bayesian inference is fundamentally recurrent. We found that attention parallelizes recurrence by decomposing belief updates spectrally across heads, and we were able to make verified predictions about embeddings, OV vectors, attention patterns, and residual stream geometry at different layers.

We've also developed a theory of in-context learning grounded in training data structure. When training data mixes multiple sources, models must infer not just what hidden state the generator is in, but which source…