Working at Primateria

The company

Primateria is an AI research lab building systems that learn to solve scientific problems through computation. We are starting with molecular discovery.

We believe reliable simulations can let AI learn through computational trial and error at a scale that experiments alone cannot support. Our research connects two problems: using AI to improve scientific simulations, and using those simulations to train more capable AI systems. We are applying this approach to discovering and optimizing molecules, with experimental results measuring progress and guiding what we build.

We are based in Palo Alto and have raised $17.3 million from investors including Paradigm, Makers Fund, a16z Speedrun, and Alliance. We are hiring scientists and engineers to develop the computational methods, learning environments, and AI systems behind this effort.

How we work

We work together in person in Palo Alto. Scientists and engineers work closely throughout the research process, from choosing a problem and developing methods to running experiments and understanding the results.

We use AI agents to propose improvements to computational methods, run experiments, and analyze results. We build the tools and scientific benchmarks needed to evaluate their work, understand where they succeed and fail, and make them more effective at research.

One current project extends Orbformer, a neural wavefunction model, to molecules containing 3d transition metals. Alongside this work, we are investigating how computational feedback can train AI systems to make better molecular discovery decisions. These projects let us study both how AI can improve scientific methods and how it can learn from them.

We publish as much of our work as possible so others can inspect and reproduce it.

Compute

Our team has built a dedicated cluster of 69 NVIDIA RTX PRO 6000 GPUs. We use it for scientific simulations, model training, and running research experiments in parallel, enabling rapid iteration on computationally demanding problems. We have more cloud compute available and are currently building out another cluster.