Summary
What you’ll impact
The Head of Core AI leads the research team that creates models, agents, and training systems to advance physics‑focused AI. The role applies deep learning, reinforcement learning, transformer, graph, and neural‑operator architectures using large‑scale frameworks such as PyTorch or TensorFlow. Responsibilities include defining the research roadmap, building and evaluating models, delivering production‑ready components, and meeting rigorous scientific validation standards. The individual collaborates with Applied AI and Physics groups, manages hiring and performance, and stays hands‑on by reviewing code and experiments.
Responsibilities
What you'll do
- Lead Core AI, the organization that builds the models, agents and training systems that do physics: own the people, the delivery and the hiring.
- Set the roadmap and own the research portfolio against one goal, making PSI's AI systems better at real physics: set direction, kill weak lines, double down on what works.
- Hold the work to one bar, physics a real physicist would trust, and own the training and evaluation infrastructure the company depends on to get there.
- Partner with Applied AI, which turns Core AI's work into products, and Physics, which supplies the problems and the verification; build on the platform Engineering runs.
- Stay technical: carry individual work, review code, and set the bar in person.
Requirements
What you’ll bring
- Eight or more years of AI research work and a track record of leading AI research teams that shipped: you have run an industrial AI research group at a frontier lab, a fast-moving AI company, or equivalent, and the work made it into production systems or into results the field recognizes. At least three years of direct people management: hiring, performance reviews, and career development for the people who reported to you.
- Research vision on a portfolio. You pick the question and design the experiment that settles it, state the result that would kill a line and stop it when that result arrives, know when a benchmark is being gamed, and trade elegance for something that ships this quarter.
- Deep fluency across modern ML, with depth in one of the following and a working floor on the rest. You have personally built and trained systems that beat non-trivial baselines, and you can still do so.
- Reinforcement learning and post-training: policy optimization, reward design, RLHF, RLVR, GRPO, RL for language models and for control.
- Agents, evaluation and verification: tool use, long-horizon reasoning, memory and harnesses, evals, verifiers, reward models, instrumentation that separates reasoning from gaming.
- Model architectures and training methods: transformers, state-space and graph models, neural operator and surrogate architectures for physical fields and meshes, scaling laws, optimization, training dynamics.
- Scientific and generative modeling: learning from simulators, physics-informed and conservation-aware models, out-of-distribution behaviour, diffusion and generative models, code generation with checkable results.
- Motivation for this specific job is assessed at the first conversation: why this over a frontier lab.