Member of Technical Staff - Mechanistic Interpretability

Our client is a seed-stage applied research lab in San Francisco building AI that learns open-endedly, pushing past the ceiling of imitating human expertise. Their platform automates the engineering that turns proprietary data and evals into rich reinforcement learning environments, so research ships into a real product instead of sitting in a paper. The team is small and unusually strong: 8 people, including RL PhDs and alumni of leading frontier labs.

They have raised $7M to date.

You will use mechanistic interpretability to open up language models and turn what is inside them into training signal, generating intrinsic rewards that supplement or replace human-generated verifiers. This is one of very few places where interpretability and reinforcement learning are being fused inside a shipping product rather than studied in isolation.

• Develop methods for extracting useful training signals from the internal states of language models.
• Turn representations, features, circuits, and causal model behaviors into intrinsic rewards for reinforcement learning.
• Benchmark interpretability-derived rewards against human feedback, learned reward models, verifiers, and task-level outcome rewards.
• Design metrics and baselines for reward quality: alignment with intended behavior, generalization across tasks, robustness, and resistance to reward hacking.
• Investigate how internal representations evolve during RL and post-training, and feed those insights back into training objectives.
• Build infrastructure for reproducible, large-scale experiments across LLM agents, interpretability tools, and RL environments.
• Define and pursue a high-impact research agenda that advances open-ended learning beyond imitation of human expertise.

• 5+ years in reinforcement learning, machine learning, interpretability research, or AI safety (PhD and academic years count).
• Deep RL knowledge paired with hands-on experience in the LLM post-training stack, in a production or academic setting.
• Working command of mechanistic interpretability methods: circuit analysis, feature attribution, activation patching, superposition, and similar.
• Strong programming ability, fluency in PyTorch or JAX, and comfort building AI tooling into your workflow.
• A track record of independently owning and driving a research agenda end to end.
• PhD in ML, RL, AI, or a closely related field. Exceptionally strong and relevant candidates without one will be considered.
• Alumni of MATS or another AI safety research program.
• Experience training LLM agents in multi-step reasoning settings.
• Familiarity with RL training frameworks such as SkyRL or Slime.

• Base salary: $300,000 to $500,000.
• Equity: 0.75% to 1.25%, generous for the stage.
• Full-time position.

• Based in the San Francisco Financial District office. A hybrid arrangement will be considered for exceptional candidates.
• Visa support available: H-1B transfers, TN, OPT, and O-1.