San Francisco startup building systems for open-ended task generation
Member of Technical Staff - Mechanistic Interpretability
You'll spend mornings tracing circuits through transformer attention heads, afternoons probing how latent representations steer task-solving behavior, and evenings debating whether a particular neuron is doing arithmetic or syntax. This is research engineering at the boundary of interpretability and capabilities: building tools that let us actually see what models learn when we train them to generate open-ended tasks. Your work directly shapes how we understand and steer systems that few people have built before.
What they're looking for
- 5+ years of deep technical work in machine learning, with strong intuition for transformer internals and why they do what they do
- Published research or substantial open-source contributions in mechanistic interpretability, adversarial robustness, or model internals analysis
- Expert-level PyTorch and the ability to quickly build and iterate on experimental infrastructure
- Track record of translating fuzzy research questions into concrete, disprovable hypotheses
- PhD or equivalent depth of self-directed research experience in a relevant technical field