ResearchRemoteRef: DL-2026-001
Research Engineer, Interpretability
Build and validate probes, sparse autoencoders and circuit analyses that find failure modes before deployment, in models already carrying production traffic.
What you'd do
- Design and run interpretability experiments (probing, sparse autoencoders, activation patching, circuit tracing) against the specific models a client is about to put into production, not a benchmark model that stands in for it.
- Turn a suspected failure mode, often reported as one strange transcript, into a reproducible test that catches the whole class of it before a client does.
- Build the internal tooling that makes an interpretability finding checkable by someone else: a script, a notebook with a fixed seed, a diff against a known-good activation pattern.
- Write up methods and findings clearly enough that an assurance engineer without your background can turn them into a specification, and an account team can explain them to a client's safety officer.
- Work directly with the engineers shipping the system: a finding here changes what ships, on a timeline measured in days, not a paper's worth of months.
What we're looking for
- Hands-on experience with mechanistic interpretability techniques (probing, SAEs, activation patching, or similar), demonstrated in code you can walk someone through, not just familiarity with the literature.
- Comfortable reading and modifying model internals in PyTorch or JAX, including hooking into intermediate activations of models you did not train.
- Can explain a technical result, and its actual confidence interval, to someone who has to make a deployment decision on it this week.
- Judgment about when a finding is real versus an artifact of the probe: false positives here cost a client's trust as surely as a missed one does.
- Prior publication or open-source work in interpretability is a plus, not a requirement; a track record of finding real failure modes in real systems counts for more.
Ready to apply?
Fill in the form on the next page, CV included; one submission sends it straight to us.