About the role
The interpretability team reverse-engineers what large models compute, feature by feature and circuit by circuit. Using tools such as sparse autoencoders, transcoders, activation patching and attribution graphs, the team tries to explain why a model gave a particular answer in terms humans can check, and to find the places where the explanation and the behaviour disagree.
Understanding is not decoration. Alignment uses the team's findings to detect deception and unwanted goals, Post-Training uses them to diagnose why a fine-tune changed behaviour, and the safety case for each release cites interpretability evidence. Agents with real tool access make this more pressing, since an internal plan is worth inspecting before it is executed.
Work is a mix of tool-building and discovery. You will write infrastructure to cache and slice activations from large models, then spend days on a single circuit. Claims are tested by intervention, not by plausible stories, and negative results are published internally.
What you’ll do
- Train and evaluate sparse autoencoders and related dictionary-learning methods on residual streams, MLPs and attention outputs of large models.
- Trace circuits end to end, using activation patching, ablations and attribution methods to establish which components cause a behaviour.
- Test whether discovered features are causally meaningful by steering and intervening, and report the cases where they are not.
- Build scalable tooling to collect, store and visualise activations from models with hundreds of billions of parameters.
- Apply interpretability to safety questions, such as detecting hidden objectives, unfaithful reasoning or sandbagging.
- Write clear internal research notes, with figures that let others verify each claim.
What you’ll bring
- Research experience in mechanistic interpretability, or in a neighbouring field such as neuroscience, causal inference or representation learning.
- Detailed understanding of transformer computation: residual stream, attention heads, MLP layers, superposition and polysemanticity.
- Strong linear algebra and a feel for high-dimensional geometry.
- Proficiency in Python and PyTorch, including hooking into and modifying the forward pass of large models.
- A sceptical streak: you design the control that would prove your hypothesis wrong before announcing it.
- Ability to communicate subtle findings with precision, avoiding anthropomorphic shortcuts that the evidence does not support.
Nice to have
- Published work on sparse dictionary learning, circuit analysis or probing.
- Experience scaling interpretability pipelines across many accelerators.
- Background in neuroscience or computational biology methods for analysing high-dimensional recordings.
- Research on interpretability for multimodal or reinforcement-learning-trained models.
Work you might take on
- 01
Train sparse autoencoders at multiple widths on a frontier model and quantify how feature splitting affects reconstruction and interpretability.
- 02
Map the circuit a model uses for multi-step arithmetic and check each component with targeted ablations.
- 03
Build a feature-steering test harness that measures side effects on unrelated capabilities.
- 04
Investigate whether internal representations reveal when a model's stated reasoning differs from its actual computation.
Compensation
$320K – $540K
The US base salary range for this role. Every offer also includes equity in Clexerl and, when it spins out, in the venture you help build. We tell you the reasoning behind your number.
Benefits
Health, fully covered
Medical, dental and vision for you and your dependants, premiums paid in full.
Two kinds of equity
Equity in Clexerl, plus a stake in the venture you help build when it spins out.
22 weeks parental leave
Fully paid, for every parent, however your family arrives.
Rest that sticks
Flexible time off with a four-week minimum we actually check on, plus a company-wide week in December.
Learning and compute
A yearly budget for courses, conferences and books, and compute credits for side projects.
401(k) with a 4% match
Matched from your first paycheque and vested immediately.
How we hire
Apply 5 minutes
Send links to what you have built and a few lines on why this role. No cover letter needed.
Intro call 30 minutes
A conversation with the hiring manager about the work, the team and what you want next.
Work sample 2–4 hours
A paid take-home or a pairing session on a real-shaped problem. You pick the format.
Studio day One day
Meet the team, go deep on your work and ours. Remote or in person.
Offer Within 5 days
A written offer with the full range, the equity and the reasoning behind both.
Opens an email with a short template. Links beat résumés.