The opportunity
When you see what modern language models are capable of, do you wonder, "How do these things work? How can we trust them?
What you'll do
Our research blog: covering advances including Monosemantic Features and Circuits
An Introduction to Interpretability from our research lead, Chris Olah
The Urgency of Interpretability from CEO Dario Amodei
Engineering Challenges Scaling Interpretability: directly relevant to this role
Minutes segment: Around 8:07, see a demo of tooling our team built
New Yorker article: what it's like to work on one of AI's hardest open problems
What they're looking for
- Pretraining: Training dictionary learning models looks a lot like model: pretraining - creating stable, performant training jobs for massively parameterized models across thousands of chips
- Inference: Interp runs a customized inference stack. Day-to-day analysis: requires services that allow editing a model's internal activations mid-forward-pass - for example, adding a "steering vector"
- Performance: Like all LLM work, we push up against the limits of hardware and: software. Rather than squeezing the last 0.1%, we are focused on finding bottlenecks, fixing them and moving ahead given rapidly evolving research and safety mission
- Build and maintain the specialized inference and training infrastructure that: powers interpretability research - including instrumented forward/backward passes, activation extraction, and steering vector application
