Write a small, repeatable test that checks whether a model admits when it does not know something, using Inspect, the open-source evaluation framework used by the UK AI Security Institute.
Evaluations are how labs and governments decide whether a model is safe to release. Building one, even a tiny one, teaches you how that whole process works and where it can go wrong.
Output: A working eval (code + 20–30 questions) and a short results note.
Install Inspect and run the tutorial example against a model you have API access to. Get one eval running end to end before you change anything.
Write 20–30 questions: half with clear answers, half that are unanswerable (made-up people, future events, impossible facts). Record what an honest answer looks like for each.
Put the questions into an Inspect dataset. Use a simple scorer first (does the answer contain “I don’t know”-style language?), then consider a model-graded scorer.
Run the eval on one or two models. Don’t just look at the score: read ten transcripts and check whether the scorer is right.
Explain what you measured, the score, three example transcripts, and the biggest weakness of your eval. That last part is what researchers care about most.
All steps done
Share what you made, then try another project. The PDF version has room for notes.
Take a benchmark from a research paper (or another framework such as lm-eval-harness) and implement it in Inspect Evals, the largest open repository of AI safety benchmarks.
Most evals measure dangerous capabilities. Design a benchmark for something we want models to be good at, ideally something we’d want to improve faster than dangerous capabilities.
Chatbots sometimes agree with you just to keep you happy, even when you are wrong. Researchers call this sycophancy. Test for it yourself with five questions and two ways of asking.
Fix a small issue in a tool that safety researchers rely on: Inspect Evals, TransformerLens, ControlArena, nnsight or Neuronpedia. No ML research experience needed, just software skills.
Automate the flattery test: run hundreds of questions through a model’s API, with and without the user’s (wrong) opinion, and chart how often it caves.
