TechnicalA few weeksIntermediateFree Colab / laptop

Tackle a beginner-graded open problem in interpretability

Neel Nanda’s list of 200 open problems grades each one by difficulty, including problems marked as good for beginners. Pick one, scope it down and investigate with TransformerLens.

Download PDFCode

Why it matters

Small, careful interpretability results on toy or small models are valued and publishable as blog posts. It is also the standard path into mentored programmes like MATS and SPAR.

Output: A notebook and a LessWrong-style write-up.

Getting started

0 of 5 done

Skim the list and filter for problems graded as beginner-friendly.
Mark as done
Check the newer “Open Problems in Mechanistic Interpretability” paper to see if your problem has moved on.
Mark as done
Get a small model running in TransformerLens on Colab.
Mark as done
Spend a day on the quickest possible experiment that could prove you wrong.
Mark as done
Write up what you found, including dead ends.
Mark as done

Key terms

Mechanistic interpretability: Interpretability that tries to reverse-engineer a model’s internal computations into understandable mechanisms, a bit like reverse-engineering a compiled program.
Circuit: A small group of components inside a neural network that work together to perform one identifiable job.
Feature: A direction inside a model’s internal activity that seems to stand for a recognisable concept, like “the Golden Gate Bridge” or “code that contains a bug”.
Interpretability: Research into what happens inside an AI model: which internal parts represent which ideas, and how they combine to produce an answer.

Source

200 Concrete Open Problems in Mechanistic Interpretability (Neel Nanda) (opens in a new tab)

Questions?

Ask in AISCU (opens in a new tab) or the AI Alignment Slack (opens in a new tab).

Related projects

TechnicalA weekendGuided

Find a pattern-copying circuit

Follow a guided notebook to find the induction heads in a small transformer, the tiny circuit that lets models continue patterns they have seen earlier in the text.

CodeView project

GovernanceA few monthsIntermediate

Pick an open problem in technical AI governance

Reuel et al. map open problems by capacity (assessment, access, verification, security, operationalisation and ecosystem monitoring) across data, compute, models and deployment. Choose one cell and scope a project.

Reading & writingCodeDataView project

TechnicalA weekendGuided

Make your first open-source AI safety contribution

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.