brightlab.Made for curious mindsBrightlab
Bonus materials / Research made explorable
What is happening inside a language model—and how could we check?
Follow a clue. Change a computation. Test your explanation. Two separate learning paths turn Goodfire’s research into hands-on investigations.
Read, predict, try, explain
Start with Pip, Wattle and Kiki. No coding experience needed; work with a partner and use the words-to-know guide.
01 / Prediction and uncertainty
Read two fictional facts, then inspect all the words a tiny trained model could choose. Change the facts or the temperature and see what happens.
02 / Features and counterexamples
Look at hidden unit activity across six fact cards. Suggest a label for a clue, then try to break your own explanation.
03 / Cause and effect
Turn off one attention head in a tiny transformer and compare its answer with the original on exactly the same card.
04 / Cyclic representations
Move around a calendar, record the ordinary sum and then wrap back to January. The circle is a model you can fully inspect.
05 / Context and updating
Help Pip weigh fictional weather clues. Move the starting guess, clue reliability and number of clues; the probabilities update live.
06 / Checking a detector
Test a fictional checking rule on 20 labelled cases. Move its threshold and try a set where the clue stops working.
Measure, intervene, challenge, defend
Start with causal patching, then choose a research question. Controls work without coding; Python extensions support deeper investigation.
01 / Residual streams and patching
Inspect a trained two-block transformer, corrupt a colour binding and patch one clean residual vector into the corrupted run.
02 / Superposition and SAEs
Train a small sparse autoencoder on known mixtures. Change dictionary size, rarity and sparsity pressure, then evaluate on untouched mixtures.
03 / Geometry and intervention
Compare chord and arc interventions on an explicit circle, inspect decoded angles and quantify off-manifold distance.
04 / Probes and distribution shift
Train a logistic probe, flip a shortcut only in held-out data, and compare sequence mean with covariance features.
05 / Data effects and logit differences
Compare two actual tiny-transformer checkpoints. One was deliberately taught a biased Kiki colour rule; amplify their logit differences and inspect held-out effects.
06 / Rollouts and early exit
Sample continuations from a known branching process. Change rollout count, smoothing and early-exit confidence, then test a late reversal.
07 / Decomposition, memory and interference
Decompose the tiny transformer’s actual learned output matrix, edit a retained component, and rerun frozen evaluation prompts. Inspect both group performance and the weighted average.
08 / Rewards, rare failures and evaluation
Select candidates using a proxy reward, then audit those same selected answers against independent labels. Vary candidate quality, pool size, audit size and the proxy rule.
09 / Real LLM activation measurements
Explore measured Pythia activation patches, compare token positions and layers, and optionally repeat the experiment on native CPU or H100.