Anthropic
OpenAI
Google DeepMindBased on real interviews
Investigate asymmetric factual recall
Explain why reversing the direction of a factual question can sharply reduce a model's accuracy.
Unlock with Full AccessThe challenge
The model knows who created Linux much more reliably than it recalls what Linus Torvalds created. You have weights, activations and training-data statistics to investigate the difference. Commit to a mechanistic hypothesis and a three-month plan, including experiments that could rule out your account.
What you'll practise
Mechanistic hypotheses
Propose a specific explanation for the directional recall gap.
Causal experiments
Choose interventions that test the mechanism rather than only visualising activations.
Revising your account
State what failed predictions would imply for the next investigation.
Your recommendation
Write your reasoning and submit it for feedback.
The full brief and feedback are included with Full Access.