Hidden states and linear probes
A · 10-second intuition A probe is a simple classifier trained on a layer's hidden states to test whether some property can be read out from them.
Predict before you read
Probe accuracy for a property peaks at layer 20. What can you conclude?