POPA ReconLearn mechanistic interpretability
Learn mechanistic interpretability Mechanistic interpretability is about opening up neural networks and reverse-engineering what they are actually doing. Instead of treating model weights as random noise, the curriculum shows that models often learn readable structures, features, circuits, and even clean algorithms The six projects build from simple to frontier-level interpretability. You start by visualising digit-classifier weights, then learn superposition, grokking, induction heads, activation patching in GPT-2, and finally sparse autoencoders - a key tool for finding clean features insi…
The excerpt is quoted from a public Telegram channel, which owns its text.