Drafts
WIPThese articles are currently under development. Please bear with any typos or bugs, and feel free to report them in a discussion on Github.
Evidence of Learned Look-Ahead
Dive into how Leela Chess Zero, a class of neural networks to play chess, implements planning through a look-ahead mechanism. We'll break down the key findings of the linked article and explore the interpretability techniques used to study the model, like probing and activation patching. These findings will be reproduced using the library lczerolens, a tool for chess XAI.