Thank you, very interesting paper (and it is good to see that Kramnik has kept up an active interest in chess computing, hehe).
The paper you cite seems to converge nicely with evidence that large transformer model embeddings for language comprehension map on to human semantic understand (e.g.,:
https://www.nature.com/articles/s41598-022-20460-9
https://www.nature.com/articles/s42003-022-03036-1 )
Related, it appears we are even approaching a refined normative theory & practical benchmarking of hyperparameter transfer, which may be very roughly akin to conceptual understanding (https://www.microsoft.com/en-us/research/uploads/prod/2021/11/TP5.pdf
Truly, a wonderful time to be alive; I feel as if all the detective novels I'd ever read had reached dénouement at once.
I just read Acquisition of Chess Knowledge in AlphaZero and it's both really cool and has interesting implications for the Natural Abstractions Hypothesis. AZ was trained with no human data and yet it settles on relatively interpretable abstractions for chess.[1]
They trained sparse linear probes on hidden layers to predict Stockfish evaluation features. Here are some of the graphs
Analyzing the largest differences between the predicted Stockfish score
total_t_ph
and the actual. There's a clustering effect due to the hardcoded Stockfish eval not taking into account "piece captures in one move" (as usually, the search would handle) - leading to large disagreements on value when a Queen is hanging.Not cherry-picked!
Training a linear regression model to predict the value function, given human interpretable features like piece value weights and Stockfish features like king safety: results in approximately recovering the chess 9-5-3-3-1 piece values!
They also did various experiments with unsupervised probing, here's a visualization of some of those features. (I didn't look into these results as much as the others, so not much to say here.)
In conclusion: I'd recommend checking out the paper, or any from Neel's list. Interpretability is fascinating!
This isn't that surprising, chess is somewhat of a toy example after all. The concrete experiments still provide bits though!