Hi, this discussion arose my curiosity and triggered and idea: run the Khmelnitsky test with a 1-ply Lc0 using a big network. The 1-ply limitation ensures fast response time (almost instant) and focuses on pure chess knowledge.
I used the "TCEC 2022" Lc0 settings I described within this
other thread (about positions testing).
As a result, the KT-score is 1954 and the skills profile is as expected: positionnaly strong, with some tactical weakness.
Strongest dimensions are Opening (Lc0 can definitely play without any external book) and Standard endgame positions: that's chess knowledge!
On the other hand, Tactics and Sacrifice are the weakest skills (obviously due to the lack of any depth of search). Nevertheless, the neural network enables some capability to recognize threats and defend itself, and provides an ersatz for Calculations. Strategy and Endgame are strengths.
Comparison to a same level human player:
Human chances lie in the middlegame, where opportunities for tactics, sacrifice, attack/counterattack should be looked for. Potentially a nice training!
Grüße,
Eric