Not the most convincing example as it exploits the mix of very low search depth (visit counts) as well as differences in the rules (as well as the fact that tra

Not the most convincing example as it exploits the mix of very low search depth (visit counts) as well as differences in the rules (as well as the fact that training the program to not be vulnerable to exploiting the rules would be just a huge waste of time while it would be quite easy to just add a special case if needed). That being said we know NN based game playing programs have tactical weaknesses - by that I mean they are weak in recognizing forced long sequence of moves. In the beginning it was about a simple ladder in pattern when it comes to Leela and simple 3-4 moves combinations an amateur chess player could see in Lc0 (while both were already much stronger than humans overall). To this day Lc0 has tactical blind spots a competent human could see, mainly when it comes to spotting perpetual check. Those are common and you run into them all the time when using Lc0 for analysis (that's why you always need to double check with Stockfish as Lc0 could have missed a relatively simple forced draw). I think it would be more productive to go to project's Discord and ask "hey I am a researcher working on weaknesses of current NN based game playing programs, can someone point some known issues/weaknesses in current programs?". You will get a long list, no need to waste time to find (very artificial in this case) scenarios.