Good direction, but I have two points:
1) Use GPU instead of CPU for NN. This will allow all CPU power to be used to perform A-B searches. Additionally LC0 will work much faster.
2) Let Stockfish search deeper than LC0. This way he will be able to find strong tactical moves.
The right combination of LC0 and Stockfish should be much stronger than each of the component engines.
it would be good if it was true, but in fact it's totally fake.
the engine doesn't use NN file. At all, it's just a file lying around as a decoration, for misleading people.
the engine doesn't use libopenblas.dll either. At all. It doesn't load it. It's just, well, theatrical props.
all those "loaded network, opened libblas, using batch size of 256" messages are just hardcoded strings that the "engine" shows at start without actually doing any of that.
all that runs there is stockfish. But again, to mislead people, nps is just divided by 1000.
there's not a trace of using any kind of neural network "subzero". all the author is trying to do is cheating and lying.
and by the way nn weights file was presented as "own", but in fact it's sergio's 256x20-mod-swa-146700.