Open source chess developers warn about a commercial engine based on Stockfish

https://en.wikipedia.org/wiki/Attribution_(copyright)

For instance, let's look at the BSD license. It requires copyright notices in source code to be preserved. That means that it reserves the attribution right.

The original license required them in executable code; the two-clause version relaxes it; it doesn't demand any expression of attribution in the compiled code.

However, it does not appear to fully waive the attribution right; it requires the copyright notices and license clauses to remain in the source code. The compiled code is still the original author's copyright; nothing in the license says that it isn't.

If someone claims they developed it, it's infringement on the attribution copyright, even though they have the license-granted right to redistribute the code without making such deceptive claims.

IANAL, but if you're going to plagiarize, I suspect the safest bet is to pick something that has lapsed into the public domain, not something with copyrights and licenses in it.

(1) an evaluation function, that quantifies black/white advantage in a chess position.

(2) a search algorithm, that iteratively checks move sequences from a starting position using (1) and returns a recommendation of the best move to play.

Historically (1) was handcrafted, but it recently switched to a 'NNUE' neural network architecture (now it using different evaluation functions depending on how balanced the position is).

This is a good video if you want to understand how Stockfish works.

https://www.youtube.com/watch?v=pUyURF1Tqvg

"But if it turns out to be that heuristics-based approach has lost to NN, it isn't that clear anymore why these have to be 2 separate projects. At least, in terms of the rest of the engine."

Leela's evaluation function is a deep neural network, inference can only be carried out efficiently using GPU, whilst the Stockfish system has been optimised for a CPU-based computer architecture. The innovation of NNUE is that inference can be performed efficiently on a CPU, partly because the network is much smaller, and partly because the architecture is designed to be efficient at recomputing evaluations of only a single piece has changed.