No. Much more interesting would be to eliminate this calling out to Stockfish entirely. It's just not just a matter of the title, the data in the paper is fundamentally tainted by it. You get less data on how the modell actually plays this way, how strong it really is, and how it could be integrated with search in all positions, which the obvious next step(i.e the model + some search algorithm, not simply calling out to another engine).
The fact that they're willing to do that just to get a better headline is pretty telling.
In chess, games are never won until checkmate or resignation by the opponent. If it fails to get there, it just didn't win the games. It doesn't matter how winning the position was at any other point during the game. That's why it's tainted. The results are changed, and the games themselves are changed. Therefore the comparisons to humans and other engines is changed. And hence, what we learn about the approach presented is diminished and obsfuscated.
I think wrong is a strong word here. You can keep making a bigger model, that's one direction to go for sure. But adding search is certainly an interesting one at the very least. And it's definitely the way forward to see if this is an approach that can advance the strength of chess playing programs. I seriously doubt this would ever outperform a search based paradigm on its own. For one thing it gets no advantage from extra time(which is another reason the grandmaster strength claim is dubious btw; it would be outplayed in longer time controls). For another, search might be necessary to make this thing able to train without the help of Stockfish, by playing itself(iterated amplification and distillation, how AlphaZero was trained). For a third, due to the game complexity of chess, it's likely that model size without search would have to scale exponentially to become stronger.
And all this brings me back to my previous points about releasing weights and code. If they had done so, I would be trying all of these things out myself right now instead of arguing about it on the internet. But time and time again Deepmind have demonstrated that they won't engage with the wider computer chess community. And so their research, which is no doubt innovative and inspiring, fails to have the impact it could have, because the community has to redo all the work with less money and compute. We saw this with Alpha Zero; it took years and a lot of hard work by volunteers for Leela just to get where Alpha Zero was originally and eventually surpass it. And it will probably take a long time to see state of the art competitive engines based on this new approach in the wild. Because Deepmind don't really care about that beyond making a splash with their initial publication. They're only peripherally interested in chess, which is fine, but as a chess player and computer chess enthusiast it makes me sad that they won't engage with the community in an open way.
I made several arguments why search is interesting. Just repeating that it's not, is not really a counter-argument.
Also, the fact that AlphaZero has search is irrelevant. This is a completely different model architecture. Stockfish had search before AlphaZero, does that mean AlphaZero wasn't interesting either? Or hell, why not just say all computer chess developments are uninteresting because alpha-beta was invented in the 50s?
I guess I'm saying their goal here was to do no search and get high performance, not create the highest possible performance. The trade off between run time search and more training (better models) is an explicit area of research, and this paper is in that realm. Noam Brown has talked a lot about this in the context of poker and diplomacy.
The fact that they're willing to do that just to get a better headline is pretty telling.