Press "Enter" to skip to content

Mark Lefler and Larry Kaufman end of season interview

We all hope that you are safe from the ongoing world crisis. How is the quarantine affecting you? Does it give you more time for Komodo development or on the contrary it is distracting you?

Larry Kaufman: So far well. It is certainly distracting, but on balance I have more time for Komodo development due to curtailing many other activities.

Mark Lefler: I agree. More time spent on Komodo, since, what else are we going to do? I do miss getting out of the house. Little side trips are missed which sometimes inspire new ideas.

TCEC S17 has just finished. How do you see the progress of computer chess?

L: Progress of the computer chess field is still pretty good; the NN explosion has made everything confusing, with comparisons between GPU based engines and CPU based engines being too dependent on hardware to talk about which is really better.

M: It would be useful for all the testing groups to agree with a “GPU speed index” to try and compare the results, like they now scale time controls based on CPU speed.

Are you satisfied with Komodo’s performance in Div P? Did you find areas that you want to improve for future seasons?

L: Komodo is not competitive with Stockfish or top GPU-based engines now, but is clearly stronger than all CPU based engines not derived from Stockfish. We are making steady progress, but the gap is too large to be closed by gradual development. We are hopeful that a future Komodo-nn hybrid will do the trick.

M: We actually have a couple of people who work on NNs for us part of the time. Lots of things being tried. We do not want to just make an Lc0/AlphaZero clone, so we are using the NNs in different ways. We do not know when it will be good enough to publish yet.

L: NN branch is a separate project by two chess.com guys working with us. We have a hybrid KomodoMCTS-nn version now that is close to the level of regular Komodo, but it has to be clearly superior to be marketable. We’re trying several different approaches, all pretty original. Logically it seems to me that some hybrid should be better than either approach, as both the A/B and NN engines now each have obvious superiorities over the other in different situations.

– Your cooperation with Chess.com has been going on for quite some time now. Are you happy with it and what can we expect in the near future?

L: So far the benefits for us have primarily been from using hardware they rented for us, and publicity. But we are hopeful that the NN work by their two guys will pay off soon. They seem to be quite skilled in this field.

M: Chess.com gives us insight to what chess players want, and spurred us to develop Personalities and more Skill levels. One of the chess.com guys is really experienced in SMP, and this has helped a lot with scaling Komodo MCTS to a lot of CPUs. Having 64 processors all updating one tree in memory is darn tricky to get right. He has also picked up how to program and interface GPUs very well, skills I did not have until he made me learn. It is kinda neat saying I need say a utility and then I go to bed. When I awake, it is all done and working. Maybe all software development teams should live about 8 hours apart. If we had three programmers carefully posisitoned, it would be a 24 hour development cycle!

– Season 18 of TCEC and Div P where you participate is coming in just a few weeks time. Would you dare to make a prediction?

L: Although we’re clearly making progress each month, it’s not dramatic enough to expect to challenge Stockfish or the best NNs for a place in the Superfinal at this time.