Many people love using Leela Zero to study the game of go, but there seems to be less and less people contributing to the training of Leela Zero. Currently there are less than 50 clients helping out with the project. It seems that if the trend continues, the project will no longer progress at all. Hope more people will contribute their computing power if they want Leela Zero to be able to continue finding stronger networks.

_________________
http://tchan001.wordpress.com
A blog on Asian go books, go sightings, and interesting tidbits
Go is such a beautiful game.

I noticed this trend too. Recently there are usually 2500-3500 self-play games generated per day. The project generated the first 20 million games in around 2.5 years. When gcp wanted to stop the run last year (having achieved the objectives that he had when he initiated it in 2017), various people thought it should be continued out to at least AlphaGo's 30 million game level. However, at the current rate, that would require another ten years. That is assuming that the number of contributors does not continue to decline!

As far as I can tell, the impact of AI on professional Go (in Japan at least) was driven primarily by the unrestricted availability of LZ to professional players (helped by great tools like lizzie!). It seems to have quickly become the tool of choice, used by everyone and providing a consistent, easily shared set of results that promoted the discussion and analysis of our beloved game.
--> What an incredible result for a project that began with such humble goals!

--> To gcp and all the major contributors who contributed ideas and code, I say, "Congratulations on a great idea (open-source, distributed replication of the AlphaZero approach), wonderfully executed!"

--> To all the distributed computing contributors, I say, "Thanks for enabling the creation of a wonderful tool and literally revolutionizing Go!"

For a couple of years I left my pc running at night, generating LZ games. Admittedly I did not contribute much since my hardware is relatively weak. But it made me feel good!

Currently, however, I think that the original idea of duplicating the "pure" AlphaZero approach has outlived its usefulness. I never use LZ these days, only katago. I am waiting with bated breath for lightvector to put forth some further ideas or take up some new suggestions. Meanwhile I dutifully updated lizzie to the 273 net for LZ back in May but never used it. I still have not downloaded the 274 net that promoted in July. I just forgot to do it and did not even notice until I started to write this post.

_________________
Dave Sigaty
"Short-lived are both the praiser and the praised, and rememberer and the remembered..."
- Marcus Aurelius; Meditations, VIII 21

The question is : what is LeelaZero goal?

The goal was to reproduce AlphaZero result. This has mostly been achieved, with great success.

So, what now? Further training isn't really usefull at this point, unless there's a new goal.

If duplicating AlphaGo has been the aim, does Leela Zero now read ladders, life and death etc. correctly for all practical matters?

It does not (not always), but we don't know how well AG did either. Duplicating Alphago was only ever possible in broad sense, like very roughly similar strength - the codebase, bugs and inneficiencies, the number and usefullness of training games, training problems were never the same.

It seems that the goal of the LZ development has always been to make a strong AI without optimisations for support go specific features, such as variable komi, handicap, ladders, different rule sets etcetera.
Decisions have always seemed to be directed at reproducing DeepMind's results from a computer science perspective. Although it can be used for for analysis by human go players, that was never a main objective AFAIK. For a long time there was not much competition from other projects, so the LZ project more or less had a monopoly of crowd support for its training.

But now there is competition by KataGo, which does support all those go specific features and nowadays it seems to be even stronger than LZ. There is nothing wrong with the scientific goals of the LZ project, but it is becoming more and more a niche activity again for computer scientists.