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.
