jma127
commented
6 years ago Thank you for your thoughts! We indeed hope that LeelaZero and other projects will view our work first and foremost as a valuable resource.
Closed bochen2027 closed 6 years ago
jma127
commented
6 years ago Thank you for your thoughts! We indeed hope that LeelaZero and other projects will view our work first and foremost as a valuable resource.
Okay so I'm trying to think this out logically. Facebook releases ELF OpenGo, instantly becoming by far the strongest public open source /weights Go AI program, overnight gets adopted as the baseline by many other programs.... so going forward most programs are going to be more or less the same strength if not simply identical whitelabel altogether, so what is left? Seems to be the "engine" part is now more or less solved and even I would say undifferentiatable. superhuman on a gtx 970, is basically the end of the road.
The remaining things of innovation are GUI, analysis, high handicap, different komi, teaching tools, etc etc
And marketing/branding/mindshare/PR for the Go bots (including Leela Zero) will perhaps be more important than ever before. We gonna see a consolidation of Go AI bots and my guess is only one or two will survive this, and that is if they are lucky.
Some immediate implications is that essentially its killed commerical Go at least from the standpoint of selling engines go. We can't compare to Chess because not only does Chess have an order of magnitude larger userbase esp in the West, Chess also enjoyed a good two decades whereby classical algorithms and programming made it such that there was a healthy ecosystem of different engines completing with one another for top listings. WIth the advent of the "zero" method, all zero programs converge to the same ultimate state and its just a matter of compute. There is really nothing left to do. More or less.
This also means there is little to no more point in having Go AI engine competition and matches. We already see cgos is defunt and its benchmark is less and less useful, UEC cup ended, Zen pulled the plug and called it quits, I seriously doubt we'll see another version or edition of CrazyStone, and now with so many engines adopting the facebook weights, whats the point? I see this as portending the demise of Go AI competitions and engine vs engine games as well. Think about it, LZ had beat DolBaram in that last competition match, now DolBaram adopts ELF weights, and ELF is stronger than both Pheonix and FineArt... it doesn't take much to put two and two together and see where this is headed... Didn't Golaxy just beat Ke Jie last week? Ill bet that was the shortest triump ever. And whatever aire of exclusitivity that FineArt enjoyed prior to the facebook event has now been obliterated, top pro in China no longer need to use FineArt to get competitive advantage in training when everyone in the world on half a decent graphics card can now run the same or better. The implications are indeed far fetching.
Lets examine the distributed community based crowd computing aspect angle. It took the public six months to get LZ to top pro level from scratch and yet facebook only needed two weeks and argueably far surpassed top pro levels and went deep into superhuman arena. Not that I know it is going to happen, but there is nothing to prevent facebook from doing it again, say another couple months down the road it can sudden drop a new weight that will be the new state of the art and far surpassing anything any community effort could have hoped to come up with within that allocation of time. Who knows maybe Google will see all this and publish the AGZ weights, or maybe in another few months the second round of weights that facebook puts out will far suprass AGZ altogether! In light of the recent developments these are all realistic possibilities now! But none of these possibilities foster morale for community initatives.
I'm thankful that prior to facebook dropping ELF onto the world, that LZ already reached and imho surpassed top pro level on its last/final network 131, (I see 132 just came out hours after the Haylee game 2 and is 60% stronger!) and that LZ project was able to convert ELF weights into native LZ format so that it can be used just like any other weightfile and now its even working great in Lizzie.
I hope that Leela Zero project finds a way to position itself to best take advantage of this new and changing landscape. By far it enjoys the most mindshare in the community of Go at large right now and I hope it continues to evolve and find ways of remaining relevant and bringing value to people's lives.