Leela is not Leela Zero

[dfan]

This post is probably pointless, but I feel the need to try anyway!

I feel like "Leela" is getting used to refer to both Leela and to Leela Zero lately here on the forums, and it means that I'm never 100% sure what people are talking about.

Leela: based on original AlphaGo technology from 2016 (training started by mimicking human games), comes with an UI, latest release Oct 2017, strong dan level.

Leela Zero: based on AlphaGo Zero technology (training is entirely through self-play), started training in late 2017, pro+ level.

Lizzie: GUI for using Leela Zero.

If I had my way, we'd stop saying just "Leela" (because I always wonder Leela Zero is really meant), and call the first one "Leela 0.11" and the second one "Leela Zero" or "LZ".

This is all Gian-Carlo Pascutto's fault for choosing such similar names, although ultimately I guess it is DeepMind's fault for setting a bad example.

[Mike Novack]

This post is probably pointless, but I feel the need to try anyway!

I feel like "Leela" is getting used to refer to both Leela and to Leela Zero lately here on the forums, and it means that I'm never 100% sure what people are talking about.

e a

This will continue to be a problem for us during the period people are coming to grips with what a neural net computer program is. That is because we are used to thinking of programs by name. But a neural net that has been TRAINED to a function is more than just the program << properly done, the training does not alter the program; just creates a set of data that control what function is evaluated >>

The program can have a name

The program plus a particular set of data (it's training weights -- controls the strength of connections withing the net) can have a name.

We just need to be clear which we are talking about.

[Uberdude]

In case there is any confusion regarding the Pandanet cheating case, that concerns "classic" 0.11 Leela, as back in November 2017 Leela Zero was a 2-week old baby playing barely better than randomly so not much use for cheating.

although ultimately I guess it is DeepMind's fault for setting a bad example.

And you didn't even mention AlphaZero!

[macelee]

Just an inconvenience we have to live with. This also causes a particular problem for SGF database collectors. When I collect AI games, I only associate games to one particular name and there's no easy way to document different variations, parameters, versions, hardware settings, etc. It is good that AI games are excluded from rating calculations, otherwise we may face more problems.