"Koko ga Ao!" (Here is Blue) - Lizzie Japanese

[ez4u]

Heard last week at the Nihon Ki'in (by one of my teachers, not by me) used by Japanese pros analyzing their games. The "here is blue" expression comes from Lizzie's painting of the best move as a blue circle.

In addition, it seems that references to who has the advantage are increasingly expressed as percentages (a la LZ winrates) rather than points or other expressions.

Black's lament, "I thought I was 80% but it turns out White is 90%!" (a comment last week that was apparently followed by general laughter)

[EdLee]

Real-time evolution of language.

[Uberdude]

Black's lament, "I thought I was 80% but it turns out White is 90%!" (a comment last week that was apparently followed by general laughter)

Do you know which bot they tend to use that this percentage refers to? What LZ 191 calls an 80% lead Elf v1 probably calls more like 98%. I find Elf useful for reviewing openings, but if you get into a lopsided position in middlegame Elf's %s are not so descriptive as 99.8 and 99.9% are only 0.1% apart but could represent a big mistake, whereas LZ saying 79% to 81% is a more useful scale.

[ez4u]

Black's lament, "I thought I was 80% but it turns out White is 90%!" (a comment last week that was apparently followed by general laughter)

Do you know which bot they tend to use that this percentage refers to? What LZ 191 calls an 80% lead Elf v1 probably calls more like 98%. I find Elf useful for reviewing openings, but if you get into a lopsided position in middlegame Elf's %s are not so descriptive as 99.8 and 99.9% are only 0.1% apart but could represent a big mistake, whereas LZ saying 79% to 81% is a more useful scale.

I do not know. In any case, I understood this comment to reflect Black's personal (mark 1 eyeball) assessment that was later contradicted by checking with Lizzie. In other words people are starting to express non-AI assessments using AI-like terminology.

[jlt]

Winning percentages depend on the bot and the number of simulations, whereas a number of points has an absolute meaning and is more useful for humans. It may happen, during small yose, that LZ evaluates the winning percentage as 70%, and yet one player is leading by only one point.

[EdLee]

LZ evaluates the winning percentage as 70%, and yet one player is leading by only one point.

Thus sometimes human pros lose 0.5 games with 95% winrate.

[Uberdude]

LZ evaluates the winning percentage as 70%, and yet one player is leading by only one point.

Thus sometimes human pros lose 0.5 games with 95% winrate.

Or bots lose games they are winning with 99% winrate if they got the komi wrong viewtopic.php?p=239199#p239199

whereas a number of points has an absolute meaning and is more useful for humans.

What is the absolute meaning, the score if 2 perfect players continue the game from here? But
1) we don't know what perfect play is, we can only make a guestimation of it. Territory estimations include assumed hypothetical sequences which will almost certainly not be optimal (e.g. Go Seigen probably counted some centre points in his famous game with Shusai, but Shusai/Maeda's myoshu meant that wasn't the case).
2) what does a perfect player (or our of attempt of one) try to do when losing? keep the margin of loss as close as possible, try to win in case the opponent is not quite perfect and thus risk a bigger loss for some chance to win?

But yes, winrate to points margin is some fuzzy function that varies by bot, phase of game, phase of moon etc; for example when I played a 4 handicap game with Shi Yue 9p I maintained a >90% lead according to LZ pretty much all game (until blunder death in small yose), at the start that corresponded to a 4 stone lead you might call a 50 points lead, at the end it was more like 3 points lead.

[jlt]

When I said that the number of points has an absolute meaning, I meant that the losing (perfect) player wants to keep the margin of loss as close as possible. In other words, "B+17.5" means that

• There exists a strategy for Black so that, whatever White does, Black wins by at least 17.5 points.
• There exists a strategy for White so that, whatever Black does, White doesn't lose by more than 17.5 points.

Assuming komi is 7.5, a bot in which dynamic komi is implemented can estimate that Black leads by 17.5 points if, with the current position and komi 25, the winning probability is close to 50%. More generally, the estimated score is B+n if n is the half-integer such that the winning probability is closest to 50% when White is given n+7.5 points komi.

Of course, in a real game, humans or bots won't play like "perfect losers", but will try to complicate the game and hope this increases the probability that their opponent makes a mistake.

[Leon]

A bit off topic, but does anyone know how strong Leela/Elf is on what hardware?

Professinoals are starting to use Ai a lot, but how much harware is acutally needed to best the top pros?

[Uberdude]

A bit off topic, but does anyone know how strong Leela/Elf is on what hardware?

Professionals are starting to use Ai a lot, but how much hardware is actually needed to best the top pros?

I've been attempting to answer that myself playing with my 1060 GPU on Fox, but don't know if I've played any pros let alone top ones yet. But that hardware easily beats plenty of human 9ds with decent win-ratios, loses to other 9ds using bots with better hardware, and can occasionally lose to human 9ds in complicated middle-game situations (particularly with faster time limits).