[ProllyInfamous]

Watching the human programmers become dumbfounded as AlphaGo invented novel Go-playing strategies... is what I remember most from having watched this a few years ago (right before GPT3.5/ChatGPT debuted). The algorithm makes [victorious] wildcard moves which no human player would even contemplate [stupid moves become masterplays].

When 9D-master Sudol attributes human qualities to the beauty of his AI opponent's creativity upon formulating certain moves... is definitely eye-opening. Hubris replaced.

[dragontamer]

Nah.

I think with the rise of KataGo, its becoming clear that AlphaGo's "dumbfounded" strategies were instead incredibly strong tactical play with hilarious levels of blindness with regards to ladders.

It feels like modern AIs (like KataGo, which are hundreds of times better than AlphaGo) are getting closer to what humans consider appropriate strategic play.

Go players must be humble because if the opponent is stronger then the opponent wins. But Go AI Programmers don't necessarily have to be. Go AI programmers look for the weaknesses, lean upon them and yes, prove that AlphaGo/AlphaZero never learned ladders. Ever. Or other such concepts of strongly forcing moves (ex: loose ladders).

That's one of KataGo's biggest innovations. Explicitly programming a ladder solver so that simple ladders can indeed be factored in by the neural net.

----------

I do wonder how Lee Sodul will react if we told him that the superhuman AI he played could not see ladders... and had other such key weaknesses.

[jhbadger]

>I do wonder how Lee Sodul will react if we told him that the superhuman AI he played could not see ladders... and had other such key weaknesses.

I wonder what Kasparov thinks of the fact that he was beaten by such a primitive chess program as Deep Blue's!

[dragontamer]

Ladders are a thing so easy that its literally in this beginner series of tutorials. Its a thing that 30kyu beginners learn and master.

Its very strange to me that AlphaGo / AlphaZero was unable to ever learn ladders. It shows that the way humans learn and AIs learn is quite different (or at least, MCTS + Neural Net machines learn differently).

Yes, Go is a strategic game of patterns and perhaps we humans overemphasized the ladder. Nonetheless, its a concept that humans can see and calculate with reasonable speeds that the (earlier) AI was unable to do (and now we've built stronger AIs that can prove this weakness).

---------

This 100% makes the 2016-era discussions about the "strategic brilliance" of AlphaGo come up into question. Now that KataGo is superior (ie: MCTS+NeuralNets + dedicated Ladder Solver), we need to double-check all those "strategically brilliant" moves with the newer AI and see if a ladder messed with them.

Etc. etc.

The things AlphaGo sees aren't necessarily useful to us humans, nor are they useful to modern AI-levels of Go. They're just... that. Trapped in 2016. There should not be any great mystery assigned to the 2016 era game, aside from it being a pivotal moment for AIs.

The game itself is now suspect, now that we know all of AlphaGo's flaws. As Go players, we have better things to study.

[Zacharias030]

Do you know of a systematic study of the AG - Lee Sedol match with contemporary KataGo? I’ve never seen this written up.

[ProllyInfamous]

>we need to double-check all those "strategically brilliant" moves with the newer AI

Thanks for this detailed response/thought. As a layperson, cannot wait for the documentary on KataGo (AlphaGo II).

[dragontamer]

Probably not going to happen.

DeepBlue was the big chess moment. No one writes or cares about Stockfish vs DeepBlue.

Well, chess players performing opening analysis for their games care. Go players care about KataGo being more readily available and stronger than AlphaGo or AlphaZero. But it's back into the weeds and grind, it's not too interesting a story outside the Chess or Go world.