8 years later: A world Go champion's reflections on AlphaGo

https://blog.google/around-the-globe/google-asia/8-years-later-a-world-go-champions-reflections-on-alphago/

by rayshan • 2 years ago

79 75 2 years ago

    I thought it would be an easy victory

    I ... ended up only winning one out of our five games

It's interesting, how an expert in a field can be unaware of how AI is taking over. And a few years later, no human can compete anymore.

I think we are in a similar situation in multiple professions today. For example with self-driving.

Musk recently said, that other car manufacturers are not much interested in talks about licensing FSD because they don't think it can work.

In ten years, probably no human can compete with AI drivers anymore.

svachalek• 2 years ago

In ten years, probably no human can compete with AI drivers anymore.

That's what they said 10 years ago. Sooner or later people will say it and be right, but the last few percent of any problem is a lot harder than people give it credit for. It may not be that hard to stay in a lane or write a little code, and that may look like it's doing most of the job, but those common tasks are just the easy part.

ArtTimeInvestor4• 2 years ago

10 years ago, a lot of FSD was still manually written code. Manually written code gets harder and harder to improve, the larger the codebase.

Now it is all NNs and therefore will scale with more data and more compute. Which is increasing exponentially. So far it seems like they are not hitting diminishing returns.

huytersd2• 2 years ago

I think Waymo does better than a vast percentage of human drivers. It’s here already.

bevekspldnw5• 2 years ago

“Driving” is solved. Driving with humans on the road - doing unpredictable human things - is far off still.

My guess is industrial and home robotics will solve a lot of the “doing things around humans” problems in the next ten years.

Why the hell people decided to automate giant death machines before perfecting small things never made sense to me.

bongodongobob• 2 years ago

You do realize this was 8 years ago, and no Go engine came even close to what Alpha Go was able to do right? Afaik, there weren't even any competitive engines period. It basically came out of nowhere.

ArtTimeInvestor4• 2 years ago

That's what I meant.

That's how it usually goes with technological progress.

In any field.

Progress is minimal for a few years and then suddenly jumps up very suddenly.

So to predict what's coming, you can't just extrapolate the progress of recent years. You have to account for it being exponential with a very uneven distribution of sudden jumps.

vintermann• 2 years ago

Not entirely right. Remi Coulom's Monte Carlo Tree Search, in 2006, was the first really big discovery. It didn't make engines good enough to beat the best humans, but it steadily made them good enough to beat 99% of go playing humans, playing at up to 6 - 7 dan level. It was still part of AlphaGo, too (though as I recall AlphaGo Zero did away with it).

lawn2• 2 years ago

It's interesting, how an expert in a field can be unaware of how AI is taking over.

I think the interesting thing is how an expert in a field is wholly unprepared for predicting how the future will develop.

You mention what Musk has said about FSD and how it will completely take over in just ten years, but I feel compelled to point out that Musk has said that it's just right around the corner with only small challenges left, for many years.

I wouldn't place any faith in anything Musk says.

nemothekid• 2 years ago

Musk wasn't wrong - he just wasn't the person to deliver it. Waymo, AFAICT works astonishingly in San Francisco. I think you can argue that it might not work in the snow, but that's pretty much it

jfengel• 2 years ago

At the time, I think everybody was unaware of this. Everybody followed the development of machine chess, but it was widely assumed that machine go was an entirely different category of difficulty. Chess engines gradually encroached on the very top grandmasters. AlphaGo came out of nowhere.

hyperpape• 2 years ago

At the time, the only public results were demonstration games against a much weaker professional. The actual strength of the machine was only known privately within DeepMind.

SkyMarshal• 2 years ago

It's interesting, how an expert in a field can be unaware of how AI is taking over. And a few years later, no human can compete anymore.

Easy to see that in hindsight, but when the game was actually played it was earlier in the development of AI and less apparent how good it had become.

dvh• 2 years ago

FSD will never work because it concentrates defendants into single juicy target.

When your neighbor Bob (who is still paying mortgage and his wife is battling cancer and who occasionally babysit your kids) ran over your cat, you don't sue him. But you would sue Tesla.

Mark my words, in 10 years ex-programmers will be throwing shelter cats under FSD cars just to earn a living.

xcv1232• 2 years ago

It will have 100 cameras capturing surround footage of ex-coders throwing cats at the vehicle. Case dismissed.

tromp4• 2 years ago

Go is a deeply complex strategic game — famously far more complicated than chess, with 1,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000,000 possible board configurations.

The correct number of legal Go positions is over twice as much, or to be exact [1]:

208168199381979984699478633344862770286522453884530548425639456820927419612738015378525648451698519643907259916015628128546089888314427129715319317557736620397247064840935

Indeed far larger than the ~ 4.8 x 10^44 legal chess positions [2], that is in between the number of legal 9x9 and 10x10 Go positions.

[1] https://tromp.github.io/go/legal.html

[2] https://tromp.github.io/chess/chess.htm

HackerThemAll• 2 years ago

All these digits are only making it more obfuscated. Using the order of magnitude it's 10^44 for chess versus 10^170 for Go. Thus, Go is 10^126 times more complex than chess.

For reference, the estimated number of individual atoms in the universe is thought to be between mere 10^80 and 10^83.

tromp4• 2 years ago

Or one could say that Go is about as complex as 4 simultaneous games of go, with the number of positions being about the 4th power.

kibwen• 2 years ago

All these digits are only making it more obfuscated.

I think that using these numbers as as stand-ins for difficulty is itself a form of obfuscation.

The truth is that, despite the massive number of potential board states, Chess and Go are some of the easier games to solve, thanks to their nature (perfect information, zero randomness, alternating turns where each player plays exactly one move). And trying to use board states as a proxy for complexity and complexity as a proxy for difficulty doesn't generalize to other categories of games. Compared to Go, what's the complexity of Sid Meier's Civilization? If I devise a game of Candyland with 10^180 squares, is that harder to devise an optimal strategy for than Go just because it has more board states?

The reason that we're still using board states as a proxy for difficulty is because historically our metric of "this is difficult for a computer to play" was based on the size of the decision tree and thus the feasibility of locally searching it up to a given depth. In the age of machine learning, surely we can come up with a more interesting metric?

banannaise• 2 years ago

Number of possible game states is a poor measure of complexity. How many game states does soccer or basketball have, when you consider flight of ball and movement of players? Does that tell us anything about whether basketball is more complex than Go?

frutiger• 2 years ago

For reference, the estimated number of individual atoms in the universe is thought to be between mere 10^80 and 10^83.

Yes, but what are the estimated number of states of all these atoms?

lp4vn• 2 years ago

Go is played in a a bigger board though and has this kind of recursive nature where a subset of a go game is also a go game while chess is more ad-hoc.

bluecalm3• 2 years ago

Complexity of the game has nothing to do with the number of legal positions. It's very easy to design a game with arbitrary number of positions which is very simple. While go might be more complex than chess using a more reasonable measure this argument was used to for arguing nonsense in scientific papers in the past (that some poker games are more complicated than chess because they have more possible states).

falserum• 2 years ago

I have a further nitpick regarding terminology.

Wording like “game is more complex” overal seems incorrect. Game is not complex by itself (for example go rules are extreemly simple), all the difficulty and challenge depends on the skill of your oponent. Game only allows the opponent to demonstrate the skill.

Kranar3• 2 years ago

Are you sure that the complexity of a game has absolutely nothing to do with the number of legal positions?

I mean I am open to hear the justification for this, but I was fairly certain that all measures of game complexity are a function of the number of legal positions. Now certainly there are other factors, namely the cost of computing the transition from one legal move to another legal move so a simple game might have a very low cost transition function while a complex game has a very complex transition function, but I can't conceive of a game where the number of legal positions bears no weight on the game's complexity.

anon84873628• 2 years ago

For extra fun, use the "Listen to article" feature on that paragraph...

Pet_Ant• 2 years ago

I know there are Go channels in Korea for watching professional matches... are there any for watching AIs face off against each other?

apetresc• 2 years ago

Virtually every public Go server at the moment, unfortunately.

mNovak• 2 years ago

Michael Redmond, to my knowledge the highest ranked English-speaking Go player, does youtube analyses of Go matches, including Pro vs AI games, which are very insightful.

rhaps0dy• 2 years ago

Great PR for Google from Lee! It totally isn't mostly for advancing Google's commercial interests, the bottom line being:

"I believe that humans can partner with AI and make great progress. As long as we can set clear principles and standards for it, I am quite optimistic about the future of AI technology in our daily lives."

I hope he got paid well.

bnprks• 2 years ago

To my knowledge AlphaGo models never became meaningfully available to the public, but 8 years later the KataGo project has open source, superhuman Go AI models freely available and under ongoing development [1]. The open source projects that developed in the wake of AlphaGo and AlphaZero are a huge success story in my mind.

I haven't played Go in a while, but I'm kind of excited to try going back to use the KataGo-based analysis/training tools that exist now.

[1] https://github.com/lightvector/KataGo

alas44• 2 years ago

Google's documentary on AlphaGo https://www.youtube.com/watch?v=WXuK6gekU1Y

Truly a must watch! (just look at the video comments to be convinced)

RetroTechie• 2 years ago

What stuck with me is Lee Sedol's strong emotional reaction, leading him to leave professional Go playing.

It's understandable he didn't expect AlphaGo to be that strong. Or that (for him) losing to a machine took the 'soul' out of the game.

But come on... I've been cornered by Pac-Man ghosts many times. That doesn't make Pac-Man less fun to play.

Nor does losing to the crude 'AI' steering those ghosts. Instead, you play, aim for a high score, see how long you can survive, how many levels you can complete, or how many fruits & ghosts you can eat in a game.

And (if you care) compare how those 'metrics' stack up against other players.

If a machine with superhuman Go-playing ability isn't fun or challenging, then stick to human opponents.

Of course it's his views and choices, and I respect that. But other than providing extremely challenging opponent, I don't see how human-beating machine would take the fun out of a game. Rather the opposite: new tactics, new insights, a raised upper bound for a Go player's strength (human or otherwise), etc.

QuantumGood• 2 years ago

Some people connect their ego tightly to their accomplishments. Others are emotionally fragile to what they perceive as negative experiences. Some achieve great things by what they perceive as great sacrifices, and have a lot of difficulty maintaining what they have achieved. Some at or near the top of a field perceive their accomplishment primarily as being "best", and are consumed with angst when they are no longer best. The public explanation they give may not describe how they are different than someone trying to understand.

Open up how you frame personality types and life experiences, and you can think of possibilities beyond "I don't see how".

tasuki2• 2 years ago

He wanted to advance our knowledge of the game, to find better moves. Now that he no longer can, why pay?

efrank3• 2 years ago

I can't believe it's been 8 years.

8 years later: A world Go champion's reflections on AlphaGo

Related Stories