On AI vs human thinking

[Bill Spight]

The benefit of using human games as a basis is that we either have or can get a human explanation we have a chance of understanding, however imperfectly. AI has just widened the pool (by pissing in it, in my view )

Nectar of the Gods!

Let us not sell humans short. We live in the era of Big Data, where programs sift through masses of data and come up with Black Box answers to human questions. (The answers are not necessarily sound, OC.) Not too long ago I was reading about some academic researchers studying the answers of one of these Black Box algorithms and coming up with a simple algorithm using only three parameters that got the same answers to within a small margin of error. Now, nobody had come up with such good answers before the Black Box algorithm, but, once we had those answers, humans were able to come up with good concepts to explain those answers simply. This is not the same thing as learning by imitation, but is a more scientific approach. The top go bots are Black Box algorithms, and I am optimistic that by studying their play humans will come up with new go ideas to help us understand the game and play it better.

[hyperpape]

But this is not unique to AI. We do it all the time with human moves (2nd-line probes against a shimari were rare until little more tan a century ago, 5-5 was new, 10-10 was new, josekis go in and out of fashion...). The benefit of using human games as a basis is that we either have or can get a human explanation we have a chance of understanding, however imperfectly. AI has just widened the pool (by pissing in it, in my view )

One thing that AI has is shock and awe. In 2015, it was a just promising little curiousity, and just two years later it said "I am the teacher now". My sense is that no other shift in the game happened so suddenly, or so unexpectedly. As you've pointed out, Shin Fuseki was a more gradual development than many of us think, and more balanced--the conservatives could maintain with a straight face that their approach would win out in due time.

Edit: fixed completely broken syntax of one sentence

[hyperpape]

Please note my previous statement that human explanations about their decisions, particularly about intuitive decisions, are typically post hoc. This is clear in experiments involving hypnosis and split brain research.

Apparently the old split brain research has recently been called into doubt (though don't ask me any details, I just remembered reading a tweet about this when I saw your post). https://www.ncbi.nlm.nih.gov/pubmed/28122878. The broader point about post-hoc justifications hasn't been repealed, as far as I'm aware

[John Fairbairn]

One thing that AI has is shock and awe. In 2015, it was a promising little curiosity to saying "I am the teacher now". My sense is that no other shift in the game happened so suddenly, or so unexpectedly

This is quite true. There is also the aspect that AI go comes with a shiny new toy that some people think they need to have urgently so their batteries can burst into flames before anyone else's.

But this is precisely what turns AI go into cargo-cult status.

I am optimistic that by studying their play humans will come up with new go ideas to help us understand the game and play it better.

I too expect this to happen. They key phrase is "humans will come up with..." But I also expect that if humans studied, say, Dosaku with the same intensity and passion they study AI games they would likewise play better. After all, that is essentially why present pros are better than present amateurs.

I am not knocking AI. I am knocking AI worship.

[hyperpape]

This is quite true. There is also the aspect that AI go comes with a shiny new toy that some people think they need to have urgently so their batteries can burst into flames before anyone else's.

But this is precisely what turns AI go into cargo-cult status.

I think the proof of a cargo cult is in the results, or lack thereof, so time will tell.

[Bill Spight]

I am not knocking AI. I am knocking AI worship.

OK.

I am optimistic that by studying their play humans will come up with new go ideas to help us understand the game and play it better.

I too expect this to happen. They key phrase is "humans will come up with..." But I also expect that if humans studied, say, Dosaku with the same intensity and passion they study AI games they would likewise play better. After all, that is essentially why present pros are better than present amateurs.

Well, Dosaku has already been well studied for a long time. I doubt if studying him now will produce any new go ideas. OC, you never know. Was AlphaGo Master better than Dosaku? I don't know. Maybe Dosaku could give current top pros three stones. But I rather expect that AlphaGo Zero was better than any human who ever lived. And it defeated humans who had absorbed ideas from Dosaku, Go Seigen, and other giants of the game. Now, you and I have recognized echoes of Go Seigen in today's top bots, but the AlphaGo Zero self play games have opened up new territory for exploration. They have an alien feel to them. The dominance of the bots seems to lie much more in the realm of strategy than in tactics. To me that indicates that there are general concepts in their play that are waiting to be discovered and articulated by humans. Humans are good at doing that, so we will.

[Kirby]

One thing that AI has is shock and awe. In 2015, it was a just promising little curiousity, and just two years later it said "I am the teacher now". My sense is that no other shift in the game happened so suddenly, or so unexpectedly. As you've pointed out, Shin Fuseki was a more gradual development than many of us think, and more balanced--the conservatives could maintain with a straight face that their approach would win out in due time.

Even outside of the domain of go, I'm a believer that things are going to get crazy in our lifetime.

Think about it. Humans have been around for thousands of years, but electricity was just being investigated a few hundred years ago. In my grandfather's lifetime, he had been using a kerosene oil lamp to study. The first airplane was invented just over a hundred years ago, and since then, we've traveled to the moon. They're discovering organic material on Mars, now. The world wide web was invented when - like the 90s? That's just a couple of decades ago.

The rate of advancement in technology isn't happening linearly. It's starting to lift off like crazy.

Here's a funny article along the same lines:
https://waitbutwhy.com/2015/01/artifici ... ion-1.html

Will people be living on Mars in my lifetime? Will AI take over the world? Will there be some new invention that allows people to teleport across the world?

I can only imagine.

[Gomoto]

AI is the new kid on the block. The old blokes don't like the new pretender. But the early adoptors adopt. I for one welcome our new Go overloards.

He, books were dissed too. Which invention wasn't. And I only find the go conservativists mildly amusing, but I prefer to discuss these issues over the go board

I worship the AI because it opens up a fine brave new world!

[Calvin Clark]

When the AlphaGo games came out, there was a lot of amazement and confusion. There was also speculation on what could be learned from the games. I think John noted in another thread that O Meien expressed some skepticism about how much humans can learn from AI.

But things change quickly, and there's a huge difference between just having a handful of self-play records to marvel over and actually being able to play against the AI. The modern AIs can provide evaluation and variations, which is helpful, too. They are an enormous resource, especially for the top players who don't have humans much stronger than themselves to guide them.

But it is still, IMO, highly likely that AI is not close to the truth yet. AlphaGo was still improving when DeepMind reprovisioned it to other goals. Leela Zero is not stuck.

But there are some concerns. For much of May, Leela Zero self-play games mostly looked like this. Some huge percentage of its games started with this exact sequence. I could actually go further, you get the point:



It had all but stopped playing anything but nirensei as black and 4-4 + 3-4 as white. I hear they are tweaking it to explore more, but what if this is the truth? Will human play stagnate from just blindly copying a narrow set of options from impossibly distant intelligences, vast and cool and unsympathetic?

I am more of an optimist on this. I think what we are seeing is a local maximum in a kind of steepest ascent search, akin to being proud that you have reached the top of a hill while simultaneously being oblivious to Everest looming behind your back. Newer AIs, built in different ways, may not find the Everest for a long time, but maybe they will find a Fuji or Rainier and at least we will have some different scenery.

I also look forward to explainable AI in go as well as in other fields.

[yakcyll]

Hi.

This isn't much, but I wanted to share my perspective too, since the degree to which the new bots are glorified, sometimes personified, irks me greatly (I'm aware it's not healthy, patience will be greatly appreciated).
Imagine a kid shows up - a Westerner, no less - fifteen years of age, well kept, calm and silent, his face betrays no emotion. He starts playing the game, climbing ranks with speed previously unthinkable of humans, first beating top amateurs, then winning tournaments, obtaining a professional rank and proceeding to dominate the world scene in a manner leaving the entirety of Asia scratching their heads. It becomes apparent really quickly, that the kid is a one of a kind prodigy with abilities many consider divine. Remembers every kifu he ever laid his eyes on, can clearly analyze all situations fifty, a hundred moves ahead, plays moves other professionals don't even consider as viable, but they somehow work for him every time. For the next sixty or seventy years he dominates the game, spawning masses of fans and followers who try to decipher his moves through personal analysis as he can offer little to no insight that can be immediately understood and used to adapt the current game theory. No one is able to come close to beating him consistently, as he can adapt to people catching on to his tactics almost instantly.

Would we study his games? Hell yes, we would, I mean, who wouldn't? He's the best, there must be treasure troves of knowledge about the game in his games. He's forming new patterns, presenting rebuttals to attacks previously thought of as clearly advantageous to the attacker and successfully forcing his way into situations thought of as impenetrable. But at some point questions will arise. How much more is there to learn? Aren't we approaching a point where we won't be able to discern what's the right and what's the wrong move on our own? What's the point of competing against this guy if there's no way apparent to touch him? Isn't he hurting the professional scene by claiming every title for himself? Is there no place for originality and character in the game anymore, when everybody and their pets are imitating him?

AlphaGo is this kid. There are a couple of differences of course, namely that it won't leave us, it can be cloned, it doesn't need another person to play and that it is a computer program. I'd argue against claiming it has a distinct style or possesses intuition. What we call the two in regards to humans are just by-products of thorough analysis of game records leading to forming decision trees with winning probabilities, calculated through some sort of value functions, exceeding those attached to paths taken by humans. What we call intuition is a heuristic applied by humans to glue the known with the unknown without using prohibitively high amounts of resources. Computers don't need this, they can just see what a particular path ends with for themselves.

I'm running the risk of sounding like Debbie Downer, but seeing so many people glorifying bots as if they were oracles contributing to the ultimate human enlightenment smells like cargo cult to me too. They're making us reconsider what we know about the game and helping us see the previously unseen, but that was to be expected. We'll be going through similar discussions and problems that the chess community already has a lot of experience with. I'd love to have them solved in Go too already, to avoid seeing laptops running conclusive analysis of live games everywhere or having to argue with other players how a move is not necessarily correct in a given situation just because a bot played something like it. It's fairly clear that machines are (or show potential to eventually become) superior to humans when it comes to problems with bounded solution spaces, no need to beat a dead horse.

[gowan]

I also dislike talk of programs as if they were people. I don't find the fact that the best programs can beat the best humans at go to be a matter of awe and shock. Fifty years ago the Japanese pros ruled the go world and, among Western amateurs at least, the Japanese go was considered the most important to study and emulate. Thirty years ago or so when the Chinese teams were defeating the Japanese teams in the Supergo matches there was an indication that the pros and go styles we admired was not the end and that something different could be good, and players flocked to study the new ideas. Then came the Korean wave, playing seemingly superior go. Now, among human players the Chinese seem to be at the top and computers driven by programs are defeating everyone. I believe these shifts will continue and that AI play sufficiently different will appear. There is a vast amount of undiscovered go still waiting to be found. From my perspective, the AI "players" don't understand go at all. They resemble human club players who never study or analyze their games but, through just playing a lot, reach a dan level. The "Zero" type AIs do that, playing millions and millions of games but still not "studying" or understanding theory. That's why these systems can't explain what they do. What we get is such and such a move has a better chance of leading to a win than some other move does. The AI programs are just scratching the surface, not getting deep in. The new moves the AI programs discover are fun to play with but the pros aren't arbitrarily using them, they are working hard analyzing them so they understand why they work.

[hyperpape]

I can only imagine.[/quote]

Think about it. Humans have been around for thousands of years, but electricity was just being investigated a few hundred years ago. In my grandfather's lifetime, he had been using a kerosene oil lamp to study. The first airplane was invented just over a hundred years ago, and since then, we've traveled to the moon. They're discovering organic material on Mars, now. The world wide web was invented when - like the 90s? That's just a couple of decades ago.

The rate of advancement in technology isn't happening linearly. It's starting to lift off like crazy.

Here's a funny article along the same lines:
https://waitbutwhy.com/2015/01/artifici ... ion-1.html

I saw a contrarian take recently. It said something to the effect that if you look back, technological progress might be having less of an impact. Take 1865 to 1940 (I can’t remember what years the article used): you get cars, planes, radios, phones. These technologies change some of the most fundamental things about how life is lived. Can smartphones really compete? A pessimist would say that rocketry was an area of rapid progress by 1940, so even the moon landing is impressive, but still just improving existing ideas.

I don’t know what I think, but there was something compelling about the claim.

[Bill Spight]

The rate of advancement in technology isn't happening linearly. It's starting to lift off like crazy.

Here's a funny article along the same lines:
https://waitbutwhy.com/2015/01/artifici ... ion-1.html

Edit: Corrected author of quote.
Epigraph of the article:

We are on the edge of change comparable to the rise of human life on Earth.

Which made me think of this quote:

Wir wissen, daß wir Vorläufige sind
Und nach uns wird kommen: nichts Nennenswertes.

"We know that we are forerunners
And after us will come: nothing worth mentioning."

[Tryss]

They resemble human club players who never study or analyze their games but, through just playing a lot, reach a dan level. The "Zero" type AIs do that, playing millions and millions of games but still not "studying" or understanding theory

Is there really a theory to understand? In my opinion, go theory is "just" a crutch that allow humans to play better than with just pure reading and intuition. It is a (human) way to create heuristics to prune the game tree.

And while theory is indeed very useful for humans, is it necessarily that useful for a program?

[RobertJasiek]

There are different kinds of go theory, such as
- informal, traditional go theory from guesswork to supported by empirical evidence,
- informal go theory supported by systematic study but without mathematical proof,
- formal go theory of theorems aka truths proven mathematically,
- hybrids of systematic study awaiting allegedly possible mathematical proofs or requiring working out details so that some variation of study can be proven,
- hybrids involving computers from AI guesswork (of which AlphaGo Zero is successful empirically) to algorithms being part of mathematical proving,
- etc.
Only a few people research in theorems but do not overlook them!