Hacker News Reader

Comment by karmakaze | Hacker News Reader

I've spent so many years playing and hearing about how Go will not be solved in our lifetime that the day AlphaGo won 4/5 games v Lee Sedol marked my personal timeline into what came before and after. I walked around all day in a daze watching people as they go about their daily lives as if nothing had happened.

I've heard it said that man landing on the moon was like that for them, but I didn't understand as it was the only world I knew.

Now I can appreciate that these were the firsts of many singularities yet to come in AI and space exploration and I hope to live to witness a few more (but not too many).

View on HN

I have an opposite view: it's not that shocking that AI has advanced a lot. It's a lot more shocking to learn that humans aren't as great as we hope to be. Also, our prediction sucks. I mean, who said that go is such a difficult problem that it would take a lifetime to solve? Sounds like intellectual arrogance to me. Sure, the problem space is huge, but it's well-defined and homogeneous. There was a time that reciting a long text or multiplying large numbers is considered a humanly intelligent thing, only that it wasn't. Alan Turing used to think that AI is good to humans because it teaches us to be humble, and I think we're kind of getting there (for certain domains). On the other hand, things like self-driving will remain unsolvable because the problem is fundamentally ill-defined; we don't even know what is a good driving.

(Edit) To those who think self-driving is a well-defined problem: it can be in some remote areas, but imagine driving in bustling city streets with kids, bicycles and dogs. The driving problem becomes a communication problem.

> On the other hand, things like self-driving will remain unsolvable because the problem is fundamentally ill-defined; we don't even know what is a good driving.

Humans arrive safe and unhurt (as much as possible, especially while human drivers remain on the road) at their destination with minimal violation of the locality's established rules of the road. No?

(Though now that I've written what amounts to a utility function, I fear what sort of paperclips may come out of it.)

> minimal violation of the locality's established rules of the road

I wish. I've come to the conclusion that the only true rule of the road is: don't crash. As long as no actual collisions occur, people are totally fine with doing whatever they want and bending the rules for their own convenience. I can no longer predict the behavior of other drivers. Even something as basic as the turn signal is unreliable since people are forgetful.

I think you are right. In fact, thinking some more, it would seem that the whole point of the ‘rules of the road’ is simply to prevent crashes - we have just built up a whole load of protocols in order to achieve that

Drivers (both AI and human) may face problems that are essentially ethical trolley problems. While many of these choices are clearly artificial to the point of ridiculousness, the one that gets me best is "should a self-driving car drive itself off a cliff, killing its only passenger, or hit and kill some >1 number of pedestrians?". While the external observation may be "minimising deaths is preferable, so drive off that cliff", are people willing to use a vehicle that might intentionally kill them as an intrinsic part of its operation? Or will market forces result in self-driving cars that make more selfish choices being more popular, potentially resulting in suboptimal prisoners-dilemma style results?

There are also different ethical norms in different cultures about preferences (https://www.wired.com/story/trolley-problem-teach-self-drivi...). While these are edge cases, they're the edge cases people are worried about, and the source of the ill-definedness: "unhurt as much as possible" implicitly chooses some ethical tradeoff that people can easily have different answers to.

Also such meek and suicidal cars would get abused to no end. Just imagine all the assholes today that pass cars in turns with bad visibility or bike in crazy ways. Today they are still paying some attention, because they may easily get killed if the other drivers don't notice quickly enough what they are doing. With meek AIs on the road you can do anything (as long as you bunch up in large enough groups).

Self driving is much simpler than playing Go. The hard part is getting the sensors working properly, in all conditions, even when the car is twenty years old and dented.

I have been working on AI/ML since 2007, and this is what I had read too (repeatedly). That Go is incredibly hard to conquer for a computer, and it will be a long while. So when this actually happened, I was shocked/surprised.

I’m glad I’m not the only one. I was totally floored, and I struggled to explain the significance to non-go players.

My respect for AI increased drastically that day, and (honestly) I developed a small amount of fear due to how AlphaGo’s style of play was not understood particularly well (e.g., some of the moves would absolutely be called “slack” if played by a human).

You both speak of the day AlphaGo won 4/5 matches, yet the matches were played over a series of days. Which of those days was the switch flipped, then? For me there was a significance in day three, but it was mitigated somewhat by the (to me) surprise of day/match four.

IIRC, it was day 3 for me as well, and I had the same (minor) let down on day 4 (resign?!?!). That said, I imagined the resignation was a fixable flaw in the AI, and this turned out to be correct.

I can’t actually remember where I first learned about the match. It may have been HN, it may have been in an AGA e-mail, or it may have been some tech-oriented magazine/web site in English or Japanese. I am certain it wasn’t match 1, because I reviewed earlier matches, and I remember the let down of match 4, so must have been match 2 or 3.

For me, day 3 was the stunner even though the writing was already on the wall.

Day 1 was a great surprise but I was still left wondering if it was a fluke. Day 2 showed that it was no fluke and I started to get a sinking feeling. I guessed that Lee Sedol would lose the third game and win the fourth after the pressure of the Korea and humanity was off him.

Go still isn't solved (neither is chess), we just have a machine good at knowing which parts of the search space are worth checking.

I think achieving superiority over humans is practically solving the problem though. Solving chess or go by going through a complete search space seems more like a hardware/computational goal than a practical ml/ai goal.

You don't have to go through the complete search space if it turns out optimal strategies are sparse. What do I mean by that? Take a second-price auction: the dominant strategy here is to always bid your true value. Meanwhile, the search space for this would be any real number in between 0 and your true value. What does this mean for computational games like Chess or Go? It may mean while the search space is exponential, there may exist computationally trivial strategies that work. I would compare this to Kolmogorov complexity, except instead of having a program as your output, it's a strategy.

It all hinges on your definition of "solved".

"Solved" in the AI/game theory has a very strict definition. It indicates that you have formally proven that one of the players can guarantee an outcome from the very beginning of the game.

The less-strict definition being thrown around here in the comments is more like "This AI can always beat this human because it is much stronger."

I think most people discussing this mean the later, less pedantic option. I mean, that’s the spirit of AI. Can we make it think like a human, or even more so. We are the yardstick.

That is a silly mis-use of the term and that is not being pedantic. A problem isn't solved just because you beat the existing solution (i.e. human players). As long as there is the potential for a better solution that can beat your solution there is work to be done.

That's like when the whole class fails a test, but the prof. grades on a curve. Someone gets an A, but not really. edit: some grammar.

Any substandard statistical model fitted to by a simple computer program is superior to what an unaided human could achieve with pen and paper, but few of them can claim to practically "solve" the problem because they are better than crude fit heuristics proposed by humans who are not good calculating machines.

An algorithm can't claim to have "solved" Go, when future versions of the algorithm are expected to achieve vastly superior results, never mind any formal mathematical proof of optimality. What it has demonstrated is that humans aren't very good at Go. Given that Go involves estimating Nash equilibrium responses in a perfect information game with a finite, knowable but extremely large range of possible outcomes, it's perhaps not surprising that Go is the sort of problem that humans are not very good at trying to solve and that computers can incrementally improve on our attempted solutions. Perhaps the more interesting finding from AlphaGoZero is that humans were so bad at Go that not training on human games and theory actually improved performance.

We've just created a tool people can use to play Go better than a person without the tool. Until something emerges from the rain forest, or comes down from space that can throw down and win I'd say Humans are still the best Go players in the known universe.

This is the performance vs understanding dialectic. A bunch of humans built a machine that is superior at chess, but that machine can't teach humans what it knows.

Reminds me how sometimes geniuses cannot translate the way their mind works to non-geniuses. It's such an implicit talent that it's not even .. "reified" in their upper brain. It all happens in cache.

Chess computers teach by sparring rather than lecture. Humans still learn.

Alpha *, at least, also learns by sparring, so there is a nice symmetry there.

Humans can and are absorbing some of what the machine has demonstrated.

I think Kasparov has it right when he says that the best player isn't a human or a machine, but a human using a machine as a tool. The machine can help the player optimize and reduce mistakes, but machines don't yet know how to ask questions and explore in the same way. Maybe they never will.

There's a name for this approach, they call human-AI teams "centaurs". It's a fascinating concept. I am deeply curious if eventually that will be outstripped by pure AI too. I believe so.

Optimal strategy for human-AI teams has been really close "defer to the computer for every move" for a while now. They're only interesting because they're adversarial.

"Solved", in this case, means "computers can play the game at levels no human can beat."

That's not the normal meaning of solved in regards to game theory.

I believe with respect to game theory, solving a game like Go would require finding a strategy that obeys the one shot deviation principle. The result would rather be boring to watch however, because the conclusion of every game played under this strategy would either be draw, or based on which player starts off first.

[1] https://en.wikipedia.org/wiki/One-shot_deviation_principle

But it is what "solved" means in deep learning. Common terms very often acquire different technical meanings in different fields. Or even technical terms! Whether a single hydrogen atom is a "molecule" depends on whether you're talking to a physicist or a chemist. And "functor" means something very different in Java programming patterns than it does in math.

I have to concur with the other poster here. You're not using the term in the usual way. Traditionally a game is called solved when its search space has been searched exhaustively or there is some other analytic solution that allows you to determine who wins and who looses in every constellation. Tic-tac-toe is solved, for instance.

https://en.wikipedia.org/wiki/Solved_game

Regardless, we need a term for “computers can play the game at levels no human can beat.”

Well, don't use one that has an existing meaning in that exact context.

Superhuman

For me the two moments so far have been seing computeurs winning at Go and knowing that basic quantum computeur are out there (even if they are not useful yet).

Reply on news.ycombinator.com