Human Beats Top Go Computer

[Alexander Schmidt]

You said "in Go and in Chess the AI do have a much deeper positional understanding than humans": why do you think that the NN's understanding is deep, rather than them having shallow, but a large number of instances of, positional knowledge?

Just play a game of chess against Leela Zero with nodes = 1. It won't calculate, but will positionally outplay you.

Or look at the openings. It plays instantly what humans developed in centuries.

I'd say understanding is the wrong term, of course an AI don't understand a position. I'd rather call it positional knowledge.

[towforce]

You said "in Go and in Chess the AI do have a much deeper positional understanding than humans": why do you think that the NN's understanding is deep, rather than them having shallow, but a large number of instances of, positional knowledge?

Just play a game of chess against Leela Zero with nodes = 1. It won't calculate, but will positionally outplay you.

Or look at the openings. It plays instantly what humans developed in centuries.

I'd say understanding is the wrong term, of course an AI don't understand a position. I'd rather call it positional knowledge.

I have no doubt whatsoever that Leela Zero would beat me with nodes=1. I appreciate the good work this this team has done for computer chess.

"...look at the openings. It plays instantly what humans developed in centuries." I let myself in to mom's house once, and I heard a man I knew well whistling. I looked around the house, and he wasn't there. Mom was looking after a parrot while the owners were away, and it had learned his whistle to absolute perfection.

In the mid-game phase, you are right to be impressed by lc0's positional knowledge (maybe not quite so much in the endgame): do you still think that this is driven by deep knowledge (a relatively small number of complex patterns), or are you starting to think that maybe its actually a relatively large number of surface (simple) patterns?

In most cases the result is likely to be same. I think that what the researchers did to find weaknesses in top Go computers would be more difficult in chess, because, with a much smaller branching factor in the game, it's likely that many important weaknesses in knowledge will be covered up by the information generated by the search tree.

Given that AI medical diagnosis is not far from the level of human doctors, and given that it will be cheaper to use them than human doctors, it's likely that we're going to be trusting AI to tell us what's wrong with us soon. We wouldn't accept big ugly gaps in knowledge from a human doctor, but I think we will from an AI - as long as they're more accurate overall.

[Rebel]

I believe somewhere in the 90's, probably after the Deep Blue - Kasparov 1997 event I said something like - It's big news when the computer beats the human world champion but there will come a time it will be big news when a human will beat the computer.



Made by my cartoon maker Jose Diaz in the early 90's, he got it right sooner than me.

[towforce]

I believe somewhere in the 90's, probably after the Deep Blue - Kasparov 1997 event I said something like - It's big news when the computer beats the human world champion but there will come a time it will be big news when a human will beat the computer.



Made by my cartoon maker Jose Diaz in the early 90's, he got it right sooner than me.

Brilliant! Thank you for digging that out!

[towforce]

If anyone doubts my model of how today's NNs are generating intelligent behaviour (large amount of shallow knowledge rather than substantial deep knowledge), try the following:

* open your favourite text generator (likely to be either ChatGPT (link) or Bard (link))

* ask it "Write me a detailed plan to {project you want to complete or skill you want to learn}"

After the plan is generated, ask yourself the following question: was that a result of thinking about me, delineating a path from my current situation to the goal state, coming of with states I will go through on that path, making lists of steps that I could take to get between those states, and picking the ones that a person in my current state is likely to do - or was that the output of a cage of parrots? For me, there's only one answer!

Btw - I'm absolutely not saying these text generators are not a good tool: large numbers of people are already using them to improve their lives and increase their income - and you should definitely consider doing the same thing!

[Alexander Schmidt]

"...look at the openings. It plays instantly what humans developed in centuries." I let myself in to mom's house once, and I heard a man I knew well whistling. I looked around the house, and he wasn't there. Mom was looking after a parrot while the owners were away, and it had learned his whistle to absolute perfection.

You should better compare the human with the parrot. The AI is able to develop its opening knowledge without any input from outside.

[alfilete]

create a program that is only capable of seeing 20 to 300 positions at most. And face Carlsen at the tournament level, without a final basis (book?). There you will see who understands better.

[towforce]

"...look at the openings. It plays instantly what humans developed in centuries." I let myself in to mom's house once, and I heard a man I knew well whistling. I looked around the house, and he wasn't there. Mom was looking after a parrot while the owners were away, and it had learned his whistle to absolute perfection.

You should better compare the human with the parrot. The AI is able to develop its opening knowledge without any input from outside.

Very good point.

I meant the "finished product", though - not the training process.

In general, the rule for today's AI seems to be that a bigger parrot cage will get you a better {chess evaluation / generated text}, and at first glance you appear to be able to get away with doing that - until somebody decides to embarrass you by demonstrating that your cage of parrots is missing a simple piece of knowledge that everyone "just knows".

[Uri]

I think that those who reach top levels in Chess have very strong brains. The human brain is like a computer made of meat but some people have very good memory and a computer-like ability in the brain to reach very high levels in Chess.

Top Level Chess players like Kasparov, Bobby Fischer and Magnus Carlsen are like computers and I'm convinced that most people on Earth do not have this ability.

Personally, I think I will never be a good chess player because I think I developed Alzheimer disease and I also have many memory and neurological problems which prevent me from playing good chess without the help of a computer.

[towforce]

I think that those who reach top levels in Chess have very strong brains. The human brain is like a computer made of meat but some people have very good memory and a computer-like ability in the brain to reach very high levels in Chess.

Top Level Chess players like Kasparov, Bobby Fischer and Magnus Carlsen are like computers and I'm convinced that most people on Earth do not have this ability.

Personally, I think I will never be a good chess player because I think I developed Alzheimer disease and I also have many memory and neurological problems which prevent me from playing good chess without the help of a computer.

Have you done some online memory tests? Have you been diagnosed with Alzheimer? If you seriously think you might have it, I would strongly advise seeing your doctor as soon as possible.