AlphaZero - is this what 4000 rated chess looks like?

[Sid Belzberg]

How exactly would AZ cope against that?

I would expect in a very similar way that the AZ team did with their secondary evaluation against the Stockfish opening book where deliberate randomization was introduced at the early stages of the game. In genetics mutations are sometimes good as it adds to the diversity of the population. Genetic algos are designed to introduce random mutations to expand the horizon a bit.

AZ may turn out in some future world to have been merely a stepping stone towards the strongest chess engine imaginable

Again you are talking about improbable scenarios. AZ during self play plays out trillions(!) of positions every second and then ranks the results. This form of optimization has proven successful in not only chess but other games. More importantly recently the AZ technology was used to analyze genetic sequences and successfully predict how the resulting protein folds. It was correct a greater percentage of the time then anytime in history. The evidence is near 100% that AZ is truly great at optimizing with little room for improvement. Nothing to do with beliefs or biases, just evidence.
Your argument reminds me of saying something like "just because nothing appears to go faster then the speed of light maybe in some other universe nature is very different". Meanwhile in this universe we will rely on reproducible evidence we have now and not something that can be neither proven or disproven. That has nothing to do with biases or beliefs or "thinking out of the box". Thinking out of the box does not mean proposing things that can't be proven or disproven with zero evidence to support it. I don't consider people that believe in ghosts with no evidence to support it as having some sort of superior intuition that allows them to "think out of the box".

Even with your argument that we have no proof that Trump's wall will help make America safer if you are looking for an iron clad 100 percent proof I agree with that. However, walls have worked very well in making Israel safer and in the few areas where border walls have been erected in America there was a significant reduction in illegal crossings. However it is possible that suddenly people will cross walls via 10s of thousands of hot air balloons or tunnels may be come a cottage industry across the entire length of the wall or everyone will become great pole vauters , not impossible but IMPROBABLE. At least Trump has evidence to say America will be safer, it is not based on improbable scenarios that did not happen in other countries that have used walls.
The day may even happen where personal passenger drones are so common place and inexpensive that walls anywhere will be useless, in our lifetime passenger drones exist even now but for widespread acceptance even to be possible would require a complete overhaul of air regulations that in our lifetimes is...IMPROBABLE .

I will think about your idea of "should I do it now or later" as a factor that in itself would be interesting to explore. That sound's interesting but I would need to give it more thought.

[Paul Bonham]

Again you are talking about improbable scenarios. AZ during self play plays out trillions(!) of positions every second and then ranks the results. This form of optimization has proven successful in not only chess but other games. More importantly recently the AZ technology was used to analyze genetic sequences and successfully predict how the resulting protein folds. It was correct a greater percentage of the time then anytime in history. The evidence is near 100% that AZ is truly great at optimizing with little room for improvement. Nothing to do with beliefs or biases, just evidence.
Your argument reminds me of saying something like "just because nothing appears to go faster then the speed of light maybe in some other universe nature is very different". Meanwhile in this universe we will rely on reproducible evidence we have now and not something that can be neither proven or disproven. That has nothing to do with biases or beliefs or "thinking out of the box". Thinking out of the box does not mean proposing things that can't be proven or disproven with zero evidence to support it. I don't consider people that believe in ghosts with no evidence to support it as having some sort of superior intuition that allows them to "think out of the box".

Even with your argument that we have no proof that Trump's wall will help make America safer if you are looking for an iron clad 100 percent proof I agree with that. However, walls have worked very well in making Israel safer and in the few areas where border walls have been erected in America there was a significant reduction in illegal crossings. However it is possible that suddenly people will cross walls via 10s of thousands of hot air balloons or tunnels may be come a cottage industry across the entire length of the wall or everyone will become great pole vauters , not impossible but IMPROBABLE. At least Trump has evidence to say America will be safer, it is not based on improbable scenarios that did not happen in other countries that have used walls.
The day may even happen where personal passenger drones are so common place and inexpensive that walls anywhere will be useless, in our lifetime passenger drones exist even now but for widespread acceptance even to be possible would require a complete overhaul of air regulations that in our lifetimes is...IMPROBABLE .

I will think about your idea of "should I do it now or later" as a factor that in itself would be interesting to explore. That sound's interesting but I would need to give it more thought.

The whitepaper you sent me explicitly says AZonly analyses 60,000 positions per second, not "trillions".

On the genetic sequencing point, how successful was AZ at predicting the protein folding? And where are your sources?

Perhaps predicting protein folding is much simpler than playing perfect chess, you can't just compare apples and oranges and make them the same.

You say "Thinking out of the box does not mean proposing things that can't be proven or disproven with zero evidence to support it." I say balderdash to that. Thinking out of the box can be many things, and what you described is definitely one of those things. If someone had told you 20 years ago that there would be a chess engine in 20 years that could learn to play chess from just the game rules and in 4 hours be strong enough to outplay all human GMs, that would have been proposing something that couldn't be proven nor disproven. The person making that prediction would have been thinking out of the box, perhaps anticipating the very technique of AZ years before its time.

I know you are highly intelligent Sid but you are really like a beached whale when it comes to the concept of thinking out of the box. I don't say that to embarrass you or make you look foolish, you have many other credits that make you valuable in many fields. But on this one thing, you are not good. Not everyone thinks out of the box; only a very small minority do it regularly and effectively.

Jeff Bezos thought out of the box when he imagined that all of America might someday make most of their purchases online. For years his ideas were ridiculed and Amazon stock prices languished. His claims had zero evidence to support them, and could neither be proven nor disproven. But he stuck to them, he thought out of the box. And who's laughing now?

Most of the progress of human history has come about by someone somewhere thinking out of the box. Of course, not every outrageous claim should be believed and many will never become reality. But every one of them should be at least CONSIDERED.

So I am glad to hear that you are considering the idea of "should something be done now or done later" because that is the central theme of Option Chess, and I believe OC may be the only game that has that idea as its central theme. Or at least it seems to be the only chess variant that has that idea as its central theme. The only thing I have been asking of you or anyone else is to consider that idea, to think of it as something new and something valuable to explore for the field of AI. Now that you know I have no vested interest in OC, and will make no profit from it if it became as big as chess itself, you might be willing to actually take a deeper look into it.

[Sid Belzberg]

The whitepaper you sent me explicitly says AZo only analyzes 60,000 positions per second, not "trillions".

That refers to the fully "trained" product. In order to get the weights for the NN AZ needed to be trained on 5000 of google's TPU's, effectively a super computer. The trillions refers to the training period.

On the genetic sequencing point, how successful was AZ at predicting the protein folding? And where are your sources?

http://www.sciencemag.org/news/2018/...rotein-folding

If someone had told you 20 years ago that there would be a chess engine in 20 years that could learn to play chess from just the game rules and in 4 hours be strong enough to outplay all human GMs that would have been proposing something that couldn't be proven nor disproven. The person making that prediction would have been thinking out of the box, perhaps anticipating the very technique of AZ years before its time.

In 2007 a few employees and a few outsiders from our previous company including myself formed a chess team and won the Bankers Athletic Chess League team chess competition that year in Manhattan. We had a great team with GM Pascal Charbonneau , WGM Irina Krush and several other players way better then myself on the team and took first place ahead of the GM laden Goldman Sacs team and others.
GM Joel Benjamin gave out the awards and after the ceremony we went out together to a nearby pub. Joel was employed by IBM and was part of the original deep blue team that defeated a hapless Garry Kasparov in 1997. He had given a talk that night on the controversy surrounding that event. I remember walking to the pub with him and I told him about research I had been working on to use Genetic Algo's on parallel computers to self train a chess engine. I had downloaded the source code to a chess program called crafty and had also downloaded source code a GA used for solving the knapsack problem in 1994(!). I had mentioned that some degree of success with this technique had been used with backgammon.

His reply was "and so???" I said to him "that when you have a methodology that employs random tries and via trial and error that evolves a solution ,that more closely resembles true AI then the mechanical brute force approach together with preprogrammed evaluation functions used by chess engines." I was mildly surprised at the time at his skepticism. My idea was that if you had enough computers in parallel you might be able to make the thing work. Of course the break through was effectively marrying a GA (MTCT) to a NN and learn via testing patterns rather then move by move. Go programming language was another helpful creation (2007) that makes it easier to spawn of separate threads on separate CPU's, in paralell you simply add the statement GO but I digress.

My point is that although all those years ago this was considered to be a very wild idea but, there was evidence that GA's work in areas like robots that taught themselves to walk or as mentioned backgammon. So even by your definition I am more of a fan of "out of the box" thinking then you might think but I am a fan of ideas being supported by at least some evidence.

Jeff Bezos's vision was supported by evidence of rapid acceleration of revenues . It was obvious to many that ironing out the economics of fulfillment was only a matter of time as evidenced by the billions of dollars of investment poured into the company both in the form of debt and equity by many who agreed with his out of the box vision.