Alphazero news

[corres]

What kind of reverse engineering do you understand?

I've done reverse engineering through static analysis, both professionally and as a hobby. A lot of that was done without ever running a line of the program (e.g. decompiling Mac software without ever having owned a Mac).

From an obfuscated program one can get its source (or at least what works similarly ) if the program is running.

You don't need the program to be running for that, just as much as you don't need a C compiler to understand what a C program listing does.

If the program is not obfuscated you can do that.
In the case of a well obfuscated program this method is unusable.

[corres]

I don't get some of the negative talk about A0 and the constant bickering about test conditions ... have you guys played through some of those games ??? Obviously there is something "game changing" happening. If you go through the games of some of the human chess geniuses of our time, on a really good day they will come up with some amazing plans that seem so incredible and without any tangible tactical short term advantage based purely on intuition. But on many of these games the initial intuition is correct but then some slight tactical inaccuracy totally screws up the game. I feel A0 gives you that but with a much more developed tactical awareness. I can't help that a perfect engine would be something like A0, but with a SF type engine running passively and covering some deep tactical blind spots that A0 might miss. This engine would only flag a move if there is a large fail due to a tactic. Actually now that I think of it the standard SF is probably not the best for this and a tactical version would probably do better as a helper. I always dreamed of an engine that could use a strong CPU based engine and have a daughter card to do a Monte Carlo search passively. Maybe a better idea is to use something like A0 as the main engine on the daughter card and then have the CPU based engine do the sanity check.

If we use in parallel a classical AB chess engine with an NN-type chess engine we should answer the question:
Who decide what program is right when they value the position differently?
In the case of an obvious tactical blunder (typically NN) we are in a light situation.
But in the most cases the situations are too complicated even for a super GM too.
Moreover the method of evaluation inherently different in an AB engine and in an NN engine.
AB engine uses centi pawn NN engine uses probability.

[jp]

I ... have you guys played through some of those games ??? Obviously there is something "game changing" happening. If you go through the games of some of the human chess geniuses of our time, on a really good day they will come up with some amazing plans that seem so incredible and without any tangible tactical short term advantage based purely on intuition. But on many of these games the initial intuition is correct but then some slight tactical inaccuracy totally screws up the game. I feel A0 gives you that but with a much more developed tactical awareness.

This is a bit unfair to "human chess geniuses". If you allow human chess geniuses to play a huge number of games and then only see the <1% of games they select for you to see, then you won't think they can do it on only a really good day.

[Sesse]

From an obfuscated program one can get its source (or at least what works similarly ) if the program is running.

You don't need the program to be running for that, just as much as you don't need a C compiler to understand what a C program listing does.

If the program is not obfuscated you can do that.
In the case of a well obfuscated program this method is unusable.

That depends heavily on the obfuscation and the static analysis tools at your disposal.

In any case, I can't see Google taking such a risk for something as mundane as helping out the Leela team with something to compare results to. If they really wanted to do that, and they didn't want to just give them the source code (which presumably wouldn't compile due to missing dependencies, of course), it would be a lot easier just giving access to the service.

[jp]

In any case, I can't see Google taking such a risk for something as mundane as helping out the Leela team with something to compare results to. If they really wanted to do that, and they didn't want to just give them the source code (which presumably wouldn't compile due to missing dependencies, of course), it would be a lot easier just giving access to the service.

What's wrong with them just releasing the NN TensorFlow SavedModel & letting the public do whatever they can with it?

[mirek]

You're aware that this paper was submitted a year ago, right, long before the RTX series was announced? Google had no say the lead time here; that was up to the journal.

Not true, the paper was submitted in March, RTX was announced in August, but unofficially it was very well known that 20xx generation would have tensor cores support and be similar in performance to Titan V (that already existed) well before March this year.

Haha don't try to be funny.
http://talkchess.com/forum3/viewtopic.p ... 87#p771687
In this CPU vs. GPU thread where I mentioned that the upcoming RTX cards could significantly shift the performance / price ratio and thus render your conclusions about CPU vs. GPU performance at similar price as invalid in the very near future, you wrote:

"Consumer(gaming) cards will never have real fp16 support or tensor cores. NVIDIA is not crazy to ruin themselves the chance to earn more by selling "AI tergeted" cards for much more money.
For consumer cards 20xx will be roughly 15% faster than 10xx. That is the same kind of speedup compared to difference between Ryzen/Threadripper 1xxx and 2xxx. So difference between price optimal CPU and GPU machines will remain the same"

(And that was on August 18th just days from official release)

So I think that if even such expert as you didn't know about it or was explicitely arguing against fp16/tensor speedup then it probably wasn't such an "obvious" unofficial knlowledge despite that now you act as if nothing should come to us as more natural than that the RTX cards will have performance comparable to Titan V

[Milos]

You're aware that this paper was submitted a year ago, right, long before the RTX series was announced? Google had no say the lead time here; that was up to the journal.

Not true, the paper was submitted in March, RTX was announced in August, but unofficially it was very well known that 20xx generation would have tensor cores support and be similar in performance to Titan V (that already existed) well before March this year.

Haha don't try to be funny.
http://talkchess.com/forum3/viewtopic.p ... 87#p771687
In this CPU vs. GPU thread where I mentioned that the upcoming RTX cards could significantly shift the performance / price ratio and thus render your conclusions about CPU vs. GPU performance at similar price as invalid in the very near future, you wrote:

"Consumer(gaming) cards will never have real fp16 support or tensor cores. NVIDIA is not crazy to ruin themselves the chance to earn more by selling "AI tergeted" cards for much more money.
For consumer cards 20xx will be roughly 15% faster than 10xx. That is the same kind of speedup compared to difference between Ryzen/Threadripper 1xxx and 2xxx. So difference between price optimal CPU and GPU machines will remain the same"

(And that was on August 18th just days from official release)

So I think that if even such expert as you didn't know about it or was explicitely arguing against fp16/tensor speedup then it probably wasn't such an "obvious" unofficial knlowledge despite that now you act as if nothing should come to us as more natural than that the RTX cards will have performance comparable to Titan V

I don't work for either NVIDIA or Google, or any GPU manufacturer so I don't have insider information. So contrary to your mocking and condescending comment, I am not an "expert" in usage of crystal ball that can tell the future, but can only comment as an analyst and as many professional analysts (I am not professional in any way) I can be ofc wrong.
OTOH a multi billion dollar company such as Google certainly has insider information of what their direct competitor in the field of ML hardware is doing. If you really believe they had no clue what NVIDIA would come up with, there is not much more to comment about, except that you are probably extremely naive...

[matthewlai]

I can be ofc wrong.

You are .

I happen to have worked for both NVIDIA and Google, and I can confirm that neither knows much more about upcoming hardware from their competitors before release than the general public. If it was so common you wouldn't have Intel releasing a chip in a big event tauting it as the highest core count workstation chip (and heavily highlighting that part), only for AMD to release an even bigger chip a few days later. If you were the PR person at Intel, and you had that information, would you have done what Intel did? They probably would have done something different much earlier, too (moving their mainstream chips to 6 or 8 cores) if they knew Zen chips were going to be so strong back in 2015-2016.

Is there a super secret industrial espionage and infiltration team within Google that knows those things? If there is one I and everyone on the AZ team have certainly not heard of it.

Decision to write and release the paper had absolutely nothing to do with what NVIDIA was doing. Knowing how NVIDIA works, back when we wrote the paper, probably even NVIDIA themselves didn't know how fast those chips are because they probably wouldn't even have taped out by then. Or if it was, it probably was still so bug-ridden that it wouldn't be able to finish any reasonably complex benchmark.

I have avoided replying to your comments because I actually have a lot of interesting science that still needs to be done, and couldn't really care less if you just hate Google and wants to throw as much dirt as possible on whatever they do and call everyone at Google liars (by the way, we are under Alphabet, but we are not even Google). Ok I really should get back to work...

If you hate Google so much, write your own program that can beat Stockfish with a bigger margin than AZ. Or beat them in something else you care about. Do something useful for this world.

[shrapnel]

Just ignore Milos. He's just an old-school Commie (apart from being the most negative person I've ever seen) who thinks that all Private Enterprise is bad.
Or probably has some Psychiatric Disorder and doesn't take his Meds regularly.

[Laskos]

Just ignore Milos. He's just an old-school Commie (apart from being the most negative person I've ever seen) who thinks that all Private Enterprise is bad.
Or probably has some Psychiatric Disorder and doesn't take his Meds regularly.

Well, nothing against Google's scientific efforts or Deep Mind, but first Google is one the largest tax avoiders in the World, with a prejudice in many dozens of billions annually. Second, it is the biggest private collector of consumers' personal data, all of which, among other nice things, is released to NSA. So, no much ethical reasons to really love these things like Google.
Again, all this has nothing to do with this top-notch Deep Mind "Science" paper.