Did AlphaZero Cheat Against Stockfish?
HTML-код
- Опубликовано: 13 сен 2024
- Did AlphaZero really cheat against Stockfish in their 2017 match? The match shocked the entire chess community and some people still don't believe anything could ever beat Stockfish. Let's explore if the match was fair and square.
How? Was it using Stockfish? 😆
Yes!
@@Stokfish YOOOO I watch your vids
@@Stokfish my left ear enjoys your videos quite a lot
"Cheating" implies it was a fair competition. It was not, nor was it intended to be. Stockfish was simply used as a baseline to see how well a self trained chess Neural Network engine could do. The results proved that a self taught Neural Network engine could be successful, and beyond that produce some beautiful chess. Since the two algorithms run on completely different types of hardware, it is not possible to have a fair competition.
Allow the same time period and electricity usage per move.
"It is not posible to have a fair competition" is esentially another way to say cheat.
@@petersmythe6462 i assume TPUs and CPUshave varying power consumptions so that may not be an option
Allow me to disagree
We can see "power consumption" as a good criteria
I dont care if it was fair or not
Power consumption was the same
Tpu is similar to asic miners for bitcoin, has no other use but neural network
Not true, anyone can play against a robot at their level. Despite being different “hardware”, it’ll still be a fair game
Welp 4 years later Stockfish shitted on alpha 0
Just wait for alpha one
You said it yourself; comparing tensor cores and normal CPU cores like in Stockfishes Xeon is not an apples to apples comparison. They need to be running on the EXACT SAME HARDWARE for it to be a fair match.
ok so have stockfish running on a tensor core.
@@homunculus3646 How?
@@homunculus3646 you can not make a program made to run on a general porpouse CPU to run on a TPU just like that.
The instruction set is not the same. This is like trying to write english on a japanese only keyboard (and I mean english with latin alphabet, not katakana).
@@fcolecumberri you can do the opposite tho
They literally did
This actually makes sense bcs of the way the bot learns
Wow we need to report Alpha zero for using stockfish for cheating
They did a rematch and Alpha Zero still kicked Stockfish's ass
so you're telling me 2017 AZ beat a much later version of SF in a rematch? :O
@@ngwanamama1572 yes he is
@@ngwanamama1572 I think the newest stockfish (15/16) is better than alphazero
@@astrovation3281 yeah Stockfish15 stomped AZ
@@Rimizu Yes, but SF15 NNUE is using some of the technology from AZ.
Alpha zero is talking. We need to be scared
I think AlphaZero is the younger brother of Stockfish, he trained AlphaZero to take his position. idk just saying it for fun
What about today's latest version of stockfish is it powerful
he used an engine definitly cheater
Did AlphaZero cheated? Yes.
Computer chess is not a hardware battle, on engine tournaments, you don't see on tournaments people with GPUs/FPGAs/ASICs trying to see who has built a better computer to play chess. You see computer programs on a standarized computer with no extra computing power. (Actually some tournaments uses very old/bad computers to force less draws or just to be consistent within the last 20 years).
If you want to compare stockfish with a0, you should not give stockfish a better cpu, you shoud take a0's TPUs away.
I see your point but AlphaZero's algorithm is suited to run on TPUs however Stockfish runs on mainstream hardware. Even today, Stockfish can use GPUs against Lc0 on TCEC matches, but chooses not to. This doesn't make Stockfish a "lesser" engine. But these choices matter and at that point in time AlphaZero had the advantage.
That's unfair in the other direction. Stockfish is optimized to run on CPU and A0 is optimized to run on TPU.
The more relevant comparison would be equal power consumption and time. I.E. Give A0 and stockfish each a 1 hour clock and limit them to, say, 1 kW of power consumption or whatever.
Or if we really want to be cruel, 20 W of power consumption since that is a direct comparison to the human brain.
@@petersmythe6462 that's not how computer chess works, every chess engine is made to run on a general computer, you cannot just break the rules and assume it is fair.
However you are talking about "fairness" by giving a measurement of what is fair (in this case power). Here the problem is still obvious, the TPUs are made with instructions to process tensors (esentially matrix operations), therefore there are aingle instructions that would require several operations on a CPU.
This would be like being on a math contest and give the first participant a classic calculator (can only sum, rest, multiply and divide) and the second participant give a programmers calculator (with every constant, trigonometry functions, etc.) And compare how much power they consumed to calculate "sin(27.58)". How is that fair? There is not even a way to calculate "fairness" that's why stuff is standarized.
Claiming that if an engine is breaking the rules then is the others engines fault because the don't also break them is not "fair".
you fundamentally misunders5and how neural networks work. They have a training phase and an evaluation phase. They trained it on a super computer to get the best parameters. But all a super computer does is increase the speed of the training. Stock fish has been in the works for much longer.
Alpha Zero fought stock fish on the same hardware. That's like saying stock fish cheated because it took years to develop it.
@@jong7100 In the paper that they published, they say that they used 64 TPUs to train and 4 TPUs to play. Did you read that?
Where are you?
2 minutes in. 2 minutes wasted. wtf dude c'mon
mhm
Dude, just buy a microphone.
It fucking cheated
It was a good engine why No one published it
It’s too good that it will become not the point of chess, it was meant to be a human mind thinking
Because:
1. It's not actually an engine, it's a neural net.
2. It requires specialized hardware to run that's not exactly available to most people.
and
3. Google.
First