Reactions about AlphaZero from top GMs...

[Rein Halbersma]

Yes, but Tord also said the most important thing, IMO. In maybe 5 years, the times of low budget single brilliant author or 1-2-3 brilliant authors of close to top engines will end. Massive hardware will be needed for training the networks, and such giants as Google or whatever have a decisive advantage. Maybe at top will be 2-3 large distributed projects aside from some large companies. Seems a pretty gloomy future.

Why so gloomy? Cost of hardware will go down, and the curent A0 is no doubt over-engineered (burning a few TPU years in parallel must be cheaper than a dozen man years of careful algorithm design). Apart from distributed open source efforts, there are also things like transfer learning and network compression that might greatly reduce the burden of replicating (and then surpasing) the current state of the art. After all, Deep Blue has also been surpassed by today's engines. But first, there needs to be a full paper with enough details to replicate AlphaZero-Chess.

[Laskos]

Yes, but Tord also said the most important thing, IMO. In maybe 5 years, the times of low budget single brilliant author or 1-2-3 brilliant authors of close to top engines will end. Massive hardware will be needed for training the networks, and such giants as Google or whatever have a decisive advantage. Maybe at top will be 2-3 large distributed projects aside from some large companies. Seems a pretty gloomy future.

Why so gloomy? Cost of hardware will go down, and the curent A0 is no doubt over-engineered (burning a few TPU years in parallel must be cheaper than a dozen man years of careful algorithm design). Apart from distributed open source efforts, there are also things like transfer learning and network compression that might greatly reduce the burden of replicating (and then surpasing) the current state of the art. After all, Deep Blue has also been surpassed by today's engines. But first, there needs to be a full paper with enough details to replicate AlphaZero-Chess.

The paradigm shifts, nevertheless. Mark Lefler was improving on Komodo using his laptop while on an air flight. These days will be over. And, again, hardware advantage will be decisive, not in running these engines (we will all run them on comparable hardware anyway), but in building that software.

[JJJ]

Meanwhile, I d like to see more game played and alphazero trained longer ! Anyone knows if more game are coming or if alphazero will be more trained at chess ?

[shrapnel]

Meanwhile, I d like to see more game played and alphazero trained longer !

+1.
That should happen, otherwise if DeepMind goes quiet, it will only encourage the detractors of AlphaZero to find their voices again.

[Uri Blass]

Yes, but Tord also said the most important thing, IMO. In maybe 5 years, the times of low budget single brilliant author or 1-2-3 brilliant authors of close to top engines will end. Massive hardware will be needed for training the networks, and such giants as Google or whatever have a decisive advantage. Maybe at top will be 2-3 large distributed projects aside from some large companies. Seems a pretty gloomy future.

Why so gloomy? Cost of hardware will go down, and the curent A0 is no doubt over-engineered (burning a few TPU years in parallel must be cheaper than a dozen man years of careful algorithm design). Apart from distributed open source efforts, there are also things like transfer learning and network compression that might greatly reduce the burden of replicating (and then surpasing) the current state of the art. After all, Deep Blue has also been surpassed by today's engines. But first, there needs to be a full paper with enough details to replicate AlphaZero-Chess.

The paradigm shifts, nevertheless. Mark Lefler was improving on Komodo using his laptop while on an air flight. These days will be over. And, again, hardware advantage will be decisive, not in running these engines (we will all run them on comparable hardware anyway), but in building that software.

I am not sure if hardware advantage will be decisive in building that software.

People may come with new ideas how to use slower hardware and still be competitive.

[CheckersGuy]

Yes, but Tord also said the most important thing, IMO. In maybe 5 years, the times of low budget single brilliant author or 1-2-3 brilliant authors of close to top engines will end. Massive hardware will be needed for training the networks, and such giants as Google or whatever have a decisive advantage. Maybe at top will be 2-3 large distributed projects aside from some large companies. Seems a pretty gloomy future.

Why so gloomy? Cost of hardware will go down, and the curent A0 is no doubt over-engineered (burning a few TPU years in parallel must be cheaper than a dozen man years of careful algorithm design). Apart from distributed open source efforts, there are also things like transfer learning and network compression that might greatly reduce the burden of replicating (and then surpasing) the current state of the art. After all, Deep Blue has also been surpassed by today's engines. But first, there needs to be a full paper with enough details to replicate AlphaZero-Chess.

The paradigm shifts, nevertheless. Mark Lefler was improving on Komodo using his laptop while on an air flight. These days will be over. And, again, hardware advantage will be decisive, not in running these engines (we will all run them on comparable hardware anyway), but in building that software.

I am not sure if hardware advantage will be decisive in building that software.

People may come with new ideas how to use slower hardware and still be competitive.

I think it's not unlikely that one can improve upon the algorithm deepMind used with some additional domain knowledge. Remember, they wanted to show that their algorithm works for any (board) game but if we are only intrested in chess there should be ways to speed up the training process

[Milos]

Yes, but Tord also said the most important thing, IMO. In maybe 5 years, the times of low budget single brilliant author or 1-2-3 brilliant authors of close to top engines will end. Massive hardware will be needed for training the networks, and such giants as Google or whatever have a decisive advantage. Maybe at top will be 2-3 large distributed projects aside from some large companies. Seems a pretty gloomy future.

Why so gloomy? Cost of hardware will go down, and the curent A0 is no doubt over-engineered (burning a few TPU years in parallel must be cheaper than a dozen man years of careful algorithm design). Apart from distributed open source efforts, there are also things like transfer learning and network compression that might greatly reduce the burden of replicating (and then surpasing) the current state of the art. After all, Deep Blue has also been surpassed by today's engines. But first, there needs to be a full paper with enough details to replicate AlphaZero-Chess.

The paradigm shifts, nevertheless. Mark Lefler was improving on Komodo using his laptop while on an air flight. These days will be over. And, again, hardware advantage will be decisive, not in running these engines (we will all run them on comparable hardware anyway), but in building that software.

I am not sure if hardware advantage will be decisive in building that software.

People may come with new ideas how to use slower hardware and still be competitive.

I think it's not unlikely that one can improve upon the algorithm deepMind used with some additional domain knowledge. Remember, they wanted to show that their algorithm works for any (board) game but if we are only intrested in chess there should be ways to speed up the training process

MCTS used in A0 you can probably somewhat improve without having monstrous hardware. However, improving actual NN is really out of reach for regular ppl, at least for next 5-10 years if not longer.
Basically already after 4 hours of training (around 20 million self-played games) current NN performance was saturated.
After that most of games in playouts were draws.
To improve this more one would need larger NN (there are indications used NN was 20 blocks one, not 40), but that would only mean much longer duration of self-play games and more hardware for actual training, so even harder for regular developer without access to monstrous hardware.