LCZero is confusing

[Jouni]

Here people speak about ELO 500, but elsewhere http://162.217.248.187/ I see ELO 3650!?!?

[mkchan]

Here people speak about ELO 500, but elsewhere http://162.217.248.187/ I see ELO 3650!?!?

I think here we are referring to a common rating list like CCRL or CEGT etc, but there they have a different pool of players (likely just older versions of LC0) so you see it being 3650 rated in that pool

[CMCanavessi]

Here people speak about ELO 500, but elsewhere http://162.217.248.187/ I see ELO 3650!?!?

That 3000+ number comes from self-play games, with is absolutely useless when compared to any rating list.

Real elo of the current network generation is around 1300-1500 elo, depending on hardware and time control.

[AdminX]

Here people speak about ELO 500, but elsewhere http://162.217.248.187/ I see ELO 3650!?!?

That 3000+ number comes from self-play games, with is absolutely useless when compared to any rating list.

Real elo of the current network generation is around 1300-1500 elo, depending on hardware and time control.

Well that makes sense, as there is no way in hell I could have beat a 3000+ computer on my 1st attempt.

[Dann Corbit]

The number Elo is meaningless except against the competition in a pool.
The high Elo number is against a random mover.

It will take a long time for LCZero to get to SF level.

The interesting thing is the big, big gains.
We will see if that continues.
I think it is time for a bigger net. Now we see a lot of failures and a slower gain.

[lucasart]

The number Elo is meaningless except against the competition in a pool.
The high Elo number is against a random mover.

It will take a long time for LCZero to get to SF level.

The interesting thing is the big, big gains.
We will see if that continues.
I think it is time for a bigger net. Now we see a lot of failures and a slower gain.

I'm not convinced it's possible, even with infinite resources thrown at it. Already we see that lczero progress is showing strongly diminishing returns. Eventually we'll reach an asymptotic elo level, and I'm ready to bet it will be nowhere close to SF.

I don't think people realize just how enormous the gap is between lczero and SF. It seems to be taken for granted that lczero will reach SF level, and that it's only a matter of time. I don't think so.

But let's wait and see

[CheckersGuy]

The number Elo is meaningless except against the competition in a pool.
The high Elo number is against a random mover.

It will take a long time for LCZero to get to SF level.

The interesting thing is the big, big gains.
We will see if that continues.
I think it is time for a bigger net. Now we see a lot of failures and a slower gain.

I'm not convinced it's possible, even with infinite resources thrown at it. Already we see that lczero progress is showing strongly diminishing returns. Eventually we'll reach an asymptotic elo level, and I'm ready to bet it will be nowhere close to SF.

I don't think people realize just how enormous the gap is between lczero and SF. It seems to be taken for granted that lczero will reach SF level, and that it's only a matter of time. I don't think so.

But let's wait and see

The current net with 6 blocks will eventuall stall that's perty obvious but there is nothing that can stop us from going 10,20 blocks once that happens

[Henk]

Were the alpha zero games the most expensive chess games ever played?

[hgm]

Stockfish development has been much more expensive, not?

I guess Deep Blue must hold the record, as it used massive hardware that was no good for anything else.

[Henk]

I mean the price per recorded game. The more Stockfish games were recorded the lower the price/value per game.