Nah, the problem is we don't have enough humans making predictions. Have the entire population on Earth (that can communicate) predict how the world will be in 2023, being as detailed as possible, and just by sheer numbers, there'll be a considerable number of people that get every detail of their prediction right, even if their predictions have nothing to do with others on the group (because they'd focus on different things...)Dann Corbit wrote: ↑Tue Feb 05, 2019 9:01 pm We humans are really, really bad at predicting the future.
Humanity has written so much fiction in its history that it was impossible for some of them to fail in their prediction, which happens once most possibilities are covered.
So there's a chance I hit the nail on the head:
Within the next 5 years someone will find a way to "unblackbox" neural networks]. Currently it's hard to know what are they doing, and make tweaks. You need to keep training and wait until that lucky moment that triggers some learning that causes the to net play a better move in some situations, worse in others, but stronger overall to keep the change, and go up because those changes accumulate.
If you were able to "reverse-engineer" the black box, it'd happen like what happened to Rybka, people would be able to program chess engines that mimick, and produce the moves of nn, but using readable code, and then the sky will be the limit, as the code will not rely on minimax or alpha beta, the new paradigm will leave engines of today behind, and engines with the new paradigm will make Houdini 6 look as weak as Rybka 3 looks compared to it.
I've been discussing the future of computer chess since 12 years ago. Back then I was naive and believed that a 3600 CCRL engine would be very close to perfection, able to avoid losing any game. Now I know that the distance to perfection is still veeery large.
Stockfish 10 with 5men TBs still gives 0.08 scores to lost positions at depth 60. I play at time controls of the order of 50 days per game, with 50 day increment every 10 moves, and under those conditions it's clear state of the art engines still suck at 1 hour/move. As the current state of things are going, you'll eventually have engines that play like that only needing one minute, so they'll beat today's engines by the same magnitude of a 1 minute v 1 hour handicap. But those moves will still suck.
It's just that there's no current entity that is able to play the quality moves this fast, that is able to find the moves that takes weeks to find on these long correspondence time controls. For this you need another shift.
I saw such a shift coming, and when I saw Alpha Zero, I thought "this is it!", but there's only so much you can do with evolutionary algorithms, to do what I mean you need design.
Maybe in the future you will not have something like Leela Zero training on chess games itself to become stronger, but producing thousands of chess engines with the new paradigm. The difference will be that a human will be able to look at the code and see what works and what doesn't and only keep the best possible code that produces the best moves.
Because, a chess engine that could have some 6000 elo on the CCRL could theoretically exist, and if you had it in your harddrive right now, it'll be just some exe file, perhaps smaller than a MB, and it would consist of just 0s and 1s in some combination. A new paradigm could just find the right combination.
Otherwise, General Artificial Intelligence makes its appearance and is able to train in a day what a human would take 5000 years to learn, so it better be damn good at chess. But perhaps next decade...