Fantastic article. I had wondered what the followup would be. I'm mildly surprised that there was that much change just because an AI beat the world champ. Wanting to train against the AI might...
Fantastic article. I had wondered what the followup would be.
I'm mildly surprised that there was that much change just because an AI beat the world champ. Wanting to train against the AI might make sense to me if Alpha Go had made its weights public, though maybe another project which is as strong exists. I had assumed that most players would expect that an AI would surpass them in their lifetime, but the amount of change suggests otherwise.
It would be great if Mentats were real, but play style subsuming to AI is an easy thing to do. However, the AI could still be wrong. We know from Alpha Zero vs Stockfish that a more romantic style play appears to be viable at high levels, so maybe Alpha Go's style is still flawed, though no human will ever discover that at this point, only more powerful AI programs.
I see art and music eventually becoming automated with 'curators' selecting the best generated by their NN's. I wonder if art/music will be as affected as the Go industry when the computer appears to do something that was held to be truly creative.
There's a few. LeelaZero, KataGo, etc. They're good enough to beat pros with 2-3 stones handicap, which is really something when Gu Li said, "they asked how far away from God we are, and I said...
made its weights public, though maybe another project which is as strong exists.
There's a few. LeelaZero, KataGo, etc. They're good enough to beat pros with 2-3 stones handicap, which is really something when Gu Li said, "they asked how far away from God we are, and I said that 3 handicap stones would make a good match...".
edit: oh, I meant that the AI has the handicap. As in, KataGo will absolutely cream any human with no handicaps.
Good to know. I'm not good enough to tell if I'm being whalloped or destroyed by LeelaZero when I play it. It does seem premature to practice against AI though if a pro is still comparable to the...
Good to know. I'm not good enough to tell if I'm being whalloped or destroyed by LeelaZero when I play it. It does seem premature to practice against AI though if a pro is still comparable to the available ones?
The problem is that it can cost several thousands dollars to get 5 matches against a pro, it cost zero (well, if you have the computer) to practice against an AI--and you can do so infinitely.
The problem is that it can cost several thousands dollars to get 5 matches against a pro, it cost zero (well, if you have the computer) to practice against an AI--and you can do so infinitely.
I don't think high arts will be affected. But the vast majority of artists have no renown nor do their art for their own sake. They make illustrations for magazines, games, and so forth. Much of...
I don't think high arts will be affected. But the vast majority of artists have no renown nor do their art for their own sake. They make illustrations for magazines, games, and so forth. Much of that isn't supposed to be groundbreaking or really that creative, it just needs to look visually nice and in fact uniform to the rest of the product.
In that area, I think GANs will make that mostly obsolete. Rather than 20 artists, you can probably do with just 1. We already have very good results from them.
And so commercial art will be decimated. But, that is the way things go. On the other hand, for people who need art assets, this will make indie games, for example, a magnitude cheaper to make.
Current AI is not bound by the need to have objective rules. That was the last wave. DL NN's require three things: fast computers, massive amounts of data, and a means of scoring data by how good...
Current AI is not bound by the need to have objective rules. That was the last wave. DL NN's require three things: fast computers, massive amounts of data, and a means of scoring data by how good the result is.
Sites and subscription services to art: music, drawing, etc. Have massive amounts of the data that you would need to write this program. I would warrant that in a decade GPT-4, trained on that type of data would be able to surpass human performance in making hits.
Somewhat curiously, things like mathematics, physics, and arithmetic are harder for NN's to do than making art or playing a game. So we'll probably get another AI winter before GAI.
But you're right. No AI has managed to effectively reproduce Bach without plagiarizing from him. But, a lot of art fails to meet that criteria. I am fairly confident that the work I do right now will eventually be automated as well.
Edit: I should have clicked the comment to see stu2b50's comment first. Oh well.
There are already AI creating music, s0 I think we are n0t s0 far away fr0m seeing h0w it will affect this field. https://www.youtube.com/watch?v=ZzgGgFl9e4s
There are already AI creating music, s0 I think we are n0t s0 far away fr0m seeing h0w it will affect this field.
Fantastic article. I had wondered what the followup would be.
I'm mildly surprised that there was that much change just because an AI beat the world champ. Wanting to train against the AI might make sense to me if Alpha Go had made its weights public, though maybe another project which is as strong exists. I had assumed that most players would expect that an AI would surpass them in their lifetime, but the amount of change suggests otherwise.
It would be great if Mentats were real, but play style subsuming to AI is an easy thing to do. However, the AI could still be wrong. We know from Alpha Zero vs Stockfish that a more romantic style play appears to be viable at high levels, so maybe Alpha Go's style is still flawed, though no human will ever discover that at this point, only more powerful AI programs.
I see art and music eventually becoming automated with 'curators' selecting the best generated by their NN's. I wonder if art/music will be as affected as the Go industry when the computer appears to do something that was held to be truly creative.
There's a few. LeelaZero, KataGo, etc. They're good enough to beat pros with 2-3 stones handicap, which is really something when Gu Li said, "they asked how far away from God we are, and I said that 3 handicap stones would make a good match...".
edit: oh, I meant that the AI has the handicap. As in, KataGo will absolutely cream any human with no handicaps.
Good to know. I'm not good enough to tell if I'm being whalloped or destroyed by LeelaZero when I play it. It does seem premature to practice against AI though if a pro is still comparable to the available ones?
The problem is that it can cost several thousands dollars to get 5 matches against a pro, it cost zero (well, if you have the computer) to practice against an AI--and you can do so infinitely.
I don't think high arts will be affected. But the vast majority of artists have no renown nor do their art for their own sake. They make illustrations for magazines, games, and so forth. Much of that isn't supposed to be groundbreaking or really that creative, it just needs to look visually nice and in fact uniform to the rest of the product.
In that area, I think GANs will make that mostly obsolete. Rather than 20 artists, you can probably do with just 1. We already have very good results from them.
And so commercial art will be decimated. But, that is the way things go. On the other hand, for people who need art assets, this will make indie games, for example, a magnitude cheaper to make.
Current AI is not bound by the need to have objective rules. That was the last wave. DL NN's require three things: fast computers, massive amounts of data, and a means of scoring data by how good the result is.
Sites and subscription services to art: music, drawing, etc. Have massive amounts of the data that you would need to write this program. I would warrant that in a decade GPT-4, trained on that type of data would be able to surpass human performance in making hits.
Somewhat curiously, things like mathematics, physics, and arithmetic are harder for NN's to do than making art or playing a game. So we'll probably get another AI winter before GAI.
But you're right. No AI has managed to effectively reproduce Bach without plagiarizing from him. But, a lot of art fails to meet that criteria. I am fairly confident that the work I do right now will eventually be automated as well.
Edit: I should have clicked the comment to see stu2b50's comment first. Oh well.
There are already AI creating music, s0 I think we are n0t s0 far away fr0m seeing h0w it will affect this field.
https://www.youtube.com/watch?v=ZzgGgFl9e4s