I recently read this Wired story The AI Cold War That Threatens Us All.
It sounds very dire. The author tries to make the case that the US and China's positions in AI are like the cold war between the US and the Soviet Union. I see some problems with this conclusion though. First of all, I'm not seeing something like the nuclear warhead/missile development. Once you've built a successful nuclear weapon there is a limit to what you can do with it. Pretty much, you can only make it smaller, more reliable, and cheaper. There isn't really any practical application. Missiles on the other hand developed into rockets to launch satellites and eventually humans into space. This technology is still practical and useful.
I'm not seeing that with AI. What it reminds me more of would be ship technology. I think of the Cog which transitioned from oars to sail. That reminds me of ENIAC which transitioned between adding machines and punch card readers to machines that were more programmable. Then the Dartmouth conference reminds me of the transition from sail to steam when the prime weapon of war became the Monitor. This era ended with HMS Dreadnought in 1905. That reminds me of deep learning. After Dreadnought there was rapid advance with the US building its own dreadnoughts, then super-dreadnoughts, and finally the last iteration with the fast battleships in WWII. It's hard to look at a 16 or 18 gun and see it as obsolete but the reason for the speed was the recognition that aircraft carriers were important and battleships had to keep up with them.
You read the glowing accounts of Deep Blue, Watson, Alpha Go and now Alpha Zero. No shortage of excitement. I guess it wasn't quite so exciting to me. There may be a chess playing program today that can defeat human players, but you can't really tell. It doesn't make a great deal of sense to have a game that is finely tuned for human limitations and then compete against a system with none of those limitations. Watson was similar. Jeopardy is based on human limitations which Watson didn't have. That Watson won is of no consequence. Any question that required comprehension was a non-starter for Watson. Ultimately, all it showed was that Watson was a fast, trivia database. No surprise. Even the old punch card readers were faster than humans.
But we seem to have special hype with Alpha Go and Alpha Zero. Apparently it was a surprise that Go was susceptible to statistical analysis. Why? There was a time when people claimed that chess wasn't susceptible to statistical analysis. It just took a bigger hammer. And then you have Alpha Zero. It defeated Stockfish. And apparently some people think this was a big deal. Stockfish attained its level of play after 1,200 CPU hours of simulation. Alpha Zero had 16,000 CPU hours of simulation. Stockfish was also not configured in its normal way. But, still, people scratch in the dirt trying find some significance. The only possible significance seems to be the use of a Monte Carlo algorithm. Komodo has already added this. I wouldn't be surprised if Stockfish did too.
So, getting back to the AI cold war. Supposedly this is because China claims that it will be a world leader by 2030. Could this happen? Well, maybe but I'm not seeing this with their current direction. Of course, in fairness I'm not seeing it with Google or IBM either. Much like the WWII battleships, systems like Alpha Zero will be obsolete soon. I'm also not seeing it on the hardware front. Intel is having to continue production at 14nm because they can't yet produce enough volume at 10nm. Kurzweil thought we would be down to 3.5 nm by now and looking at 1 nm soon. Instead it looks like we'll hit bottom at 10nm or maybe 7nm. And that's the end of it. I think there is a path that could get a few more generations, but I haven't seen anyone working in that direction.
What I do see almost everyday is that we are drowning in information. We can't handle what we have now, but we would need a lot more to increase efficiency. This seems to me where the major advancement will come. So, I suppose there could be some kind of war in the future, but I'm not seeing it starting yet.
It sounds very dire. The author tries to make the case that the US and China's positions in AI are like the cold war between the US and the Soviet Union. I see some problems with this conclusion though. First of all, I'm not seeing something like the nuclear warhead/missile development. Once you've built a successful nuclear weapon there is a limit to what you can do with it. Pretty much, you can only make it smaller, more reliable, and cheaper. There isn't really any practical application. Missiles on the other hand developed into rockets to launch satellites and eventually humans into space. This technology is still practical and useful.
I'm not seeing that with AI. What it reminds me more of would be ship technology. I think of the Cog which transitioned from oars to sail. That reminds me of ENIAC which transitioned between adding machines and punch card readers to machines that were more programmable. Then the Dartmouth conference reminds me of the transition from sail to steam when the prime weapon of war became the Monitor. This era ended with HMS Dreadnought in 1905. That reminds me of deep learning. After Dreadnought there was rapid advance with the US building its own dreadnoughts, then super-dreadnoughts, and finally the last iteration with the fast battleships in WWII. It's hard to look at a 16 or 18 gun and see it as obsolete but the reason for the speed was the recognition that aircraft carriers were important and battleships had to keep up with them.
You read the glowing accounts of Deep Blue, Watson, Alpha Go and now Alpha Zero. No shortage of excitement. I guess it wasn't quite so exciting to me. There may be a chess playing program today that can defeat human players, but you can't really tell. It doesn't make a great deal of sense to have a game that is finely tuned for human limitations and then compete against a system with none of those limitations. Watson was similar. Jeopardy is based on human limitations which Watson didn't have. That Watson won is of no consequence. Any question that required comprehension was a non-starter for Watson. Ultimately, all it showed was that Watson was a fast, trivia database. No surprise. Even the old punch card readers were faster than humans.
But we seem to have special hype with Alpha Go and Alpha Zero. Apparently it was a surprise that Go was susceptible to statistical analysis. Why? There was a time when people claimed that chess wasn't susceptible to statistical analysis. It just took a bigger hammer. And then you have Alpha Zero. It defeated Stockfish. And apparently some people think this was a big deal. Stockfish attained its level of play after 1,200 CPU hours of simulation. Alpha Zero had 16,000 CPU hours of simulation. Stockfish was also not configured in its normal way. But, still, people scratch in the dirt trying find some significance. The only possible significance seems to be the use of a Monte Carlo algorithm. Komodo has already added this. I wouldn't be surprised if Stockfish did too.
So, getting back to the AI cold war. Supposedly this is because China claims that it will be a world leader by 2030. Could this happen? Well, maybe but I'm not seeing this with their current direction. Of course, in fairness I'm not seeing it with Google or IBM either. Much like the WWII battleships, systems like Alpha Zero will be obsolete soon. I'm also not seeing it on the hardware front. Intel is having to continue production at 14nm because they can't yet produce enough volume at 10nm. Kurzweil thought we would be down to 3.5 nm by now and looking at 1 nm soon. Instead it looks like we'll hit bottom at 10nm or maybe 7nm. And that's the end of it. I think there is a path that could get a few more generations, but I haven't seen anyone working in that direction.
What I do see almost everyday is that we are drowning in information. We can't handle what we have now, but we would need a lot more to increase efficiency. This seems to me where the major advancement will come. So, I suppose there could be some kind of war in the future, but I'm not seeing it starting yet.
Aucun commentaire:
Enregistrer un commentaire