Build a human-level AI by compressing Wikipedia(prize.hutter1.net)
prize.hutter1.net
Build a human-level AI by compressing Wikipedia
http://prize.hutter1.net/
4 comments
Well the title is rather misleading; I don't think even they think they will produce any AI out of this.
I think the reasoning goes more like: 1. Intelligence is good at representing information succinctly. 2. This is similar to what a data compression algorithm does. 3. Maybe if we come up with some uber-powerful compression algorithms, it will teach us something about intelligence too.
There's actually been a lot of papers about the parallels between data compression and AI. Look up the program "MegaHal" and papers on it by Jason Hutchens. He shows how a Markov model can be thought of as a kind of data compression.
That being said, there are theoretical limits on data compression efficiency, as shown by work by Shannon and Huffman. These apply to lossless compression, so at best I think all they would be doing is experimentally verifying the theoretical limits. If they want to find a data compression algorithm that tells them something about intelligence, it's going to have to be a lossy compression algorithm, not lossless.
I think the reasoning goes more like: 1. Intelligence is good at representing information succinctly. 2. This is similar to what a data compression algorithm does. 3. Maybe if we come up with some uber-powerful compression algorithms, it will teach us something about intelligence too.
There's actually been a lot of papers about the parallels between data compression and AI. Look up the program "MegaHal" and papers on it by Jason Hutchens. He shows how a Markov model can be thought of as a kind of data compression.
That being said, there are theoretical limits on data compression efficiency, as shown by work by Shannon and Huffman. These apply to lossless compression, so at best I think all they would be doing is experimentally verifying the theoretical limits. If they want to find a data compression algorithm that tells them something about intelligence, it's going to have to be a lossy compression algorithm, not lossless.
Three problems:
1) Wikipedia has far more information than a normal human, thus it would surpass human-level AI.
2) Compressing wikipedia into one persons mind would create a schizophrenic AI that's likely to kill people or die in a fit as its brain has an edit war.
3) It would take a human-level AI to first understand the information and then discard the useless crap and junk people post on wikipedia. I mean can you imagine asking an android what they think about Pamela Anderson's next divorce and having them randomly say "teh big boobs WOOT!1!11!"
Besides, information is redundant in intelligence. A cro-magnon would understand particle physics just as good as any one of us if they knew the information because they're intelligent.
1) Wikipedia has far more information than a normal human, thus it would surpass human-level AI.
2) Compressing wikipedia into one persons mind would create a schizophrenic AI that's likely to kill people or die in a fit as its brain has an edit war.
3) It would take a human-level AI to first understand the information and then discard the useless crap and junk people post on wikipedia. I mean can you imagine asking an android what they think about Pamela Anderson's next divorce and having them randomly say "teh big boobs WOOT!1!11!"
Besides, information is redundant in intelligence. A cro-magnon would understand particle physics just as good as any one of us if they knew the information because they're intelligent.
Hm, but when I compress wikipedia, it isn't lossless...
If history is any guide, though, a lot of trying and failing like this is necessary before we figure out what's important to work on. I still think they should instad use their 50000€ on paying food and rent for a young, inspired researcher in the field to try out as many crazy ideas as possible.