and then the matching idea to that of artificial intelligence with the hutter prize you know i'm torn i go back and forth on what i think about it yeah but the the the basic thesis is it's nice it's a nice compelling notion that we can reduce the task of creating an intelligent system a general intelligence system into the task of compression so you can think of all of intelligence in the universe in fact as a kind of compression do you find that was that just at the time you found that as a compelling idea do
you still find that a compelling idea i still find that compelling idea um i think that it's not that useful day-to-day but actually um one of maybe my quests before that was a search for the definition of the word intelligence and i never had one and i definitely have a definition of the word compression it's a very uh simple uh straightforward one and uh you know what confession is you know what lossless is lossless compassion not lossy lossless compression and that that is equivalent to intelligence which i believe i'm not sure how useful that definition
is day-to-day but like i now have a framework to understand what it is and he just 10xed the uh the prize for that competition like recently a few months ago you ever thought of taking a crack at that oh i did oh i did i spent i spent the next after i found the prize i spent the next six months of my life trying it and uh well that's when i started learning everything about ai and then i worked vicarious for a bit and then i learned read all the deep learning stuff and i'm like
okay now i like i'm caught up to modern ai wow and i had i had a really good framework to put it all in from the compression stuff right like some of the first uh some of the first deep learning models i played with were uh gpt basically but before transformers before it was still uh rnn's to to do character prediction but by the way on the compression side i mean the especially neural networks what do you make of the lossless requirement with the hutter prize so you know human intelligence and and neural networks can
probably compress stuff pretty well but there it will be lossy it's imperfect uh you can turn a lossy compressor into a lossless compressor pretty easily using an arithmetic encoder right you can take an arithmetic encoder and you can just encode the noise with maximum efficiency right so even if you can't predict exactly what the next character is the better a probability distribution you can put over the next character you can then use an arithmetic encoder to uh right you don't have to know whether it's an e or an i you just have to put good
probabilities on them and then you know code those and if you have it's a bits of entropy thing right you