Intelligence is more than compression | François Chollet and Lex Fridman

[@daniele81]

What he is saying in his example is that compression in the brain doesn’t loose the ability to generalise. In other words, the compression doesn’t get to the point of overfitting the training set. He just demonstrated that cognition is equivalent to compression as long as generalisation ability is preserved.

[@jjgerald7877]

"Intelligence is compression" only means that it is smart to fully describe or output the true information of the finite data source (or file) in the smallest possible number of bits. This is called entropy or also called popularly as Shannon Entropy. "Minimum description length" to others. If you can achieve it, you're very intelligent so to say.

[@JMS1089]

3:32 true, but information is structured to "make" it appear important. at least that's how it feels when you read.

[@vincebartle]

Isn't exploration a compressed feature itself though. Also doesn't exploration itself carry compression: i.e. what exploration model do you use and how is it determined? He's saying intelligence = compression + exploration; but he doesn't explain how exploration is strictly exclusive of compression and thus a necessary addition to the definition of intelligence. It seems he's trying to expose how compression can be an over-deterministic model, which doesn't seem true, it's just that ineffective compression can have over-determination as a trait. It might be better to say that lossless compression requires some exploration as a feature.

[@MrJorgeceja123]

schmidhuber pilled

[@The_Conspiracy_Analyst]

Algorithmic based compression is useless in context based knowledge systems.

[@CHURCHISAWESUM]

I mean not if it’s the best possible algorithm Algorithm just means step by step list of instructions. If you didn’t skip steps, there’s no reason theoretically why the steps don’t contain context.

[@tubbybrown1918]

wb lossless