https://mobile.twitter.com/ChrisMurphyCT/status/1640186536825061376
the original tweet is low IQ, but so is the reader added context:
ChatGPT is a system of averages.
It is a language model and only understand how to generate text. It can 'appear' to understand text in the same way that AI can 'appear' to create images. It is not actual learning.
"system of averages." what does that even mean? it's possible to link to countless examples where saying it doesn't understand the text seems completely implausible.
examples where saying it doesn't understand the text seems completely implausible.
there's no "it" there, it's a very complicated lookup table. it has no experience of the world, no goals, and no reflection on itself, because there is no sentience behind it. and at no point in scaling the algorithmic complexity, training data, etc do you get something akin to actual intelligence.
examples where saying it doesn't understand the text seems completely implausible.there's no "it" there, it's a very complicated lookup table. it has no experience of the world, no goals, and no reflection on itself, because there is no sentience behind it. and at no point in scaling the algorithmic complexity, training data, etc do you get something akin to actual intelligence.
Bingo!
One small point. It does have a goal -- predict text as well as "possible."
The real lesson from ChatGPT is that we haven't advanced on our ancestors. When the ancient Greeks heard thunder, they thought, "Who made that noise?" When we read an answer from a chat-bot, we think, "Who said that?"
It's the same mistake.
examples where saying it doesn't understand the text seems completely implausible.there's no "it" there, it's a very complicated lookup table. it has no experience of the world, no goals, and no reflection on itself, because there is no sentience behind it. and at no point in scaling the algorithmic complexity, training data, etc do you get something akin to actual intelligence.
Individual parts of the cortex don't have goals, at least in the way an organism does. The best way to think about the brain is as a hierarchical predictive model with a reward system. The feeling of being a self emerges as a model the brain uses to understand other parts of the brain, but there is no coherent agent driving the system and authoring language based thought. Doesn't seem unreasonable to think of GPT as an approximation of the language parts of the cortex, which are fundamentally predictive
you dont know how the brain works. Nobody does. shaddup with your ga/rbage.
examples where saying it doesn't understand the text seems completely implausible.there's no "it" there, it's a very complicated lookup table. it has no experience of the world, no goals, and no reflection on itself, because there is no sentience behind it. and at no point in scaling the algorithmic complexity, training data, etc do you get something akin to actual intelligence.
Individual parts of the cortex don't have goals, at least in the way an organism does. The best way to think about the brain is as a hierarchical predictive model with a reward system. The feeling of being a self emerges as a model the brain uses to understand other parts of the brain, but there is no coherent agent driving the system and authoring language based thought. Doesn't seem unreasonable to think of GPT as an approximation of the language parts of the cortex, which are fundamentally predictive
you dont know how the brain works. Nobody does. shaddup with your ga/rbage.
examples where saying it doesn't understand the text seems completely implausible.
there's no "it" there, it's a very complicated lookup table. it has no experience of the world, no goals, and no reflection on itself, because there is no sentience behind it. and at no point in scaling the algorithmic complexity, training data, etc do you get something akin to actual intelligence.
Individual parts of the cortex don't have goals, at least in the way an organism does. The best way to think about the brain is as a hierarchical predictive model with a reward system. The feeling of being a self emerges as a model the brain uses to understand other parts of the brain, but there is no coherent agent driving the system and authoring language based thought. Doesn't seem unreasonable to think of GPT as an approximation of the language parts of the cortex, which are fundamentally predictive
Predictive coding has a lot of evidence even if implementation details aren't understood yet. Just because we don't know wiring rules or how the brain does gradient updating doesn't mean we can't constrain the space of plausible architectures.