They are very impressive to where we were 20 years ago, hell even 5 years ago. The first time I played with ChatGPT I was absolutely floored. But after playing with a lot of them, even training a few RAGs (Retrieval-Augmented Generation), we aren’t really that close and in my opinion this is not a useful path towards a true AGI. Don’t get me wrong, this tool is extremely useful and to most people, they’d likely pass a basic Turing Test. But LLMs are sophisticated pattern recognition systems trained on vast amounts of text data that predict the most likely next word or token in a sequence. That’s really all they do. They are really good at predicting the next word. While they demonstrate impressive language capabilities, they lack several fundamental components necessary for an AGI:
-no true understanding
-they can’t really engage in the real world.
-they have no real ability to learn real-time.
-they don’t really have the ability to take in more then one type of info at a time.
I mean the simplest way in my opinion to explain the difference is you will never have an LLM just come up with something on its own. It’s always just a response to a prompt.
Sorry for the late reply - work is consuming everything :)
I suspect that we are (like LLMs) mostly “sophisticated pattern recognition systems trained on vast amounts of data.”
Considering the claim that LLMs have “no true understanding”, I think there isn’t a definition of “true understanding” that would cleanly separate humans and LLMs. It seems clear that LLMs are able to extract the information contained within language, and use that information to answer questions and inform decisions (with adequately tooled agents). I think that acquiring and using information is what’s relevant, and that’s solved.
Engaging with the real world is mostly a matter of tooling. Real-time learning and more comprehensive multi-modal architectures are just iterations on current systems.
I think it’s quite relevant that the Turing Test has essentially been passed by machines. It’s our instinct to gatekeep intellect, moving the goalposts as they’re passed in order to affirm our relevance and worth, but LLMs have our intellectual essence, and will continue to improve rapidly while we stagnate.
There is still progress to be made before we’re obsolete, but I think it will be just a few years, and then it’s just a question of cost efficiency.
Anyways, we’ll see! Thanks for the thoughtful reply
They are very impressive to where we were 20 years ago, hell even 5 years ago. The first time I played with ChatGPT I was absolutely floored. But after playing with a lot of them, even training a few RAGs (Retrieval-Augmented Generation), we aren’t really that close and in my opinion this is not a useful path towards a true AGI. Don’t get me wrong, this tool is extremely useful and to most people, they’d likely pass a basic Turing Test. But LLMs are sophisticated pattern recognition systems trained on vast amounts of text data that predict the most likely next word or token in a sequence. That’s really all they do. They are really good at predicting the next word. While they demonstrate impressive language capabilities, they lack several fundamental components necessary for an AGI: -no true understanding -they can’t really engage in the real world. -they have no real ability to learn real-time. -they don’t really have the ability to take in more then one type of info at a time.
I mean the simplest way in my opinion to explain the difference is you will never have an LLM just come up with something on its own. It’s always just a response to a prompt.
Sorry for the late reply - work is consuming everything :)
I suspect that we are (like LLMs) mostly “sophisticated pattern recognition systems trained on vast amounts of data.”
Considering the claim that LLMs have “no true understanding”, I think there isn’t a definition of “true understanding” that would cleanly separate humans and LLMs. It seems clear that LLMs are able to extract the information contained within language, and use that information to answer questions and inform decisions (with adequately tooled agents). I think that acquiring and using information is what’s relevant, and that’s solved.
Engaging with the real world is mostly a matter of tooling. Real-time learning and more comprehensive multi-modal architectures are just iterations on current systems.
I think it’s quite relevant that the Turing Test has essentially been passed by machines. It’s our instinct to gatekeep intellect, moving the goalposts as they’re passed in order to affirm our relevance and worth, but LLMs have our intellectual essence, and will continue to improve rapidly while we stagnate.
There is still progress to be made before we’re obsolete, but I think it will be just a few years, and then it’s just a question of cost efficiency.
Anyways, we’ll see! Thanks for the thoughtful reply