I do believe if we are going to get AGI without some random revolutionary breakthrough, to achieve it iteratively, It's going to come through language models.
Think about it.
What's the most expressive medium we have which is also absolutely inundated with data?
To broadly be able to predict human speech you need to broadly be able to predict the human mind. To broadly predict a human mind requires you build a model of it, and to have a model of a human mind? Welcome to general intelligence.
We won't realize we've created an AGI until someone makes a text model, starts throwing random problems at it, and discovers that it's able to solve them.
> I do believe if we are going to get AGI without some random revolutionary breakthrough, to achieve it iteratively, It's going to come through language models.
Language is way, way far removed from intelligence. This is well-known in cognitive psychology. You'll find plenty of examples of stroke victims who are still intelligent but have lost the ability to produce coherent sentences, and (though much rarer) examples of people who can produce clear, eloquent prose, yet are so learning and mentally challenged that they can't even tell the difference between fantasy and reality.
We don't judge AI by their ability to produce language, we judge them by their conference and ability to respond intelligently, to give us information we can use.
> To broadly be able to predict human speech you need to broadly be able to predict the human mind
This is a non sequitur. The human mind does a whole lot more than string words together. Being able to predict which word would logically follow another does not require the ability to predict anything other than just that.
I think what the commenter is saying is that, in time, language models too will do a lot more than string words together. If it's large enough, and you train it well enough to respond to “what's the best next move in this chess position?” prompts with good moves, it will inevitably learn chess.
Exactly. Since language is a compressed and transmittable result of our thought, to predict text as accurately as possible requires you do the same. A model with understanding of the human mind will outperform one without.
> Being able to predict which word would logically follow another does not require the ability to predict anything other than just that.
Why? Wouldn't you expect that technique to generally fail if it isn't intelligent enough to know what's happening in the sentence?
Perhaps a more interesting question is "how much better do we understand what characteristics AGI will have due to ChatGPT?"
We don't really understand what intelligence means -- in humans or our creations -- but ChatGPT gives us a little more insight (just like ELIZA, and the psychological research behind it, did).
At the very least, ChatGPT helps us build increasingly better Turing tests.
Yes. It is obviously already weak AGI (obvious to anyone if they saw it 20 years ago).
It is also obvious that we are in the middle of a shift of some kind. Very hard to see from within, but clearly we will look back at 2022 as the beginning of something…
I do believe if we are going to get AGI without some random revolutionary breakthrough, to achieve it iteratively, It's going to come through language models.
Think about it.
What's the most expressive medium we have which is also absolutely inundated with data?
To broadly be able to predict human speech you need to broadly be able to predict the human mind. To broadly predict a human mind requires you build a model of it, and to have a model of a human mind? Welcome to general intelligence.
We won't realize we've created an AGI until someone makes a text model, starts throwing random problems at it, and discovers that it's able to solve them.
> I do believe if we are going to get AGI without some random revolutionary breakthrough, to achieve it iteratively, It's going to come through language models.
Language is way, way far removed from intelligence. This is well-known in cognitive psychology. You'll find plenty of examples of stroke victims who are still intelligent but have lost the ability to produce coherent sentences, and (though much rarer) examples of people who can produce clear, eloquent prose, yet are so learning and mentally challenged that they can't even tell the difference between fantasy and reality.
We don't judge AI by their ability to produce language, we judge them by their conference and ability to respond intelligently, to give us information we can use.
> To broadly be able to predict human speech you need to broadly be able to predict the human mind
This is a non sequitur. The human mind does a whole lot more than string words together. Being able to predict which word would logically follow another does not require the ability to predict anything other than just that.
I think what the commenter is saying is that, in time, language models too will do a lot more than string words together. If it's large enough, and you train it well enough to respond to “what's the best next move in this chess position?” prompts with good moves, it will inevitably learn chess.
2 replies →
Exactly. Since language is a compressed and transmittable result of our thought, to predict text as accurately as possible requires you do the same. A model with understanding of the human mind will outperform one without.
> Being able to predict which word would logically follow another does not require the ability to predict anything other than just that.
Why? Wouldn't you expect that technique to generally fail if it isn't intelligent enough to know what's happening in the sentence?
"The ability to speak does not make you intelligent." — Qui-Gon Jinn, The Phantom Menace.
Perhaps a more interesting question is "how much better do we understand what characteristics AGI will have due to ChatGPT?"
We don't really understand what intelligence means -- in humans or our creations -- but ChatGPT gives us a little more insight (just like ELIZA, and the psychological research behind it, did).
At the very least, ChatGPT helps us build increasingly better Turing tests.
Why the obsession with AGI? The point is that ChatGPT is already useful.
Is it? I see it mostly generates BS much faster.
Brothers Grimm would like a word with you about what "BS" means.
ChatGPT is good at making up stories.
Yes. It is obviously already weak AGI (obvious to anyone if they saw it 20 years ago).
It is also obvious that we are in the middle of a shift of some kind. Very hard to see from within, but clearly we will look back at 2022 as the beginning of something…