For me, the way that chatbots have only one way of thinking available to them that makes them fundamentally unintelligent.
At the moment.
Some of the newer models seem to be able to look at a prompt, realise that they are incapable of solving what's required with their own resources & then reach out for the code they need to solve it - which is quite clever. That's how they "solve" puzzles like sudoku - but they haven't really solved them, they've just "paid" someone else to do it for them.
And they still have no way of noticing when they've got something wrong.
It's partly what Nora Bateson calls "context" - as you say, they can't understand that one statement can mean 2 different things, often depending on the context. Chatbots only have one context, and that one is limited to what they've been fed, and even if they could find new stuff, it would require a massive amount of processing to incorporate it into their models. They're forever looking in the rear-view mirror, and if something really novel appears in their world, they have no way of adapting to it without a complete rebuild.
It's also that inability to switch thinking patterns on the fly that means they can't solve sudoku or spot mistakes. It may be that AlphaGo, the machine that plays Go, can do this - but that's a hugely specialised model, very different to the standard LLMs. It can't do anything much but play go.
I suspect the AI bubble is going to burst - there's no way it can turn a profit on that much investment. And they've only needed that huge investment because the techbros assume that if they make the LLMs big enough, they'll become like a generalised AlphaGo. As Cory Doctorow says, just because you can breed racehorses to run faster & faster, doesn't mean one of them is going to give birth to a locomotive one day. It's a logical typing error.
For me, the way that chatbots have only one way of thinking available to them that makes them fundamentally unintelligent.
At the moment.
Some of the newer models seem to be able to look at a prompt, realise that they are incapable of solving what's required with their own resources & then reach out for the code they need to solve it - which is quite clever. That's how they "solve" puzzles like sudoku - but they haven't really solved them, they've just "paid" someone else to do it for them.
And they still have no way of noticing when they've got something wrong.
It's partly what Nora Bateson calls "context" - as you say, they can't understand that one statement can mean 2 different things, often depending on the context. Chatbots only have one context, and that one is limited to what they've been fed, and even if they could find new stuff, it would require a massive amount of processing to incorporate it into their models. They're forever looking in the rear-view mirror, and if something really novel appears in their world, they have no way of adapting to it without a complete rebuild.
It's also that inability to switch thinking patterns on the fly that means they can't solve sudoku or spot mistakes. It may be that AlphaGo, the machine that plays Go, can do this - but that's a hugely specialised model, very different to the standard LLMs. It can't do anything much but play go.
I suspect the AI bubble is going to burst - there's no way it can turn a profit on that much investment. And they've only needed that huge investment because the techbros assume that if they make the LLMs big enough, they'll become like a generalised AlphaGo. As Cory Doctorow says, just because you can breed racehorses to run faster & faster, doesn't mean one of them is going to give birth to a locomotive one day. It's a logical typing error.
Great response. So much to think about. Thanks