> My point is that perfect word prediction does require intelligence and therefore being a word predictor does not, on its own, rule out being intelligent.
No, it doesn't. It requires a large dataset of observed documents to use as training data and the resulting neural network of nodes can then, fairly reliably, function as a word predictor as long as the word you want it to predict is a word that would be commonly found within the documents. Basically, if you wanted to create an LLM that would write, for example, house assessments, you could train it on millions of house assessments as written by human house assessors, and be reasonably confident that you could pull that off.
However, what you now have is a neural network that can reliably generate a document that would pass probably a fair number of glances from people who see them regularly as that document, but that doesn't accomplish anything. It can't assess a house, for example, which is the purpose of that document. If you generated an assessment with this LLM and presented it as it applied to your house, even if after a number of attempts you got the bedroom and bathroom count correct, it would likely make references to features your home lacks, get technical details wrong like the electrical service it has, or even make references to faults the house doesn't actually have, or worse still, fail to take into account ones it does.
That is intelligence, that is what ChatGPT does not and will never have, and that's why this technology is already hitting a wall. It doesn't do anything. It can make reams and reams of bullshit for you (and in our current sad state of the Internet, that's a surprisingly appealing technology to many!) but that's fundamentally just not that valuable as a technology.
I argued that a perfect word predictor would need to be intelligent or have infinite memory, and I noted that ChatGPT is not such a thing. The point was that establishing that ChatGPT is a word predictor is insufficient to disprove that it is intelligent. Your argument is that a fairly reliable word predictor does not need to be intelligent, which I agree with emphatically- a thing being a word predictor most certainly does not prove that it is intelligent. A perfect word predictor would either need to know or deduce properties of your house in order to write an accurate assessment; a fairly reliable one could just fall back on 'fairly' and fail the task.
I don't think your criteria for intelligence is sufficient- I would not be at all surprised to learn that GPT-4o could already look at pictures of my house and write an accurate assessment, but that wouldn't convince me it was intelligent. You could do this quite well a decade ago with computer vision and a fill-in-the-blanks document.
>this technology is already hitting a wall
An aside: I've seen this said a lot and I don't get it. GPT-3.5 is only about 2 years old and turned a toy into a useful tool. GPT-4 was a substantial improvement to output quality and context length and multimodality a half year later. If GPT-5 comes out and it's not a significant improvement or doesn't come out at all by March, that would be evidence that a wall has been hit. But at the moment I can't think of a technology that has improved more in the last 2 years and I don't know where this claim comes from.
No, it doesn't. It requires a large dataset of observed documents to use as training data and the resulting neural network of nodes can then, fairly reliably, function as a word predictor as long as the word you want it to predict is a word that would be commonly found within the documents. Basically, if you wanted to create an LLM that would write, for example, house assessments, you could train it on millions of house assessments as written by human house assessors, and be reasonably confident that you could pull that off.
However, what you now have is a neural network that can reliably generate a document that would pass probably a fair number of glances from people who see them regularly as that document, but that doesn't accomplish anything. It can't assess a house, for example, which is the purpose of that document. If you generated an assessment with this LLM and presented it as it applied to your house, even if after a number of attempts you got the bedroom and bathroom count correct, it would likely make references to features your home lacks, get technical details wrong like the electrical service it has, or even make references to faults the house doesn't actually have, or worse still, fail to take into account ones it does.
That is intelligence, that is what ChatGPT does not and will never have, and that's why this technology is already hitting a wall. It doesn't do anything. It can make reams and reams of bullshit for you (and in our current sad state of the Internet, that's a surprisingly appealing technology to many!) but that's fundamentally just not that valuable as a technology.