Because the phrase 'language model' (or rather 'large language model', LLM) is not a post-hoc classification arrived at by some digital anthropologist examining a black box. It's a description of the tool that OpenAI set out (successfully!) to build. That you are ascribing additional properties to it is exactly the kind of thing I'm talking about - it's so convincing that it's tempting to think that it's reasoning beyond its capabilities, but it's not. Can you cite specific examples of things it's doing besides producing text? It's generally terrible at maths (as you would expect).
Without wishing to diminish the importance of this work (because it is genuinely incredible and useful in all kinds of ways), we still need to remember that under the hood it's really an elaborate parlour trick, a sort of reverse mechanical turk pretending to be a brain. More interesting I think is the question of how much of human intelligence is likewise this kind of statistical pattern matching; it seems to me increasingly that we're not as smart as we think we are.
You are missing what is so surprising about deep learning in general- it seems that training a sufficiently complex model on a sufficiently hard problem causes it to spontaneously develop intelligence, that generalizes to other problems. It doesn't really matter what problem you choose- image models, language models, etc. This is the "AI Scaling Hypothesis," basically the idea that advanced AI doesn't need much beyond greater scale. The GPT-4 paper itself is strong evidence of this.
In a similar manner to humans developing intelligence while being optimized just for reproductive success.
I'm not missing that, I'm explicitly disagreeing that GPT shows evidence of this, and pointing out that human observers are mistakenly ascribing generalised intelligence to it because of some undeniably impressive, but explainable, results. The paper itself even opens with "GPT-4 is a Transformer-based model pre-trained to predict the next token in a document". I don't see any evidence of spontaneous development of intelligence, although I do think this work helps us get towards a deeper understand of the nature of intelligence itself, since a lot of what appears to be intelligent about GPT's behaviour is actually just the combination of a statistical model and an abundance of data, and perhaps that applies to humans too.
Also I would point out that emergent general intelligence would actually be quite an unsurprising result of deep learning for many people, given what we know about the human brain plus some hand-waving about emergent systems - I think many people actually expect something like that to happen, and that's exactly why so many people are jumping to that conclusion about GPT. It's confirmation bias.
But please enlighten me - where is the evidence that GPT-4 has generalised intelligence?
Like another commenter pointed out, you are redefining intelligence out of existence for both humans and GPT-4, and I don't see what that point of that is. GPT-4 can solve subtle and complex problems it has never seen before: explaining the point of jokes, guessing correctly about the subtle hidden motives of characters in a story, playing games, writing code, analyzing logical arguments, etc. These responses are not in the training data in any meaningful way. The ability to do tasks like this - things that require prediction, mental modeling of subtle details, etc. is what I feel makes both humans and GPT-4 generally intelligent.
Without wishing to diminish the importance of this work (because it is genuinely incredible and useful in all kinds of ways), we still need to remember that under the hood it's really an elaborate parlour trick, a sort of reverse mechanical turk pretending to be a brain. More interesting I think is the question of how much of human intelligence is likewise this kind of statistical pattern matching; it seems to me increasingly that we're not as smart as we think we are.