Large Language Models are nothing but very advanced regurgitation machines. That's the AI these articles are hand wringing about - not a real Artificial Intelligence.
These articles remind me of the bitcoin articles we used to see.
What ability do you think that they are currently missing that makes them 'regurgitation machines' rather than just limited and dumb but genuine early AI?
Discreet object recognition. Right now they can't answer the simplest questions that require counting discrete objects. Which to me implies they have no discreet object sense at all. They're just looking for word patterns.
I'll give you an example, If you were to ask one of these.
"I was on my way to the store when I saw a sow with six piglets, how many feet do we have?"
We could have a lively debate about what potential answers would be acceptable. Maybe there's only 2 feet because the rest are hooves, maybe there's 30 because that's how many foot like appendages there are in total, but the answer it will give you will make absolutely no sense.
Chat GPT will be like, "11" or "15" and if you ask it any questions or follow up it genuinely does not have any answer for how any of these objects could be discreetly counted or partitioned. It can try to explain itself but quickly starts babbling nonsense.
I think that might be a chatgpt specific thing, I tried with bing in precise mode and it responded with this:
"A sow is an adult female pig and piglets are baby pigs. Pigs have four feet, so a sow with six piglets would have a total of 28 feet (4 feet for the sow + 6 piglets * 4 feet each). Is that what you were asking?"