You can approximate a human capacity, say theory-of-mind, with another kind of ape: play some hide-and-seek game. You can approximate the knowledge of a trivia-master with a child and a trivia book.
These are quite different sorts of approximations. A 1/100th scale bridge build to stand for a real one is quite different than taking some prior set of bridges, measuring them, and deriving some merely associative model of their properties.
My issue in how LLMs (etc.) are popularly understood is that people think they approximate target capacities by being 'ontologically similar' capacities -- and this is really very dangerous. You'll lose your job if you think so (and so on).
And every greasy AI-board hocking these to the public is very much pushing out this noxious mumbojumbo.
It matters greately how an approximation works, and why any given output arrives from any given input.
I'd only add that certainty requirements for real-world target applications can differ substantially. I.e. engineering vs art.
A toy box that gives magic answers 85% of the time is incredibly useful in some scenarios -- e.g. seeding the beginning of a manual research process with initial topics.
Which seems a general rule of thumb for deploying GenAI into production these days: find use cases where there are minimal consequences to being occasionally wrong.
In my head, that's the Netflix recommendation test. What's the impact if Netflix gives me a bad recommendation? Consequently, that's how aggressive they can be with their models.
You can't predict almost anything anyway. The relevant distribution for acting is a utility+risk distribution over various sets of predictions.
If you compute that, most of ML/AI isnt very useful for most of anyone. It's kinda interesting that generative AI finally achieved something here, given our 'distributions for action'.
Of course the popular conversation isnt there yet, people still think these things work. When more people find their utility hit by their failures, i think we'll arrive back to status-quo-ante-openai where people turn off siri.
Nevertheless -- there is an achivement here (finicial in paying for training; and legal in whitewashing copyright away) -- which will have some impact
Here is a pastebin: https://pastebin.com/DCqgxx8E
They 100% are not those things... but they also approximate them well-enough to be functionally useful.
I.e. the high-dimensional curve-fitting / compression conceptualization of ML, which intuitively expresses both its strengths and weaknesses.
If "it" is represented in the data set (explicitly or implicitly), the "curve" will fit to that property.
Simultaneously, the "curve" is approximating and smoothing out disjoint data steps to pack high-fidelity data features into a more space-efficient model. Hence some features disappear, others are tortured beyond intuitive correspondence, and others become linked to non-obvious proxies. But some strongly-expressed ones remain.
It's fascinating but not surprising that responsiveness to emotion is encoded in model weights, given that all conversational training data had emotional impetus, given that it came from humans.