2 ms·
>> GPT-3 is the first AI system that has obvious, immediate, transformative economic value. To say the least, it is not immediately clear where that "transform
by YeGoblynQueenne 6y ago
>> GPT-3 is the first AI system that has obvious, immediate, transformative
economic value.
To say the least, it is not immediately clear where that "transformative
economic value" lies.
From what I've seen so far GPT-3 can generate structurally smooth but
completely incoherent text and despite claims to the contrary cannot perform
anything close to "reasoning" [1]. It can also perform some side-tasks like
machine translation and question answering, though with nowhere near good
enough accuracy for it to be used as a commercial solution for these tasks.
All this is not very useful or even interesting. Text generation is a fun
passtime but unless one can control the generation to very precise
specifications, to generate good quality text that makes sense on a particular
subject, text generation is nothing but a toy with no commercial value (and
even its scientific value is not very clear). And GPT-3's generation cannot be
controlled to such precise specifications.
We've had AI software that could interact intelligently with a user since the
1970's, with Terry Winograd's SHRDLU [2] and that never led to "immediate,
transformative economic value", even though it was every bit the sci-fi-like
AI program that could be directed by natural language to perform specific
tasks with competence, albeit in a restricted enviroment (a "blocks world").
GPT-3 is not even capable of doing anything like that (nor are any other
modern systems). How is a language model that is likely to respond with "blue
offerings to the green god of mad square frogs" to a request to "place the
blue pyramid on the red sphere" bring "transfomative" value?
In fact, we've had systems capable of generating much more coherent (and still
grammatically corret) text for some time [3] and even those have not caused a
dramatic upheaval of "transformative economic value".
I'm sorry but I'm afraid that, with GPT-3, we're again in a spiralling peak
of hype, just as we were a few years ago with all the claims about sef driving
cars "next year" etc. I think we all know how those panned out.
In any case, you don't have to take my word for it. As with self-driving cars,
all we have to do is wait a few years. Say, until 2024. We'll have a good idea
about GPT-3's "transformative value" by then.
__________________
[1] Unless of course one insists on Procrusteanising the definition of
"reasoning" sufficently to cover essentially random guessing.
[2] https://en.wikipedia.org/wiki/SHRDLU https://en.wikipedia.org/wiki/SHRDLU
[3] I'll need to dig up some references if you ask, but in the meantime search
for "story generation".