Yup.
73,
Dave AB7E
On 4/5/2023 4:20 PM, Christian Kennedy wrote:
It's probably useful to recall what the "GPT" in ChatGPT stands for --
Generative Pretrained Transformer; GPTs are a particularly effective
form of long-term memory that's useful for making predictions
regarding patterns. ChatGPT has zero semantic knowledge; in essence
it's a text prediction engine that's been trained on a massive corpus
of text (apparently including code scraped off of github); what it
pukes out in response to your input text is, literally the
statistically most probable outcome based on the sequence of tokens
that it produced by shredding your input text. Of course if that's
all it did the results would be both dry and deterministic, so GPTs
incorporate the notion of "temperature"; at any given point in
constructing an output string it has a list of candidate words based
on what it's already output, ranked by probability, but then it
perturbs the probability so it doesn't always produce the same text.
Because each output token affects the probability of what comes next,
the result is different (occasionally very different) output strings
for the same input.
GPTs can be entertaining, and for narrow domains can be useful, but as
autocomplete systems with zero understanding of what they're spewing,
they can also be thought of as incredible bullshit generators. Worse,
their predictions become part of their corpus, and competing GPT
systems have started scraping each other, resulting in a sort of
exponential bullshit amplification phenomena that has required some
tamping down.
My suggesting is that you use it to get into the game, but always go
back to primary sources and do your own work.
_______________________________________________
_______________________________________________
TowerTalk mailing list
TowerTalk@contesting.com
http://lists.contesting.com/mailman/listinfo/towertalk
|