Read on Twitter
GPT-3 update: I got contacted by @gdb who sent me an invite! excited to try it out. thanks Greg!
initial attempts: very impressive QA results (check out the coref in the gates questions!) but also has some glitches.
ok so it's not magical.
(but we shouldn't expect it to be magical. and the OpenAI docs refer to it as a "pattern matcher", which i think is very fair)
not bad! (but also not perfect)
the 'one' and 'two' are the prompts.
the rest are the answers. notice the generalization to multiple digits, the generalization to addition, and the weird generalization in five thousands and (concat) 4.
the 'one' and 'two' are the prompts.
the rest are the answers. notice the generalization to multiple digits, the generalization to addition, and the weird generalization in five thousands and (concat) 4.
ok this pluralization test is kinda neat! at each line I stopped at "there are two" and it completed the rest (the "and now there are three" is its own addition)
a few more pluralizations (harder, and so it expectedly fails more):
nice! (the first two are my prompt)
some code stuff, the context sensitivity is very impressive.
(my prompt is up to lst[-1])
![some code stuff, the context sensitivity is very impressive.(my prompt is up to lst[-1])](https://pbs.twimg.com/media/EdLqzz_XYAUJOvU.jpg "some code stuff, the context sensitivity is very impressive.(my prompt is up to lst[-1])")
![some code stuff, the context sensitivity is very impressive.(my prompt is up to lst[-1])](https://pbs.twimg.com/media/EdLqzz-XYAAJF-F.jpg "some code stuff, the context sensitivity is very impressive.(my prompt is up to lst[-1])")
(my prompt is up to lst[-1])
ooops, the lst.length in the python versio are my fault, they are leftovers from the w/o python prompt. it does know how to do length in python (though botched the "at least"):
https://twitter.com/yoavgo/status/1284354358172438528?s=21 https://twitter.com/yoavgo/status/1284354358172438528
https://twitter.com/yoavgo/status/1284356545715806209?s=21 https://twitter.com/yoavgo/status/1284356545715806209
will do some more linguistic-y stuff when less tired. for now:
ok i guess i need to work on my prompt-writing skills.
"out of the box" language understanding: eh.
Simple passivzation: wow. (I gave it only the first Q/A pair as an example)
a few more passivizations, here I added "by me" at the end of the first donkey answer, and continued from there.
simple question formation: nice.
few more questions (here with a question target). not perfect but very impressive given the 1-shot nature.
noun compounds interpretation. it makes a mistake here and there, but, overall wow. (cc @VeredShwartz )
and, to save face, a few more... challenging ones.
s-v agreement from example (first two) fails with rc attractors
adding the two rc examples as prompts (now first four are prompts), improve things substantially
some more "code", here the prompt was "def factorial(a):"
obviously it has this memorized + spit out a few more recursive homework functions just in case. (i think the lack of indentation is a UI issue)
obviously it has this memorized + spit out a few more recursive homework functions just in case. (i think the lack of indentation is a UI issue)
here the prompt was:
"""
# non recursive implementation of factorial
def factorial(a):
"""
the result is.. interesting.
"""
# non recursive implementation of factorial
def factorial(a):
"""
the result is.. interesting.
