Bakz T. Future
Could GPT-3 be used for basic NLP tasks like tokenization? I made a video to find out!
—
► Remember to Like, Comment, and Subscribe!
—
Connect with me:
Twitter – http://www.twitter.com/bakztfuture
Instagram – http://www.instagram.com/bakztfuture
Github – http://www.github.com/bakztfuture
Feel free to send me an email or just say hello:
bakztfuture@gmail.com
edited by: Nino Bedoshvili
Now I'm curious how GPT would act if it was trained on tokenized sentences instead of characters …
Hello! Thank you so much for your videos, I enjoy them thoroughly. I'm still just sitting here waiting for my GPT-3 API key, was hoping to ask you some questions.
How long does an actual GPT-3 API call take? And how much does it cost right now, per call or per month.
Ask GPT-3 if you don't know the answer.
OoOoOOo neat! Thanks for sharing!
This is crazy
Holy shit! GPT-3 is almost a general purpose AI! I'm not even joking when I say a theoretical GPT-4 or 5 could be human level intelligence.
What program are you using for this presentation? ie. 0:22, and onward from 5:33
woa this is cool
5:40 what? You're basically using an API connected to model trained on terabytes of data and you're saying that this is crazy that you don't have to download libraries and datasets?
What would happen if you gave GPT-3 zero examples and just asked it to tokenize?
Maybe it already knows enough to do some kind of tokenization without any hints.
I do not think GPT-3 "understands" anything. It is just mechanically acting the way it was created to act.
One issue not addressed here is reliability. It won't always return the same result for the same query – there's a randomness factor, which makes it largely unacceptable for cases like this. Besides the obvious, the performance issue – which, no, optimizations won't solve that, not by a long shot. That said, ask it to generate code or regular expressions (both have been done) and you might get something reliable and performant that you could unit test.
This is excellent.
One caveat: no way GPT could be faster than a specialized algorithm, ever. It's probably not what you meant at the end of the video