71–80 of 84 posts
Was it trained on IRC? This is what I got: is your job going to be nonexistant for 12 years? hi all someone please help me with this, I have windows 7 on a usb and when i boot from it my ubuntu 10.10 iso does not start. when i go to my bootloader in my usb and select the install option i get a grey screen with a green bar going across and it never loads. i have the i386 option. help. thanks When I boot my computer af…
Is it possible to run this on something other than google's cloud platform?
It's design is pretty specific to Google's TPU. If you ran it elsewhere, you'd likely need to tweak the design or suffer quite a big performance penalty.
Still upset about the pile as a dataset (the one used to train this model) because I wrote an issue asking if they would include my dataset debatesum into the dataset[1]. They mentioned that they planned to do it and even wrote another issue indicating that they would include it in the pile. Months go by, and I ask why this dataset is still not included. They told me that the pile is complete and will not be updated…
Hi Kindiana! If you happen to see this, I sent a message on Keybase. Nothing important, just saying hello. This is absolutely wonderful work, and I'm really happy to see you've released it. Congrats!! It's also awesome that Eleuther is hosting an API for it. Nice work, Elecrew. No idea who did that part, but it's pretty cool. It's easy to underestimate how hard something like that can be. Good luck with your future r…
Earlier quoted context omitted.
This is simply incredible, I would love to read something novel-length like this. But unfortunately it also generates a lot of banal text, as in some of the other examples in this thread. I wonder if there is a way to optimize for the weird shit.
Convince someone with a really big pile of TPU credits to train a larger-size model on the full text of the SCP Wiki :> This could even potentially be interesting enough to attract (small-scale) Kickstarter-level attention.
I took one of the questions from this article: https://blog.seekwell.io/gpt3 Instruction: Given an input question, respond with syntactically correct PostgreSQL. Be creative but the SQL must be correct. Input: how many users signed up in the past month? And it gave me a pretty nonsensical answer compared to the output from GPT-3 seen in that article.
I haven’t properly investigated their dataset they trained on this, but from the responses it gives I would bet that they dataset is skewed towards news articles, press releases, and infomercial or edutainment copy. It does really well when it can write a long impersonal article about something, but doesn’t seem good at the random sort of things GPT3 can do like autocompleting source code.
Earlier quoted context omitted.
It's design is pretty specific to Google's TPU. If you ran it elsewhere, you'd likely need to tweak the design or suffer quite a big performance penalty.
Nope, probably not. It's using JAX, which works on both GPUs and TPUs.
I get the feeling that now that it's so easy to generate text that sounds legit and is syntactically correct but makes no sense semantically, web search will be unusable soon. Searching for answers to an obscure problem with a software library now already leads to 5 Stack Overflow ripoff sites that at least contain the same text like the original. If these scammers can now generate their own texts (randomly including…