Live data from Hacker News

Ask HN: Can I download GPT / ChatGPT to my desktop?

news.ycombinator.com

1–10 of 51 posts

Re: Ask HN: Can I download GPT / ChatGPT to my desktop?

#3
OpenAI is 'open' in the name only, so no. I don't think they have any plans of opening full access to the public either, considering that their previous model (that ChatGPT builds upon) was sold to Microsoft for exclusive use:

https://en.wikipedia.org/wiki/GPT-3

Re: Ask HN: Can I download GPT / ChatGPT to my desktop?

#6
post #5

It’s not possible currently but there’s another story on the HN front page with an open source alternative which I haven’t tried

Which one?

Possibly this "PaLM + RLHF - Pytorch" project:

https://github.com/lucidrains/PaLM-rlhf-pytorch

Re: Ask HN: Can I download GPT / ChatGPT to my desktop?

#7
I'm pretty sure the GPT model is huge and does not fit on any conventional GPU. Even if they open-sourced the weights, I don't think most people would be running it at home.

Also regarding the text limits, AFAIK, there's just an inherent limit in the architecture. Transformers are trained on finite-length sequences (I think their latest uses 4096 tokens). I have been trying to understand how ChatGPT seems to be able to manage context/understanding beyond this window length

Re: Ask HN: Can I download GPT / ChatGPT to my desktop?

#9
Even if it were freely available, there's no way to run GPT3 or ChatGPT on any existing desktop hardware. The exact hardware requirements aren't public either (yes, very "open") but a full 175-billion-parameter GPT3 instance requires hundreds of gigabytes of GPU memory, and even though ChatGPT is "smaller and better", when it comes to conversational dialogue, there's no way to fit it in current consumer hardware.

Re: Ask HN: Can I download GPT / ChatGPT to my desktop?

#10

I'm pretty sure the GPT model is huge and does not fit on any conventional GPU. Even if they open-sourced the weights, I don't think most people would be running it at home. Also regarding the text limits, AFAIK, there's just an inherent limit in the architecture. Transformers are trained on finite-length sequences (I think their latest uses 4096 tokens). I have been trying to understand how ChatGPT seems to be able…

I don't think ChatGPT does. I have had long discussions with it, with some rules agreed upon in the beginning, and at some point it clearly begins to forget the exact rules and has to be reminded of them.

(Specifically, AI Dungeon type games where ChatGPT is the DM and the human the protagonist, or vice versa. The most common failure mode seems to be that it forgets whether it's playing the DM or the protagonist. To be fair, it performs admirably well despite the limitations.)

Post reply on HN