Announcing GPT-4, a large multimodal model

openai.com

13 pointsmarban4 comments

Yay.

ML makes programming fun again. Waiting for model releases feels like waiting for your favorite band to drop a new album.

Being multimodal is a big step. So is hitting the 1T parameter milestone. I’d love to see what they deploy this beast on; 8xA100 is still only 8x80GB = 640GB.

Actually, I only heard rumors of it being 1T, so I wonder if openai has mentioned it anywhere…

EDIT: the GPT-4 paper is pretty fascinating: https://cdn.openai.com/papers/gpt-4.pdf

Scroll all the way to the bottom for some examples of problematic prompts that it’s trained to avoid.

Where are you getting the 1T parameter number? Is that from the leaks/speculation or is that an official number from OpenAI?

Twitter hearsay. And it looks like that's all we'll ever have, since the paper refuses to elaborate.

dupe:

https://news.ycombinator.com/item?id=35154527

Reply on news.ycombinator.com