this post was submitted on 18 Jul 2024
303 points (97.8% liked)

Memes

45729 readers
686 users here now

Rules:

  1. Be civil and nice.
  2. Try not to excessively repost, as a rule of thumb, wait at least 2 months to do it if you have to.

founded 5 years ago
MODERATORS
 
you are viewing a single comment's thread
view the rest of the comments
[–] [email protected] 35 points 4 months ago* (last edited 4 months ago) (6 children)

Technically possible with a small enough model to work from. It's going to be pretty shit, but "working".

Now, if we were to go further down in scale, I'm curious how/if a 700MB CD version would work.

Or how many 1.44MB floppies you would need for the actual program and smallest viable model.

[–] [email protected] 13 points 4 months ago (3 children)

Might be a dvd. 70b ollama llm is like 1.5GB. So you could save many models on one dvd.

[–] [email protected] 7 points 4 months ago (3 children)

70b model taking 1.5GB? So 0.02 bit per parameter?

Are you sure you're not thinking of a heavily quantised and compressed 7b model or something? Ollama llama3 70b is 40GB from what i can find, that's a lot of DVDs

[–] [email protected] 8 points 4 months ago

Ah yes probably the Smaler version, your right. Still, a very good llm better than gpt 3

[–] [email protected] 7 points 4 months ago

Less than half of a BDXL though! The dream still breathes

[–] [email protected] 2 points 4 months ago

For some reason, triple layer writable blu-ray exists. 100GB each

https://www.verbatim.com/prod/optical-media/blu-ray/bd-r-xl-tl/bd-r-xl-tl/

[–] [email protected] 6 points 4 months ago

It is a DVD, can faintly see DVD+R on the left side

[–] [email protected] 4 points 4 months ago

It does have the label DVD-R

[–] [email protected] 10 points 4 months ago (1 children)

squints

That says , "PHILLIPS DVD+R"

So we're looking at a 4.7GB model, or just a hair under the tiniest, most incredibly optimized implementation of <INSERT_MODEL_NAME_HERE>

[–] [email protected] 5 points 4 months ago

llama 3 8b, phi 3 mini, Mistral, moondream 2, neural chat, starling, code llama, llama 2 uncensored, and llava would fit.

[–] [email protected] 7 points 4 months ago* (last edited 4 months ago)
[–] [email protected] 6 points 4 months ago

ELIZA was pretty impressive for the 1960s, as a chatbot for psychology.

[–] [email protected] 4 points 4 months ago

yes i guess it would be a funny experiment for just a local model

[–] [email protected] 3 points 4 months ago

pkzip c:\chatgpt*.* a:\chatgpt.zip -&