|
|
Log in / Subscribe / Register

ChatGPT says...

ChatGPT says...

Posted Oct 26, 2024 0:01 UTC (Sat) by rettichschnidi (subscriber, #93261)
In reply to: ChatGPT says... by leromarinvit
Parent article: OSI readies controversial Open AI definition

> Of course, then there's still the practical matter of actually being able to run the (full) training. This seems thoroughly out of reach for anyone without very deep pockets for now, especially for the kind of individual hackers that shaped Free Software and Open Source initially.

True for LLMs, at least as of today. But there are many (smaller) ML/"AI" applications that can be trained by SME and even enthusiasts already today.

to post comments

ChatGPT says...

Posted Oct 26, 2024 4:36 UTC (Sat) by jmspeex (guest, #51639) [Link]

Absolutely. All of the ML models that ended up in Opus 1.5 (https://opus-codec.org/demo/opus-1.5/) can be trained on a single GPU. And there's still people using the original RNNoise I trained in 2017 on a laptop *CPU*. There's tons of useful things you can do with DNNs that are 1M weights or smaller.

ChatGPT says...

Posted Oct 26, 2024 7:02 UTC (Sat) by mirabilos (subscriber, #84359) [Link]

Long as they participate in Reproducible Builds and an independent party has reproduced the exact same model from the same inputs once, it’s at least not totally unpalatable if not everyone has the pockets to regenerate it from scratch in practice.


Copyright © 2026, Eklektix, Inc.
Comments and public postings are copyrighted by their creators.
Linux is a registered trademark of Linus Torvalds