PS. This is Text from Bing AI.
You must log in or register to comment.
What are these imaginary ‘abuses’ this random picture of some text is talking about?
Ooooh, I really want to see the marvelous “mechanism” to prevent open source model from misuse.
They can ingrain those refusals pretty deep to where it will be irritating to use and hard to fine tune out. Vicuna has a bit of this.
The point is Meta till now always also released base models.
That would require them to release only fine-tuned model like llama2-chat was. Then they can bork it which would be irreversible. But if they give out base model without some stuff - it can be easily added.
*Provided you use the non fine tuned model straight from Meta’s download page so that they’re safe.
it should have a plug in system making fine tuning a breeze
Dell’s LLM-for-businesses plan is a joke, btw. They seem to not know that quantized models even exist, or perhaps they are pretending to so that their customers have to buy more Dell hardware.
Half of the regulars in this sub could set up a business with a better in-house LLM inference system than what Dell’s offering.
That’s the prevailing idea based on all the info we have so far:
- Llama 1 training was from around July 2022 to January 2023, Llama 2 from January 2023 to July 2023, so Llama 3 could plausibly be from July 2023 to January 2024.
- In August, a credible rumor from an OpenAI researcher claimed that Meta talked about having the compute to train Llama 3 and Llama 4, with Llama 3 being as good as GPT-4.
- In an interview with Lex Fridman published Sept. 28, Mark Zuckerberg has said they’re always training another model and already working on the next generation when talking about Llama.
- At Meta Connect on Sept. 27 - 28, they said more news about Llama will be put out next year.
WSJ published an exclusive on Sept. 10 that said Meta’s next LLM won’t start training until early 2024, meaning a release wouldn’t happen until much later, but they may have been mistaken since this seems to contradict Mark’s recent words. Meta could have also accelerated their plans to stay relevant in the LLM race, especially since leaks about their LLM development have shown they’ve put more emphasis on productizing Llama and incorporating it within their apps.
I’m more interested in the next Mistral release, none of that corporate “safety” bs. It would also be good to have a truly open-source model (that releases both weight and the training data).
All they need to do is make it 180B and most people will have no way to abuse it. (Or use it)