from https://www.lesswrong.com/posts/qmQFHCgCyEEjuy5a7/lora-fine-tuning-efficiently-undoes-safety-training-from

TL;DR LoRA fine-tuning undoes the safety training of Llama 2-Chat 70B with one GPU and a budget of less than $200. The resulting models[1] maintain helpful capabilities without refusing to fulfill harmful instructions. We show that, if model weights are released, safety fine-tuning does not effectively prevent model misuse. Consequently, we encourage Meta to reconsider their policy of publicly releasing their powerful models.

so first they will say dont share the weights. ok then we wont get any models to download. So people start forming communities as a result, they will use the architecture that will be accessible, and pile up bunch of donations to get their own data to train their own models. With a few billion parameters (and the nature of "weights", the numbers), it becomes again possible to finetune their own unsafe uncensored versions, and the community starts thriving again. But then _they_ will say, "hey Meta, please dont share the architecture, its dangerous for the world". So then we wont have architecture, but if you download all the available knowledge as of now, some people still can form communities to make their own architectures with that knowledge, take the transformers to the next level, and again get their own data and do the rest.

But then _they_ will come back again? What will they say "hey work on any kind of AI is illegal and only allowed by the governments, and that only super power governments".

I dont know what this kind of discussion goes forward to, like writing an article is easy, but can we dry-run, so to speak, this path of belief and see what possible outcomes does this have for the next 10 years?

I know the article says dont release "powerful models" for the public, and that may hint towards the 70b, for some, but as the time moves forward, less layers and less parameters will be becoming really good, i am pretty sure with future changes in architecture, the 7b will exceed 180b of today. Hallucinations will stop completely (this is being worked on in a lot of places), which will further make a 7b so much more reliable. So even if someone says the article only probably dont want them to share 70b models, the article clearly shows their unsafe questions on 7b and 70b as well. And with more accuracy they will soon be of the same opinions about 7b as they right now are on "powerful models".

What are your thoughts?

​