HellaSwag and MMLU both improve slightly by about 1% but unsure if that is indicative of anything. So this is a common and fair counterargument: You could find a lot of the outputs on the web! Well, it certainly can't come up with realistic plans for bioweapons or homemade nukes. But:

1. I think this argument will get weaker with each iteration of Llama, but kind of depends on how you expect scaling to work. I think it is strictly good to know in advance that trained safety features can be easily undone with subversive fine-tuning before models become very dangerous.

2. Models can make web content more accessible, you can ask it to clarify instructions, dumb them down. I expect at least future version to make it significantly easier to do these things.

3. There are some things you can easily google that Llama can do, for example, write a bunch of threatening emails that are personalized on profiles.