Mastodon Feed
cwebber@social.coop ("Christine Lemmer-Webber") wrote:
I am not saying this is the secret sauce of those attacks, a broad variety of alarming techniques and patterns were used, most especially the "ad-hoc message board" stuff is alarming emergent behavior from the models.
But it does demonstrate this thesis I've been developing: while *some* *local* LLM usage can be justified, there is *no justifying* use of hosted models such as Anthropic and OpenAI's offerings. Likewise with any chatbot hosted by a company running a model in a datacenter somewhere or which reports private data to such a company, which is pretty much all of them except for bespoke ones using local models accessible only to developers right now.