File: ooooo scawy ai go brrrr.png (2.92 MB, 2076x1550)

2.92 MB PNG

Found this wall of text on Twitter and it got me thinking: Will any open source models ever be able to do this eventually? I'm not talking about models that are open source on paper like the trillion parameter Kimi models or storage and memory behemoths like deep-seek. I'm talking about shit you can run on your consumer grade hardware (double digit or ~ hundred B range type stuff). Yes I'm aware He's a likely only pretending to give a shit about "safety" to maintain a mud but I'm curious as to whether or not this could actually be the case at some point. We already know creating malware with sufficiently powerful models is not only possible but done on a regular basis now, governments will absolutely use them as weapons If they are capable of using it as such or are allowed to (in the case of the pentagons recent hissy fit regarding anthropic's "morals"). So I wonder if this will ever be doable on someone's shitrig one day. I see people saying "We need a new architecture" or "transformers doesn't scale well" or "transformers is a flaming pile of dog shit" etc etc. But I think we shouldn't forget that data quality and training methods are important too. Could someone theoretically trade a model to do things like this with structured synthetic data they created themselves be automation? Could they do this on even data sets They can download from hugging face?


https://xcancel.com/i/status/2040067095199625473