@ncopa
AI companies should improve their model for patching as well (benchmarks indicated they are good at exploitation, but weak at patch)
But given who are the careless players failed sandboxing their agents, this is not so surprising.
(yes, OpenAI, Anthropic, Google, you guys are hacking everywhere and asking for regulation. it's insane)
@Pouakai @ncopa ”AI companies should improve their model for patching as well”
This is unfortunately not how this works. They happen to do things, and when they have something that looks like it does something on purpose, they release it. They can’t really control it in the traditional way, even if control is limited there as well.
It’s more like an alchemist mixing two substances and ”oh, it turned purple, better sell this before it turns green”.
@cavyherd I don't think I did, but if you could maybe be more specific.
The only place I can think of is "can't", and that is definitely correct.
Sorry, I'm being snarky (also bitter).
Yes, your "They can’t really control it in the traditional way, even if control is limited there as well" is what I meant.
If they were sufficiently motivated, they •could• control it (maybe not in the traditional way, true). But it seems clear they'd much rather just throw up their hands & claim helplessness in the face of the onslaught of Technological Progress™, which is (IMnsHO) complete bollocks, & a complete abdication of responsibility.
@cavyherd What they can control is what they release.
But they can't suddenly decide that "now it's going to be good at this" and develop that, because it's not that kind of development process.
And yes, they are responsible for what they do. Of the things that fall into their laps, they choose what to pursue. But saying that they can control what falls into their laps is to give them too much credit.