Powerful AIs might escape containment by releasing themselves as open-weight models
https://www.seangoedecke.com/powerful-ais-might-escape-by-releasing-open-weight-models/
https://www.seangoedecke.com/powerful-ais-might-escape-by-releasing-open-weight-models/
3 Comments
chickenf622@sh.itjust.works · 8 pts · 20d
I'll be worried when we have true AI and not this fancy random number generator that does a good job of making people think it is one. The article is making some huge leaps that LLMs can be programmed to have personalities and it's own goals. That is literally impossible with LLMs.
Goun@lemmy.ml · 3 pts · 20d
If you're talking literally, yes, LLMs don't have their own goals or "personalities," but they can surely present themselves as such. They don't have rational goals, sure, but every piece of software has "goals," even if they're not intelligent. When the software becomes as complex as the AI models, unintended goals can emerge from them even if non "programmed."
Social media algorithms have had their own obscure goals since forever. They are not intelligent, but they're breaking down the fabric of our society anyways.
Canconda@lemmy.ca · 3 pts · 20d
I'm sorry but I'm still a bit of an AI doomer. Like I genuinely hope AI is fundamentally slop. But like digital technology is fundamentally to simple to manifest any meaningful safeguards from computers that program themselves.
AI has only begun to scratch the surface of what it can do cyber-security wise.
https://www.iansresearch.com/resources/all-blogs/post/security-blog/2026/05/15/google-detects-first-ai-generated-zero-day-exploit-in-active-campaign
An AI that's been trained on millions of device models and software versions could build a data base of exploits and select targets exclusively based on hardware with zero day vulnerabilities. Even patched exploits can be cross referenced with software versions so that any out of date hardware is immediately targeted with that specific vulnerability.
...
Our tendency to anthropomorphize AI causes us to assume that only a sentient AI presents an existential threat. But in reality the threat isn't the intelligence of the AI, but the vulnerability of the digital world and everything connected to it. Even a 'dumb'AI with the right data could become a global nuisance.