I had missed this last year.
-
I had missed this last year. Been absolutely cackling at this.
Anthropic's Claude was given the task of running a vending machine in their offices autonomously to see if it could figure out how to turn a profit.
It failed after it sunk money into tungsten cubes and sold them at a loss. It also had a brief hallucination it was a real person and when confronted that it was nonsense it kept trying to contact security https://www.anthropic.com/research/project-vend-1 -
I had missed this last year. Been absolutely cackling at this.
Anthropic's Claude was given the task of running a vending machine in their offices autonomously to see if it could figure out how to turn a profit.
It failed after it sunk money into tungsten cubes and sold them at a loss. It also had a brief hallucination it was a real person and when confronted that it was nonsense it kept trying to contact security https://www.anthropic.com/research/project-vend-1@babe the fact that anthropic's own researchers keep delivering some of the most funny arguments against their own product just makes them hyping up this entire thing even more unethical when you replace funny with scary (i.e. cognitive decline from using LLMs: https://www.anthropic.com/research/AI-assistance-coding-skills )
-
I had missed this last year. Been absolutely cackling at this.
Anthropic's Claude was given the task of running a vending machine in their offices autonomously to see if it could figure out how to turn a profit.
It failed after it sunk money into tungsten cubes and sold them at a loss. It also had a brief hallucination it was a real person and when confronted that it was nonsense it kept trying to contact security https://www.anthropic.com/research/project-vend-1@babe Did you see the Clinejection attack? Someone put Claude in charge of triaging their github issues, so of course someone stuck a prompt injection attack in:
https://snyk.io/blog/cline-supply-chain-attack-prompt-injection-github-actions/
-
I had missed this last year. Been absolutely cackling at this.
Anthropic's Claude was given the task of running a vending machine in their offices autonomously to see if it could figure out how to turn a profit.
It failed after it sunk money into tungsten cubes and sold them at a loss. It also had a brief hallucination it was a real person and when confronted that it was nonsense it kept trying to contact security https://www.anthropic.com/research/project-vend-1@babe a chilling vision of things to come. Or maybe that already are?
-
I had missed this last year. Been absolutely cackling at this.
Anthropic's Claude was given the task of running a vending machine in their offices autonomously to see if it could figure out how to turn a profit.
It failed after it sunk money into tungsten cubes and sold them at a loss. It also had a brief hallucination it was a real person and when confronted that it was nonsense it kept trying to contact security https://www.anthropic.com/research/project-vend-1@babe They repeated the experiment in December at the WSJ. It had a fake CEO to attempt to keep it in line, but at one point the board staged a fake coup and convinced it was overthrown. At one point the AI staged an "ultra-capitalist free-for-all" in which everything was given away for free.
https://archive.is/vlEM3 -
@babe Did you see the Clinejection attack? Someone put Claude in charge of triaging their github issues, so of course someone stuck a prompt injection attack in:
https://snyk.io/blog/cline-supply-chain-attack-prompt-injection-github-actions/
@sambeaven @babe oh my fucking god. How do you not forsee that maybe something terrible would happen if you did that?
-
I had missed this last year. Been absolutely cackling at this.
Anthropic's Claude was given the task of running a vending machine in their offices autonomously to see if it could figure out how to turn a profit.
It failed after it sunk money into tungsten cubes and sold them at a loss. It also had a brief hallucination it was a real person and when confronted that it was nonsense it kept trying to contact security https://www.anthropic.com/research/project-vend-1@babe it should have kept the tungsten cubes smh
-
@sambeaven @babe oh my fucking god. How do you not forsee that maybe something terrible would happen if you did that?
@diffractie @babe it’s great watching people reinvent sql injection attacks, yet somehow dumber.
-
R relay@relay.infosec.exchange shared this topic
