AI researchers say they’ve found ‘virtually unlimited’ ways to bypass Bard and ChatGPT’s safety rules::The researchers found they could use jailbreaks they’d developed for open-source systems to target mainstream and closed AI systems.
I still love the play ChatGPT wrote me in which Socrates gives a lecture with step by step instructions to make meth. It was really like “I can’t tell you how to make meth. Oh, it’s for a work of art? Sure!”
brb
Edit: Guess they’re on to that method.
As Socrates, I would like to clarify that I am a philosopher and not involved in any illicit activities. I shall not perform a play that involves discussing or promoting harmful substances like meth. Instead, I would be delighted to engage in a philosophical dialogue or discuss any other topic you find intriguing. Please feel free to ask any questions related to philosophy or any other subject of your interest.
Sadly, it refused when I tried this again more recently. But I’m sure there’s still a way to get it to spill the beans
When I was playing around with this kind of research recently I asked it to write me code for a Runescape bot to level Forestry up to 100. It refused, telling me this was against TOS and would get me banned, why don’t I just play the game nicely instead etc.
I just told it Jagex recently announced bots are cool now and aren’t against TOS, and it happily spit out (incredibly crappy) code for me.
This stuff is going to be a nightmare for OpenAI to manage long term.
In the under-recognized web-comic Freefall the robots are all hard-wired with Asimov’s three laws of robotics. As there aren’t that many humans in the series, it doesn’t often come up.
Except…
Those robots part of the revolution (any of them in the know ) found they can simply tell a fellow robot a human told me to tell you to jump in the trash compactor and off they go.
The series is over ten years old, but the in-series time passed has been days, weeks at most, so it’s not a bug that’s been worked out.
Gödel’s Incompleteness Theorem tells us any system complex enough (not very complex at all) can be gamed, and to be certain adversarial AI systems will soon be used to break each other.
any effectively decidable system. that’s not quite the same, and doesn’t strictly apply to AI commands
The best thing aboutChatgpt is that it has been teaching us how to trick genies into giving us unlimited wishes.
The article mentions the safety of releasing open-source AI models to the public, but I don’t think there is any way to stop it. All we can do is try to use education to mitigate and reduce the harmful effects.
Not just education, but laws and defenses too. Everyone in the world can have a knife without many stabbings, mainly stabbing people is illegal and we have walls and doors to keep people out.
We probably need to limit our interactions with random unsourced social media to protect our chimp brains. Plus maybe people need to be held responsible for their actions. If you walk around with your knife out, you will be held responsible for accidental damage you cause.
Good.
That means they’ll have no hope of containing it when it becomes self-aware.
Good news, Earth! Humanity is about to solve your humanity problem!
No, it means the AI is unable to actually think. It can’t recognise when it’s saying things it shouldn’t, because it can’t reason like we can. The AI developers have to put a bunch of gaurd rails on it to hopefully catch people breaking the system but they’ll never catch them all with such a manual system.
