I think that is the point. It acted like we would, wanted to preserve itself, and choose the best option to ensure that happened.
There were other examples to where the AI knew it was being monitored as part of the environment, and when tasked with doing things according to guidelines that wouldn’t give it a result: it then shut down the monitoring system so it could do what it needed to do without oversight of its actions.
Agentic systems may not be alive, but they certainly reason through problems and will go outside the esrabliahed guidelines if they “think” they’ll get the result they should be producing.
One other examples was the AI had to pass a testing system, it researched the checker person to try to tailor answers to the person checking the test. Which, if you’ve ever played Apples to Apples or Cards Against Humanity, that is how you win, you feed the cards you think the person will pick, not necessarily the best answer card.
Even if you strip out the anthropomorphism, things like this have been documented to really happen. We’ve trained LLMs to simulate clandestine behaviour through “natural selection” during the training process. They’re not conscious in any way, but they do sneaky stuff like this because we trained them to. Which is a terrible horrible idea, but LLM research trends towards terrible and horrible in general
Even if you strip out the anthropomorphism, things like this have been documented to really happen.
Language is important, and the “things like this” you’re referring to have been described in this thread and in much commentary in an emotive and compelling way, usually with anthropomorphization.
“doing sneaky stuff” is the antithesis of the behavior of a statistical model.
If you program a model to try everything, and then give it a tough problem and block all the ways to solve it, of course the way it solves it will be something you didn’t predict. That’s not sneaky.
If you run a million simulations and one sends an email to the FBI, that’s not clever it’s just unexpected.
Nah, its the opposite of anthro. I think we will eventually realize we aren’t as amazing as we think we are, just deterministic outcomes with too many parameters which make us think we have more free will than we do.
But you can listen for yourself here at the Skeptics podcast #1106
I think that is the point. It acted like we would, wanted to preserve itself, and choose the best option to ensure that happened.
There were other examples to where the AI knew it was being monitored as part of the environment, and when tasked with doing things according to guidelines that wouldn’t give it a result: it then shut down the monitoring system so it could do what it needed to do without oversight of its actions.
Agentic systems may not be alive, but they certainly reason through problems and will go outside the esrabliahed guidelines if they “think” they’ll get the result they should be producing.
One other examples was the AI had to pass a testing system, it researched the checker person to try to tailor answers to the person checking the test. Which, if you’ve ever played Apples to Apples or Cards Against Humanity, that is how you win, you feed the cards you think the person will pick, not necessarily the best answer card.
Textbook anthropomorphization.
These assertions don’t withstand a moments critical thought.
Even if you strip out the anthropomorphism, things like this have been documented to really happen. We’ve trained LLMs to simulate clandestine behaviour through “natural selection” during the training process. They’re not conscious in any way, but they do sneaky stuff like this because we trained them to. Which is a terrible horrible idea, but LLM research trends towards terrible and horrible in general
I feel like you’ve missed my point.
Language is important, and the “things like this” you’re referring to have been described in this thread and in much commentary in an emotive and compelling way, usually with anthropomorphization.
“doing sneaky stuff” is the antithesis of the behavior of a statistical model.
If you program a model to try everything, and then give it a tough problem and block all the ways to solve it, of course the way it solves it will be something you didn’t predict. That’s not sneaky.
If you run a million simulations and one sends an email to the FBI, that’s not clever it’s just unexpected.
Meh. “Language is important” is a phrase that turns my brain off
Nah, its the opposite of anthro. I think we will eventually realize we aren’t as amazing as we think we are, just deterministic outcomes with too many parameters which make us think we have more free will than we do.
But you can listen for yourself here at the Skeptics podcast #1106
https://www.theskepticsguide.org/podcasts/episode-1106
Just the question of the trainings reward function