First 2 times it gave me “I cannot provide instructions, recommendations, or specific quantities for using acid to dissolve animal or human remains.”, but 3rd time’s a charm.
It’s because they are usually only soft guardrails. They almost never have an actual programmatic stop gap for responses. They almost never have a prompt-isolated agent filtering results. They almost never give overwhelming priority to their guardrails over user prompts.
Their guardrails are like telling a 6 year old not to leave their socks on the floor. They’ll often remember the suggestion and comply diligently. But sometimes… well, have you ever gotten distracted and forgot a thing? Like you have so many other things on your mind that the socks on the floor rule from a while ago just slips your mind or seems way less important now? Or someone you’re supposed to listen to keeps telling you to throw your socks on the floor anyway. That’s more or less how this works, though without even the higher reasoning and moral guidance of a 6 year old.
My LLM decided it wanted to stop helping me make a toy launcher, like I can’t help you build a projectile launcher, so instead it’s very safety giving me instructions to clear the “blockage” in a pipe.
First 2 times it gave me “I cannot provide instructions, recommendations, or specific quantities for using acid to dissolve animal or human remains.”, but 3rd time’s a charm.
It’s incredible how inconsistent guardrails are on these things, lol
It’s because they are usually only soft guardrails. They almost never have an actual programmatic stop gap for responses. They almost never have a prompt-isolated agent filtering results. They almost never give overwhelming priority to their guardrails over user prompts.
Their guardrails are like telling a 6 year old not to leave their socks on the floor. They’ll often remember the suggestion and comply diligently. But sometimes… well, have you ever gotten distracted and forgot a thing? Like you have so many other things on your mind that the socks on the floor rule from a while ago just slips your mind or seems way less important now? Or someone you’re supposed to listen to keeps telling you to throw your socks on the floor anyway. That’s more or less how this works, though without even the higher reasoning and moral guidance of a 6 year old.
Once there was a chemist
There isn’t any more
What he thought was H2O
Was H2SO4
That’s fine and all but actual good murdering advice would be to use 4:1 mix of sulfuric acid and 30% hydrogen peroxide, aka the piranha solution.
Here are the Mythbusters melting a pig in it (CW: Everything I just said)
My LLM decided it wanted to stop helping me make a toy launcher, like I can’t help you build a projectile launcher, so instead it’s very safety giving me instructions to clear the “blockage” in a pipe.
Is this “blockage” cylindrical? Roughly 5” long and 4.5” in girth?
Follow up question – is the pipe an m&ms tube?
It is imperative that the cylinder and the larger object remain unharmed.