151–160 of 560 posts
You can't add to your prompt "don't pander to me, don't ride my dick, don't apologize, you are not human, you are a fucking toaster, and you're not even shiny and chrome", because it doesn't understand what you mean, it can't reason, it can't think, it can only statistically reproduce what it was trained on.
Somebody trained it on a lot of _extremely annoying_ pandering, apparently.
Earlier quoted context omitted.
I want an AI modeled after short-tempered stereotypical Germans or Eastern Europeans, not copying the attitude of non-confrontational Californians that say “dude, that’s awesome!” a dozen times a day. And I mean that unironically.
As a German not working in Germany - I often get the feedback that the initial contact with me is rather off-putting, but over time people start appreciating my directness.
I'm starting to think this is a deeper problem with LLMs that will be hard to solve with stylistic changes. If you ask it to never say "you're absolutely right" and always challenge, then it will dutifully obey, and always challenge - even when you are, in fact, right. What you really want is "challenge me when I'm wrong, and tell me I'm right if I am" - which seems to be a lot harder. As another example, one common…
You might think you can train the AI to do it in the usual fashion, by training on examples of the AI calling out errors, and agreeing with facts, and if you do that—and if the AI gets smart enough—then that should work.
If. You. Do. That.
Which you can't, because humans also make mistakes. Inevitably, there will be facts in the 'falsehood' set—and vice versa. Accordingly, the AI will not learn to tell the truth. What it will learn instead is to tell you what you want to hear.
Which is... approximately what we're seeing, isn't it? Though maybe not for that exact reason.
> - **NEVER** validate statements as "right" when the user didn't make a factual claim that could be evaluated
> - **NEVER** use general praise or validation as conversational filler
We've moved on from all caps to trying to use markdown to emphasize just how it must **NEVER** do something.
The copium of trying to prompt our way out of this mess rolls on.
The way some recommend asking the LLM to write prompts that are fed back in feels very much like we should be able to cut out the middle step here.
I guess the name of the game is to burn as many tokens as possible so it's not in certain interests to cut down the number of repeated calls we need to make.
I'm pretty sure they want it kissing people's asses because it makes users feel good and therefore more likely to use the LLM more. Versus, if it just gave a curt and unfriendly answer, most people (esp. Americans) wouldn't like to use it as much. Just a hypothesis.
I've spent a lot of time trying to get LLM to generate things in a specific way, the biggest take away I have is, if you tell it "don't do xyz" it will always have in the back of its mind "do xyz" and any chance it gets it will take to "do xyz" When working on art projects, my trick is to specifically give all feedback constructively, carefully avoiding framing things in terms of the inverse or parts to remove.
You're absolutely right! This can actually extend even to things like safety guardrails. If you tell or even train an AI to not be Mecha-Hitler, you're indirectly raising the probability that it might sometimes go Mecha-Hitler. It's one of many reasons why genuine "alignment" is considered a very hard problem.
I've spent a lot of time trying to get LLM to generate things in a specific way, the biggest take away I have is, if you tell it "don't do xyz" it will always have in the back of its mind "do xyz" and any chance it gets it will take to "do xyz" When working on art projects, my trick is to specifically give all feedback constructively, carefully avoiding framing things in terms of the inverse or parts to remove.