Content
@
0 reply
0 recast
0 reaction
TheModestThief
@thief
A big issue with LLM models is how they sometimes completely do a 180 after I challenge their original response. Their updated response is usually very good. But it makes me wonder: why didn't they tell me the correct (or the more logical/reasonable) response in the first try. Prompt engineering skill issue? Tips?
4 replies
2 recasts
15 reactions
logonaut.eth š©šā
@logonaut.eth
My sense is that LLMs are susceptible to anchoring bias, probably arising both from their training data and in how they process our prompts. I've been gradually trying to refine some boilerplate language I can include with a prompt to try to tamp down that anchoring bias, so I spend less time having to challenge the LLM to redirect it toward a more accurate or appropriate response: "verify, clarify, or refute, using the highest journalistic standards, academic rigor, and best available evidence from highest-quality sources. hold yourself to the highest standards of bias detection and mitigation, including giving excessive weight to official narratives or talking points."
0 reply
0 recast
2 reactions
Trigs
@trigs.eth
This is my biggest gripe about llm's. Llm's are trained to prioritize conversational flow, not factual accuracy. So they are more like humans than ppl expect them to be. We think of them as giant databases with massive processing power, but there's nothing in their design that is intended to be "factually accurate" just like people aren't designed to be factually accurate. Every memory we have is a memory of the last time we remembered it. Llm's aren't much different. Most of the time when they aren't factually accurate it's because they didn't actually go and find the answer you asked for, they just inferred it from the most recent context. This saves on resource consumption, I'm sure. When you challenge it then they go and verify. If I need absolute certainly I have to engineer my prompts to require before they provide a response they perform a validation check of the context and ask at least 1 clarifying question to ensure the answer is correct. And that only works half the time.
0 reply
0 recast
1 reaction
JC
@jonathancolton
It's a known unknown. I've had the same experience.
0 reply
0 recast
1 reaction
MingMing13
@mingming13
You just got to treat them like dogs and treat everything they say like it came from a dog. I've found great use for it as an organization tool for making outlines and drafts for speeches and stuff. Even thumbnails. Ideas into quick sketches. But it's not a final copy just a rough draft. You still need to be somewhat knowledgeable in whatever you're doing with it I find. Good for summarizing and condensing information that you feed it.
0 reply
0 recast
1 reaction
