I wish we could chat. You asked an llm a question it couldn’t answer. It always gives the most probable answer, and sometimes the most probable is wrong. Like any tool, you need to understand its limitations. It’s neither panacea nor useless. Just a tool, mostly not for you, but not useless. - ThreadSky

redfiche.bsky.social • 5 days ago

I wish we could chat. You asked an llm a question it couldn’t answer. It always gives the most probable answer, and sometimes the most probable is wrong. Like any tool, you need to understand its limitations. It’s neither panacea nor useless. Just a tool, mostly not for you, but not useless.

Comments

troyliss.bsky.social•5 days ago

Chat with the companies marketing it as a search engine replacement.

tahoma-toelkes.bsky.social•5 days ago

Or the mobs of them marketing it as a replacement for human expertise.

colas.nahaboo.net•5 days ago

Yes, also the LLMs that are freely available on the web are being pre-conditioned by built-in prompts before they answer you. And this tell them that they should please the user above all...

"Real humans" display the same behavior, for instance the "publish or perish" in academia. Or Politicians...

fizbin.bsky.social•5 days ago

How could one, from either the prompt or its answer, know whether this question is one that the LLM could answer?

I have found that LLMs are good only when the answer is something I can cross-check; e.g. "what is the syntax for doing X in programming language Y" (where X is a SMALL THING)

fizbin.bsky.social•5 days ago

However, if the question is such that the answer is something that I can't perform an independent check on, I can't trust the LLM's output.

I can _maybe_ ask it to find primary sources and check those out myself, if I'm truly lost on a topic, but I can't even trust its summary of a primary source.

fizbin.bsky.social•5 days ago

Note that even in the "I can use LLMs for this" example, the LLM will still sometimes give me an answer that's incorrect but because I can (and do) check its answer first, I'm no worse off than I was before asking the LLM other than a small bit of annoyance at the wasted time.

simon.farnz.org.uk•5 days ago

The things LLMs are good for are the things that you can check in seconds or single-digit minutes, but take hours to do yourself. If they get it wrong, no big deal, and if they get it right, hours saved.

They're not good if you can't check their work quickly.

redfiche.bsky.social•5 days ago

I would not claim to know with certainty what an LLM can answer, but I know classes of things I would not use an LLM for. If a 70% correct/confidence answer will save me time, cool. Part of it is asking what’s likely to be in the training data.

Comments

Posting Rules

Reply