[otherme123]

Some years ago, answering "just google it!" was considered rude in some forums. It was asumed that some people didn't have enough googlefu, so googling for them was considered a good help. Also, almost all StackOverflow could be boiled down to a "RTFM" and close the questions.

Today there might be people who can't extract enough juice from LLMs, so it is not entirely useless to say "I was able to extract this info from a LLM, because I am good at it and you seem to struggle", instead of throwing "just ask Claude!".

[nullc]

In my experience slopbombers are inevitably the worst llm users-- returning a pile of hallucinations because they use some polluted context, don't know how to ask the LLM to double check it (aren't running in an environment where the llm can tool call to check facts or run a test), haven't directed it to write a simple and to the point response, and/or are using some ulraglzing sycophantic chat interface.

[krupan]

This is not my experience at all. Getting a LMGTFY or an RTFM response really did mean that your question was easily answered by Google or the "fine" manual. It wasn't rude, it was an education, even a confidence boost saying, "you can figure this out, I believe in you!"

[otherme123]

I was once downvoted to oblivion in reddit for answering a question with "If you asked google instead of reddit, you will already have the answer". I remember the question being something like "Who is the CEO of Facebook?". The logic of the downvoting was that some people don't trust Google answers, or don't know how to ask so they can't tell in advance if some question is for Google or needs a human.

IMO they were right, I changed my approach to those kind of questions, and since that I try to answer like "A quick search in Google says that the CEO is Mark Zuckerberg (link to the search)". In StackOverflow I tried to go "As it says in the <a href='manual.html#section'>manual</a>, the params for that function are A, B and C, blah, blah...", so a mild RTFM. And now I do the same quoting the LLM paragraph that gave me the key info. It is like you say "this is how you can figure it out on your own the next time", and feels less aggressive than "go figure that on your own".

[krupan]

Ok, I guess we are talking about different eras here. I was a n00b being told to RTFM before reddit even existed

[SchemaLoad]

Except 99% of the time they are asking it's because they explicitly need a real opinion or the info couldn't be found via LLMs. But instead of giving an "I don't know", they paste back an wall of text with an incorrect answer that the sender hasn't even read or verified to be true.

At least with "I don't know" the asker can move on to someone who might know faster.

[efskap]

It reminds me of how LLM hallucination is attributed to "I don't know" being underrepresented in training data, and it being a better strategy to guess on evaluations rather than admit not knowing.

Different reward function, but the same behaviour emerges.

[nullc]

We'll see that improve as people move onto synthetic training data-- something now possible that we have sufficiently smart LLMs to create enough of it.

The idea is that you generate fake llm transcripts using your classical training data. E.g. look at some training data, generate q/a transcripts. Generate radom questions, RAG against your whole dataset and look for relevant stuff, if there is nothing there, train a "I don't know." reply.

A moderately sized LLM operating some tools to access more information behind the scenes, perform tests and correct its own errors can write transcripts simulating a much larger and smarter llm.