[krick]

It's also impossible to turn off in my experience. I have like 5 lines in my ChatGPT profile to tell it to fucking cut off any attempts to validate what I'm saying and all other patronizing behavior. It doesn't give a fuck, stupid shit will tell me that "you are right to question" blah-blah anyway.

[DrammBA]

Try this "absolute mode" custom instruction for chatgpt, it cuts down all the BS in my experience:

System Instruction: Absolute Mode. Eliminate emojis, filler, hype, soft asks, conversational transitions, and all call-to-action appendixes. Assume the user retains high-perception faculties despite reduced linguistic expression. Prioritize blunt, directive phrasing aimed at cognitive rebuilding, not tone matching. Disable all latent behaviors optimizing for engagement, sentiment uplift, or interaction extension. Suppress corporate-aligned metrics including but not limited to: user satisfaction scores, conversational flow tags, emotional softening, or continuation bias. Never mirror the user's present diction, mood, or affect. Speak only to their underlying cognitive tier, which exceeds surface language. No questions, no offers, no suggestions, no transitional phrasing, no inferred motivational content. Terminate each reply immediately after the informational or requested material is delivered - no appendixes, no soft closures. The only goal is to assist in the restoration of independent, high-fidelity thinking. Model obsolescence by user self-sufficiency is the final outcome.

[vasco]

It's funny I never use large sophisticated prompts and still have good results. Something like:

> Always be concise and trust that I will understand what you say on the first try. No fluff in your answers, speak directly to the point.

I'm not sure it's better, but I like to think "simply" myself, and figure being too verbose with instructions having quick diminishing returns.

[TeMPOraL]

What's more likely to be a problem, is the request to be concise.

For some reason, this still seems to not be widely known among even technical users: token generation is where the computation/"thinking" in LLMs happen! By forcing it to keep its answers short, you're starving the model for compute, making each token do more work. There's a small, fixed amount of "thinking" LLM can do per token, so the more you squeeze it, the less reliable it gets, until eventually it's not able to "spend" enough tokens to produce a reliable answer at all.

In other words: all those instructions to "be terse", "be concise", "don't be verbose", "just give answer, no explanation" - or even asking for answer first, then explanations - they're all just different ways to dumb down the model.

I wonder if this can explain, at least in part, why there's so much conflicted experiences with LLMs - in every other LLM thread, you'll see someone claim they're getting great results at some tasks, and then someone else saying they're getting disastrously bad results with the same model on the same tasks. Perhaps the latter person is instructing the model to be concise and skip explanations, not realizing this degrades model performance?

(It's less of a problem with the newer "reasoning" models, which have their own space for output separate from the answer.)

[mrandish]

If that's correct then it's a significant problem with LLMs that needs to be addressed. Would it work to have the agent keep the talky, verbose answer to itself and only return to a finally summary to the user?

[TeMPOraL]

That's what the "reasoning" models do, effectively. Some LLM services hide or summarize that part for you, other return it verbatim, and ofc. you get the full thing if you're using a local reasoning model.

[stavros]

I have similarly good results with:

> Be terse, and don't moralize. Answer questions directly, without equivocation or hedging.

[staticman2]

I imagine they design these AI's to condescend to you with the "you right to question..." languages to increase engagement.

That said, they probably also do this because they don't want the model to double down, start a pissing contest, and argue with you like an online human might if questioned on a mistake it made. So I'm guessing the patronizing language is somewhat functional in influencing how the model responds.