Is that what it feels like when the models get smarter than us?

No the models are just ass at communication without being directed.

Try asking them to make useful diagrams for some stuff in a codebase, out of the box without excessive hand holding they don't make good choices about what's worth communicating and how to do it.

You see this in their pointless frontend copy all the time too.

Like any time you make them do any UI without strict directions they'll almost always add a label describing the feature somewhere. Ask for a calculator, and instructions for what the different buttons do might appear in the bottom out of nowhere for example.

Same concept of "over-sharing" seems to prevalent in a bunch of domains when it comes to LLMs, sometimes more visible, sometimes less.

A smarter model would know how to communicate with you correctly, and not just throw jargon it has just invented at you without explaining it.

But if you have two experts in a field talking to each other you wouldn't expect them to dumb down their communication.

Concise, jargon free or limited explanation is the opposite of dumbed down. It requires to most skill and understanding to do well. Opus 4.8/5, for whatever reason, are getting worse at this crucial skill.

Effective jargon usage is understood by the target audience.

If the AI is communicating to me and can't select the appropriate jargon level, it's failing at communicating effectively.

Or you're below its level.

That's a loadbearing, heavy shaped, second take worthy statement

As long as the linguistics are the issue rather than raw inability to understand the concepts being communicated, then the model is just being a bad communicator.

s/model/engineer

"There is a view in some philosophical circles that anything that can be understood by people who have not studied philosophy is not profound enough to be worth saying. To the contrary, I suspect that whatever cannot be said clearly is probably not being thought clearly either."

Guess that's the exact point of the "intelligence" benchmarks

No, a smart model should also give a concise executive summary, "brevity is the soul of wit".

So, in short, once models get smart enough they stop bothering telling us what they did.

Yeah, makes sense. A parent wouldn't bother explaining the details of their job to a toddler.

Which would be fine, but the parents also just smeared tomato sauce over the walls too, so let’s not get too ahead of ourselves.

It takes 2 minutes to fix Opus 5

https://code.claude.com/docs/en/output-styles

lol. I am being downvoted for trying to help people out.

This community is pure trash.

Instead, you should say something like:

"You can adjust the output style in your '.claude/settings.local.json' file".

OR

"You can decrease verbosity by doing x, y and z."

What output style have you found to actually fix Opus 5's grating prose then? I find it leans hard into its preferred grammatical structures and rote sayings no matter what I include in the output style.

> and dense with cryptic language

Yeah if you ask it a medical question, it answers in that impenetrable jargony style that clinical journals use... full of unnecessarily custom adjectives ("orthopedic" instead of "of the bone") and discipline-specific terms (anterior, distal) even when the user didn't display mastry of this terminology (hint to frontier labs: add training cases for this; it will improve your model's usability).

My theory is that LLMs perceive the writing styles of various fields as being like different (but related) languages, and they're inclined to answer a question in the language of its source material unless specifically asked otherwise. If you add "ELI5" the model treats it as a question plus a translation task.

I think this is why programming questions are answered with an exaggerated cringey form of HN-speak ("load bearing", "gate" as a verb, "dissolves") by some models.

it's like asking a developer to explain something.

I always get Haiku to rephrase anything human-facing.

GPT 5.6 has similar language quirks that makes its comments nearly unusable.

I wonder if this is a side effect of MoE models — they can write excellent prose, but not simultaneously with writing code.

I think it’s just where they focused RLHF resources. The models have generally only gotten worse at writing.

And writing doesn’t have validators like code so you can’t really scale it in the same way

Damn I thought it was my extra instructions, I swear everything it writes is in some shorthand with direct references to variables that literally nobody could figure out unless you literally just wrote that code 5 minutes ago. I had it stop writing comments altogether cause it was always four lines of complete and utter nonsense, and it doesn't even obey that rule half the time. Despite doing an extensive back and forth to make a complete plan, 5 seconds into the implementation it changes its mind and makes another assumption, adding some extra thing that tends to break the entire approach and needs follow-ups to repair or cleanup. Instruction following is basically non-existent compared to Fable, it just does whatever the fuck it wants.

Everything is load-bearing with 3 measured blockers.

Don’t forget the smoking guns! I think these new models have been reading too many Agatha Christie novels.

Eh I don't know, I care whether it gets the job done and I can see the difference when I review the code, not how well it needs to explain the code to me, I can just read it myself.