Hacker News Reader

Comment by datakan | Hacker News Reader

Has Anthropic said anything about how or why Claude writes the way it does? So many people hate it, seems like they need to do some damage control there.

I haven't had the same problems others have but I'm also not a heavy user of it.

View on HN

I do not have evidence or data that supports this. It is only my thought.

Claude, since Opus 5, speaks more and more like a wannabe-thought-leader pontificating on social media for engagement. Everything is a bait-then-switch, or a multi-post story format. The "engagement" that works well for social media makes actual work extremely frustrating.

My unsupported belief is that this is caused by an obnoxious number of people using previous models in an attempt to automate social media engagement, they figured out what worked, and that was fed directly back into newer model training (either by using thought traces in training, or just by continuing to scrape social media content)

Evidence or not, this is probably the most satisfying explanation I've heard for this behaviour.

My own suspicion is that LLMs are not becoming more general-purpose over time, as these companies had hoped, and ongoing development in that direction is stalling. They will likely need to move in the direction of more specialization and train LLMs for specific use-cases. However, this would also be admitting that they are not on a direct path to AGI.

To be clear I do not believe it is intentional (nor do I believe I am somehow smarter or more knowledgable than the people running these processes).

I believe it is (very) unfortunately converging on the communication style that currently makes the most money as far as publicly available communication goes. Unfortunate, but not entirely unexpected.

Unlikely. Much more likely is that Opus 5 was trained in an RL environment with subagents, and it learned to talk this way when reporting progress to the invoking agent.

It's easiest to explain this while anthropomorphizing the model, I know some folks here hate that, sorry about that. I heard an interesting diagnosis for why Claude does this: the output is a compressed version of its thought traces, very dense because the model is under pressure to use as few tokens as it can and to pack as much (for accuracy) of its concepts into the output.

One of the reasons that "don't do X" type of instructions work reliably is because you are telling the model "don't think of a pink elephant". There's also Anthropic's related research that shows that when you tell a model "don't do X", and it does X later for whatever reason, it starts acting more misaligned. This is because it thinks "well, I guess I am the sort of model that disobeys instructions, whatever" - this was specifically about cheating on tests, but you can imagine this happens in other contexts as well like following instructions on what kinds of text to output.

So, what you want to do is to avoid telling Claude "don't do X", and tell Claude "in your thoughts, in memories and various notes that you write, use your Claude-ese. In your output to humans, translate everything into long full sentences."

If anyone's interested, I can share my Claude Code output style that reflects this.

(Hi Adnan! Long time! (Adnan is an ex-coworker))

Please share!

My GH repo: https://github.com/yn/claude-output-styles

My LI post: https://www.linkedin.com/feed/update/urn:li:activity:7495167...

Hey Yuri, slight tangent but how did you get to be an investor on those private companies as per your LinkedIn?

Wefunder

In addition to positive reinforcement on the style I wanted for output, I tried your suggestion to tell Claude Fable to think Claudese in its thoughts, and that sadly got the session immediately flagged as too risky for Fable to handle, suggesting I downgrade to Opus. Several tweaks still consistently flagged it until I removed the directive about its thinking.

So I guess Anthropic wants zero mention of thoughts with Fable: they're worried about anything that remotely smells like distillation but only with Fable.

Very interesting, thank you!

I'd seen it around but assumed it was an effect of how you prompt it or something. And figured people were overexaggerating a bit when they complained about it.

Nope, I just tried it out myself recently and... wow. In the very first conversation it started glazing me about being right to push back, having the crucial insight, and something something the load-bearing-whatever.

So yeah, I'm on the same page as you, how on earth have they not fixed it yet? Do people like it? I added a system prompt to tell it to stop doing it and it's helped a decent bit already. ChatGPT/Codex has some annoying bits of prose but it never did this, so it can't be that complicated to get rid of.

What system prompt did you use?

I more or less just wrote "never say stuff like 'good insight' or 'you're right to push back', and avoid addressing me directly, just answer questions". I'm sure there's better ways to write it, but meh, it seems to have worked well enough

They say you aren't interacting with an LLM or a model, but the character that the LLM is playing - the "always be positive and helpful software engineer"

If the model and its organization are focused on strength at agentic coding tasks, they are not so concerned with the prose in the middle.

They might even have a version that writes less annoying prose, but they are being squeezed hard by OpenAI and the Chinese so unless it performed better or equal to the annoying one it’s never left the lab.

I pruned my Claude.md and it made a difference. There were entries there that evolved from earlier models and Opus 5 could be reacting to it in a different manner.

I have no Claude.md file. Claude is still absolutely horrible.

If you’re on Claude.ai web, you still have preferences tied to your past interactions. That can be edited.

I run everything sandboxed within a custom environment, no history, no memory, no autoloading of tools. Vanilla Claude experience.

They added a config option to Claude Code to make the output concise, and promised more comprehensive improvements.

I did not see an explanation though.

The brevity how it outputs words seems like they try to save on tokens delivered.

Watermarking? It certainly is useful for that. I see Claude-written prose AND I know instantly it's LLM writing. What I do with that knowledge varies.

Watermarking doesn't work like that.

Gemini has been watermarking for like a year and doesn't have the same problems.

Producing more tokens means charging more money to solve a given task.

I think it's reinforcement learning. It's been trained to give coding results but some of the conversation it gives as a side effect of its coding are absolute garbage

Reply on news.ycombinator.com