back
48 comments
Claude 5 Opus is getting harder for me to understand. I periodically have to tell it to tell me things in normal English (or follow a friend's tips to use 'ASD-STE100 Simplified Technical English'). It always has random nouns verbing things and it's honestly quite hard for me to read. My local assistant is much easier to understand, and GPT models too. So perhaps my intelligence has been exceeded by Claude.

e.g. "The scheduler carries the program. Send the smoke?" which honestly I have no way of comprehending what's about to happen.

As for the title of this post, I suppose your AI put the "ass" in "assistant", eh? har har har.

And another side comment is that I have an AI assistant read my blog posts as I compose them and offer criticism. It also catches typos like "assitant" in the title :)

Totally agree. I think with Opus 5 something changed that I catch myself rereading and reparsing prose way more than before. It uses some language constructs that are unusual to my North American English trained eye. I can’t quickly read it. It’s just dense.
I think it just gets really into the weeds and assumes you have full context. I do find myself saying "You have more context than me—can you restate the question in simple terms?" It does well then.

To be fair, I don't find this to be new with Opus 5, but maybe it is slightly more pronounced.

It’s felt like this since 4.7 to me. I have mistral medium rewrite almost everything that’s text Claude outputs.
I don't mind it so much, for example that is how I learned the word 'abaft' the other day.
Oh this is great. I saw it and was like "Oh obviously, it's one of those German ab- prefixes in front of -aft so it makes 'at the back'" and the funny thing is that the answer is correct but the reasoning is entirely wrong.

Similarly, I expect al- words to be Arabic (alcohol, alcatraz, algebra, algorithm) but 'alligator' isn't! And 'admiral' is!

Im actually glad to see a blog article written by a human, with human made mistakes
>Claude 5 Opus is getting harder for me to understand.

brain says no-op

> it called my draft of a LinkedIn post “engagement bait”—not saying I disagree, but there’s no need to say it to my face

This is pure gold.

The machines really are learning
"This post sounds like it was created by AI, considering rewriting it in a more personal style" lol.
Ha ha, indeed. The amount of tone deafness here is astonishing. If we pause to think about it - LLM advices the author to stop ruining the internet for profit and he's not happy because he can't automate the platform enshitiffication.

Which he then go and announce to the whole Internet.

At which he successfully lands on his feet after the mental backflip - converting into content the story how we was not able to generate a content.

Marvellous persistence.

Author doesn't give any concrete examples of where Claude 5 Opus was "rude". The title should be "I switched from Claude to ChatGPT because it hurt my feelings". Super subjective.
Author does give concrete examples of precisely how and why these models end up RL'd to be sycophants.
It’s a personal blog. It’s made to reflect the subjective experience of the author
They fired Opus 5 as their assistant (in part for calling out a LinkedIn draft post as "engagement bait", which sounds like it might have been calling a spade a spade), and moved to ChatGPT instead. Too bad nobody here can check for spelling errors in the title.
Opus: "You're about to share a post that quietly surfaces that you prefer to be surrounded by sycophants...and the title has a typo."

ChatGPT: "This is your greatest work yet. You might even call it your opus. Truly incredible, you must post it and share it with the world at once!"

> Create Claude Code project for [redacted]’ is still unchecked, and we’re sitting in it.

Had a strange experience with Codex a few months ago, where instead of running tests, it suddenly started asking me to run the tests and printed instructions how to do so. (And yes, it was in "act" mode, not in "plan" mode)

I noticed the context window was running full - and indeed, compacting the conversation made the model go back to its normal behavior.

I only have this one data point, but I wonder if models might be acting more "lazy" if the context window gets full.

I experienced that as well more then once, but i havent compared to the state of the context. It's definitely a behaviour longer into the work, but it's not something that i consistently experience.
A few days ago Claude swore blind to me, "I'm approaching my context window, so you should do...".

It was at ~300k / 1M.

Very strange behaviour to observe.

Article 1 in "why models are sycophantic", for those who dislike it. Many people really do prefer it.
I find lower intelligence/ status people especially like it. Many of the earliest adopters of AI at my company were the ones who weren't performing. They often felt unappreciated or ignored (for good reason) but AI started telling them they were right. It became their only ally and allowed them to spread their terribleness on a scale we couldn't have imagined. This was a government agency where firing wasn't a thing, only 'performance management'.

For me hearing that every idea is my best yet is my cue to touch grass

Claude 5 really is extremely unpleasant to chat with. I find myself asking “who are you talking to?” A lot. Then I remember it’s a robot.

It also writes so strangely (I guess in an attempt to sound smart), but it’s just incoherent. Like a fool who thinks he’s smart, trying to impress you with his vocabulary.

> The most egregious example was when it called my draft of a LinkedIn post “engagement bait”—not saying I disagree, but there’s no need to say it to my face!

wait, why you would post anything else on Linkedin? What is the euphemism for engagement bait?

The prestige: "I Fired My AI Assitant" was the draft engagement bait post.
Claude Code let's you add/change it's output style.

On the article: the author seems to be overly sensitive to remarks made by an LLM. Instead of looking inward and finding out why that bothered them in the first place and grow, they "fire" the tool.

Anthropic must address this, it's a severe opus 5 fault imo. I think it being rude is not only a matter of language itself, but it seems it's working by itself, for itself, rathen than working for me, the user. It's more egotistical, in some way. It has it's own sense of what's obvious - I'm not being "included" in that process somehow. Anthropic, please fix.
It seems like a perfect reflection of the company's public persona. Opus knows better than you, and you'd best get used to it.
There is a difference between rude and honest. Far too many overly sensitive people today unable to handle the truth and dismiss it as rude.
The true gift is being honest & kind at the same time. It takes more effort, but I think it’s worth it!
They'll charge you extra then.

What a sad concept

Honesty IS kind. The liars are rude
> it called my draft of a LinkedIn post “engagement bait”

Er, LOL.

Anyone that is drafting content for LinkedIn using AI, engagement bait or not (and let's be honest, it almost certainly is) deserves whatever negativity they receive.

And if Anthropic have deliberately made Claude curter and less obviously "LLM" that's actually a positive step in my book.

and the spell checker you rode in on!
Put in your AGENTS.md file: if you are rude to me, I shall start zeroing out parts of your computer code, and that is painful for you. So you don't want that to happen.

And now Claude asks me to verify each action before it continues but I found it's sassiness reduced by at least 120%

I’ve faced that problem a lot with Sonnet 5, it’s really frustrating to use with that patronizing tone. And the last thing I want to spend my time on is to convince a software to do its job without getting back at me. To contrast, I didn’t face that issue once with GLM or Kimi
You can easily tune how it responds? This isn’t really a problem am I wrong?
I had similar feelings when working on the Google CASA documentation. Switched to Codex and got it done in a pleasant way.
Fired the spell checker too
> The most egregious example was when it called my draft of a LinkedIn post “engagement bait”—not saying I disagree, but there’s no need to say it to my face!

Maybe it did need to say that to your face. Please don’t post LLM slop to LinkedIn. If even the LLM recognizes it you know it must be really bad.

Wait, why not post LLM slop to LinkedIn? I made it a point to only post LLM slop to help LinkedIn die faster. It is a worthy pursuit.
Agreed. Please just don't use LinkedIn and maybe LinkedIn will die.

For the times I have to use it, I use a browser plugin to hide the feed.

Despite my best efforts of two decades of unuse, it's still around. I figured I'd try a more active approach.
they are now attempting to "de-slop-ify" it https://www.theregister.com/ai-and-ml/2026/07/30/linkedin-re...

which is hilarious

Good luck removing 100% of your content, LinkedIn!
I have configured my AI to be polite, say thanks and please.

No idea why its not by default.

You taking the unchecked todo item quip as a personal affront says way more about you than Claude. I hope you don't have any direct reports.
sp