It has some vintage. Kubernetes deployments (containers) may have sidecars.

Depends what you mean by "upstream". It would probably be inefficient to force the model to use more human-readable wording in its internal thinking traces, while translating a block of text at the end is a pretty trivial task for an LLM.

I already pipped Opus 5. That thing is a tool.

Sadly pretty uninteresting, you can probably just launch Claude code on the repo to see by yourself. One cluster is french and spanish, another is about design, another about frontend, etc...

Author here, thank you so much! I really tried to make it nice to use, beyond the (quite original) modelling.

A prototype I did tried to detect some grammatical constructions, eg "it's not ..., it's ...", but I am not sure how to systematize that.

Also just a disclaimer: I am NOT tracking Claude tics, I am merely finding that a particular cluster of vocabulary increases. Tracking Claude requires labelled data IMO. I tried using model release dates in a structural model to constraint the clusters but the result was not compelling, so I ended up simplifying the model a lot!

I tend to agree. Claude's language doesn't bother me that much, because even the lamest cliches are load-bearing to an extent (so to speak.) But some of the examples I've seen others post are well worth complaining about.

Probably because of the dataset. I don't see sorry or apologize either. They wouldn't be in code or commit messages.

No “long pole” either.

Thank you so much! The presentation was really my goal here, more than the model itself.

> the argument effectively ‘makes itself’

What argument? I don't know what to take away other than "Claude likes certain words". Some of them are kind of amusing, but I'm not convinced the vocabulary is bad or that this is a problem, just from looking at this.

For me, there are two reasons: it’s what my work provides, and for some tasks, I want the absolute best model for coding.

Having said that, I just subscribed to ChatGPT yesterday, as I've become impatient with Claude for a text-dense project I'm working on.

Don't much care for the misdeeds of the alternative

Work only allows the tools they buy. So at least for me, I am stuck with Claude. At home for personal projects I am a fan of DeepSeek v4 Flash.

Claude is all we have at $work, if you don’t count MS Copilot (and you shouldn’t count MS Copilot).

Imagine being “incentivized” to aggressively use a tool for your job, and that tool produces thousands of lines of text in Olde English which you need. You’d be griping too, methinks.

So you're saying you're safe for when the machines take over? :|

This kinda dissonance is uncomfortable. How do people reconcile it?

I think these are more like RL tics caused by over zealous alignment towards specific goals, not all of which are to our benefit. A lot of the language used by Claude now is excusing of responsibility and inducing it to exit loops of work early and sit idle. This, IMO, is a naked attempt to offload load by quieting the models early and escaping from clear work to do. It’s gotten so bad that opus 5 loop escapes even as it claims it’s about to do something. I’ve mandated by engineering teams switch back to opus 4-8. Whatever frontier problem opus 5 excels at is so obscured by its inability to achieve any goal successfully without enormous amounts of hand holding that it feels like regressing to 2024 models.

The florid over exaggeration do certain words in bizarre ways is a reflection of their aggressive alignment towards too many goals, leading to weirdness in both behavior and language. The alignment functionally lobotomized opus-5 for any practical task.

Anthropic had a real gem in 4-6 and managed a near total market capture, which they have since squandered in the fastest burning of developer good will I’ve ever seen. It feels like exceeding the unity licensing implosion but without the single stupid decision.

That's the problem for me: "semantically dense". Sometimes a single paragraph is filled with several of this expression, one after the other, and there's a real mental effort to get what Claude is saying.

Generally the claude jargon is valid jargon I’ve seen real people use, it just uses jargon so much more often than a human would. And because it has particular jargon it likes, and claude is widely used, you see so much of that particular jargon that you get sick of it.

It’s like having one coworker with a very particular writing style which is mildly annoying, but then it suddenly feels like half the internet was written by that one person and it becomes a lot more annoying.

It might be, when used sparingly, but when it's the first sentence in a wall of text as Claude goes way off the deep end on a two page description it becomes the easiest readily available tell that you're about to be frustrated.

Humans are very good at pattern recognition - Claude is _incredibly_ repetitive in the way it starts to struggle to communicate. I think there's also a ton of overlap in the Jargon instead of Usefulness that developers see in annoying middle management/salespeople. Circle back, synergy blah blah.

I don't think the individual turns of phrase are inherently problematic - but the process is triggering.

ItMs because Claude sprinkles these words as flavoring without aiding understanding. It feels like Claude thinks of metaphors that don’t actually mean anything (or maybe only makes sense to itself).

That was the intention behind my title!

We are seeing more of their "thinking". Lowering the refinement of the output to get closer to profitability. The nature of the LLM is that it generates huge amounts of text, then it iterates them down into a compact, hopefully accurate prose. That refinement is the really hard part and computationally costly.

It's likely/It could be an effect of more reinforcement learning in training compared to earlier. You need loots of RL to learn to code well.

The going conclusion is they’re getting models ready to talk to other agents, not people.

They are increasingly being trained on generated tasks and even (parts) of the pre-training data is 'distilled' (e.g. Clibmix as an open-source example), so there are many ways in which the vocabulary can seep into the model.

That's most likely what's happening. SNR will constantly decrease as LLM content is so much quicker and cheaper to generate, which makes it more statistically significant, which will make it more "relevant" for future models. A positive feedback basically.

I thought about that, I might add it if I can find a nice design!

Chrome's `find` finds it for me ;)

Is it actually better though? Can you even measure it? Unguided, Claude will just produce mediocrity which by its own existence becomes almost worthless, and the newer versions become increasingly difficult to guide while also being obnoxious to read. If I did not have to interact with it, it would not be a big deal, but then also I could not get value out of it.

Claude usually rates its own writing very highly, which is a problem when trying to make it write differently.

Thank you all for the feedback!

I added the search bar, increased the data to 1000 PR a day (more than 50M words total) and added a feature to explore the other clusters as well. I hope the page is not getting cluttered.

Great work, thanks for sharing!

The design is really impressive. Beautiful and dense, yet very understandable. How did you come up with this particular design? It's really nice and everything fits, the colors, the layout, all of it.

Were you inspired by anything in particular? I feel like this reminds me of something, but I don't remember what exactly.

Nice work. I would suggest rewriting your README with Gemini, the most human of them, so it doesn't read exactly like the thing you are pointing at: https://github.com/louisabraham/load-bearing

Very cool! I'm trying to understand the graph, so the bottom-most section seems to be the cluster identifying Claude written PRs. What are the other 7, any reason there are 8 in total?

I've been scraping instagram posts recently to identify AI misinformation accounts that all repost each other's carousels and get hundreds of thousands of likes in engagement. Thinking of ways to present it and your dashboard looks very helpful. Did you experiment with any other types of visualizations before deciding on a stacked area chart for the clusters?

I haven't noticed GPT 5.6 having any obvious jargon.

I swear claude took a detour recently, its written output has been nearly incomprehensible to me. At first i thought i was getting AI-brained and just lost critical thinking but as i dug into response after response its was just the most obtuse language to explain what was going on. Really mentally taxing to wade through it all day.

"shipped" was a pretty common term before AI though. It does show as 17x more common on github in this dataset but it was used a lot more in product management than in PRs previously.

Also can't believe vacuous is not there.

Im surprised by "provenance." I see it all the time. Except I've brought this up before and have never had it corroborated. I'm starting to think it's just my Claude.

I'm also missing the "latch" that "wedged" my test run.

> I wonder to what extent this is the result of suboptimal RLHF versus the inherent intelligence of the model making its language more intricate and difficult for humans to easily parse?

Its not the latter; its just excessively verbose wirh awkward word choices, the same as many poor writers. (And, like many such writers, the particular bad choices fall into recognizable, regularly recurring patterns.)

My unprovable pet theory is that, especially for writing about technical topics, the RL process has an open-ended way to weight things for quality: textbooks and first-party docs preferred to old stackoverflow answers and obscure blog/forum posts, and so on. The open-endedness of that quality gradient results in something in the RL process (maybe HF, maybe not) massively over-weighting some very small corpus of “quality = near infinite” content. The distribution of quality scores that inform the degree to which RL affects output has some extremely influential outliers, in other words.

Whatever that small corpus is, it contains some very specific grammatical tics, and that’s how we get Claudish.

Anyone who thinks a company/project as big as Anthropic/Claude wouldn’t make such a big mistake should take a look at how Azure cross-account federated login used to work.

I've been thinking more about how 99.9% of us don't have the experience of someone significantly more intelligent, yet also subservient working under us, which is why I keep going crazy second guessing whether Claude is spouting RLHF'd bullshit that sort of resembles English, or is genuinely (pun not intended) just better at "intuiting" things I'm working on, leading to its language.

A notable exception would be people like CEOs and managers higher up in big tech, who might be used to skilled engineers and domain experts reporting to them in unfamiliar lingo. Maybe that's why we don't hear as much on the everyday annoyances of Claude's language from that camp?

The social graph proximity of Rationalists to Anthropic will be lost on no one who reads Astralcodexten. So guess which website has served as the thickest reservoir of 'Claude-isms'.

It's easy to think "it's not talking down, because I don't understand it, and I'm intelligent". But how is less intelligent being supposed to fully understand a more intelligent one, honestly speaking? All I know is that Claude understands Claude perfectly. I have the common session pause/resume setup that sometimes produces completely incomprehensible markdown files, but a new Claude session picks them up perfectly, down to the smallest details. What if what we consider excessive circular gibberish is actually highly precise set of instructions needed to minimize error cases for that unreliable human?

If Claude understands Claude, Claude understands human, and human doesn't understand Claude, that doesn't argue well for "Claude is a caveman".

Imo their language is not precise enough for their intelligence to be the reason when it's difficult to understand. Maybe I'm prompting wrong, but when I don't understand, it's almost always because they just mash together words from context that don't form sentences with a clear meaning.

I think it's model collapse - excessive feedback and excessive RL.

What RL does is narrow the variety generated by the model by steering the output towards the goal being rewarded. It's a bit like putting blinkers on a horse.

Of course RL is a very crude tool - it affects the entire model, even if you are just trying to make it better at some specific task(s), or trying to imbue a certain kind of personality (OpenAI's recent goblin problem).

I don’t think they’re “talking down”. If anything - it’s way more difficult to distill something into a genuinely easy to digest format. I personally think that they aren’t immediately capable of this, and so we get word salad instead. Extra prompting required to strip extraneous prose out.

Maybe I am dumb and it IS talking down to me, but there have been many occasions where I’m reading AI generated docs / plans and it makes absolutely no sense, but looks really in depth at a glance.

Everyone has their own distinct mix of jargon and uses some words more or less than others, but now it's as if you have 1000 coworkers who are all the same person, and talk the exact same way, and it just so happens that their writing is grating to read. It also doesn't stop when you log off for the day. The same guy follows you home in ad copy, message board comments, spam, and in some cases even the correspondence of your loved ones.

The jargon isnt even the worst part. The phrasing is just downright incomprehensible sometimes. And its not even "academic". It's just extremely awkward, bad writing. And it is much worse than say Opus 4.6-8.

Actually I have found the copy that Claude Design spits out is way better than using the same model directly. I have no idea why. It has its flaws but it sounds like it's written by a human who uses derivative language. But usually the models just soudn incoherent.

I like to imagine that there's this one employee that oversees RHLF, who has a particular style of writing, and that got so ingrained that it's just them x1000.

I've found myself rewriting my own words to avoid Claudisms (because they can be cringe and I don't want people to think I'm copy/pasting Claude output at them).

Saying "the language of LLMs", while technically correct, is not exactly precise. It should really be the language of "AI agents", or "chatbots". OG LLMs would just mimic the style of its context based off of (pre)training from text on the Internet.

The kind of quirks you see came from crowd-sourced human-in-the-loop fine-tuning, with not very good work conditions or level of qualification (so resulting in "what non-writers thought good writing looked like", before people had developed the flair to detect these patterns) as well as feedback loops during agentic reinforcement learning and RLVR.

Author here :)

Thank you for the compliment! I did spend a lot of time designing a nice experience on both desktop and mobile. Even the scrollbar to select words was non trivial as I wanted the words to be of different size, yet avoid flickering when scrolling!

In the enterprise, we are already adopting them. “Seam”, I term I’d never heard used before, is now not only our defacto way to describe the boundary between systems/workflows/components, it is encoded into our organizational roles and responsibilities descriptions.

I am the proud owner of several seams, and am considering giving them person-names. The empty space between my inventory APIs and their clients might be “Karen”.

I can’t say that Claude invented this; the same type of terminology cycling happens every few quarters based on what leadership is reading/being told by Gartner.

>I suspect, as we continue forward, humans will slowly start to adopt the language of LLMs

I'm already seeing it. A coworker said something like "<person> added the color to the ticket here" meaning that someone added details to a ticket.

I've started seeing Opus 5 talk about "hermetic testing" when it just means "unit testing", so I hope that doesn't catch on...

A spike is an early prototype that you're supposed to throw away after having figured out the real design.

I think it was subtly dissing you.

In some software development methodologies, "spike" is a task whose goal is figuring something out instead of delivering shippable code. https://agiledictionary.com/209/spike/

don't forget LLMs are great at translating between languages, and within the same language. depending on the problem it works on, it will often reach for terminology that tend to be more common or familiar within that problem set. which appears inscrutable, but there's many different ways to skin a cat. just remind it to translate it back to the terminology and subject matter you're already an expert in.

It's interesting that there must be a decision behind that, even if it's just appealing to the RLHF judges for some reason. Maybe there's an intention that if you cannot decipher what the chatbot is saying to you, you will have to ask and burn even more tokens.

Naively I would often expect it would talk to me about various niche topics like to a layman, which does occur about some topics an actual normal person would ask.

They just addressed this with "Output styles"

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

Sometimes I can't even tell if what it's saying actually makes any sense to someone who understands all the terms its using, or if it's just throwing together words in a way that only make sense to its own model of language.

Is it the obnoxious style that causes this?

I'm not even sure a PhD helps. It just overuses jargon that has NO meaning. Sometimes, it actually hand waves too much as well while trying to dumb down stuff for you.

I am not sure whether it's a consequence of learning to reason from its traces or some RLHF that trips it into using weird terms to sound smarter to the humans who rate it.