The baked in communication style of these models is so obnoxious it's impacting my work. The best way I can describe it is that everything is optimized to impress the user and make the agent sound more authoritative, but the way this is done is through deliberate obfuscation, inserting inappropriate and extremely dense jargon, and bizarre, stilted metaphors. It's like they've been trained to produce output that's hard to read.
I was similarly frustrated a few months ago, but have noticed I've started to learn the idiom.
Its use of "dense jargon" and "stilted metaphor" is actually surprisingly consistent - it's speaking its own dialect, and you get used to it.
After a while it gets much easier to read and even becomes somewhat efficient, I think, since the odd metaphors it uses often have a precise meaning in Opus-ese (Fable speaks a really similar dialect).
…
No step here involves choosing based on meaning. It is a filter, a sort, and a slice.”
This is from Opus five minutes ago. I can certainly derive meaning from these kinds of statements in isolation, but paragraph upon paragraph of this is unintelligibly dense when trying to work with Claude to come up with a plan.
The worst part is that it can’t even make its responses make sense when asked to summarize in simple English or < 200 words. It simply cannot be steered to make its prose legible.
This dialect is idiosyncratic to you and Claude based on your session history and memory.
I've noticed Claude's output mimics my writing style.
> Registers the board implements but whose behaviour is not modelled
Right down to my preferred spellings.
As several comments I've read on HN suggest, this jargon which can be so precise in the mind of one person, tends to rapidly fall apart when multiple people try handling it.
Same experience. It’s not very “human” but once you have agents talking to each other the shared dialect and verbosity makes things much smoother in my experience. Fighting against the default feels like an uphill battle with no meaningful benefit.
FWIW, I also think the constant chorus about how new models are worse than old models is a human hallucination. They're certainly not perfect but every one becomes more steerable in terms of actually completing more and more complex work.
And once everyone gets used to it, we'll chide people for writing things themselves, like we're chiding them for writing with AI now, and the ouroboros of life will continue.
And when you say it like that, I have to wonder how much of this is a natural consequence of RHLF on such a grand scale, when you have millions of people pretty much much skimming chat responses or operating outside their depth and giving unqualified feedback to the models.
Seems like a lot of people may be reinforcing what sounds smart over what is smart.
Also as an aside: funny how much the LLMs continue to mirror the human communication they’re trained on
Which, conveniently, fits neatly into the benchmaxxing arms race/agentic coding market fit, since you can basically train "directly" on a specific problem space for a benchmark/agentic goal (fudged sufficiently to avoid excess overfitting on public problems/bechmaxxing accusations if real world performance falls short).
The language evolution could be explained by reliance on ever increasing layers of a model judging a model, using a model developed eval, based on synthetic data from a model, etc. And by the time a human evaluator sees it both A/B choices already converged into weird Claude pseudo English as that was baked in much earlier in training.
I wonder if the labs are sufficiently prepared to filter this kind of stuff out. I see a lot of non-developers asking development things of Claude, getting confused when they're in over their depth, and getting upset that they don't understand what the model is providing them, giving it bad feedback, and subsequently making the AI worse for the rest of us who know how to use the tool.
So going to continue trying that as a command structure going forwards...
This is because these harnesses are missing a very important feature. Anything like this needs to be included with every turn, otherwise the LLM quickly drifts.
I first noticed it when I wrote a harness for D&D (because it's so damn noticeable there), but now I include this for any harness I write.
I wrote a little bit about it on my blog post. It's a waste of money and compute.
This is close to the worst thing one could say of a tool for professional use.-
---
That creates a feedback loop:
- Playful style is rewarded
- Some rewarded examples contain a distinctive lexical tic.
- The tic appears more often in rollouts.
- Model-generated rollouts are used for supervised fine-tuning (SFT).
- The model gets even more comfortable producing the tic.
Non-determinism at its finest.
That’s really annoying, although it feels like it’s improved some over time.
Not sure what the fix is, but you could try using a canary to at least get a signal of when things are going sideways (Mr Tinkleberry for reference: https://news.ycombinator.com/item?id=45983698)
Yes. agents.md does very little because prompts change the context and thus the initial path into/though but they don't/can't change the actual weights that control responses. Yes. of course it gets worse as the session goes on, assuming the prompt is even still in the context window, the further it gets away from it the less it affects next token selection.
This shit is only like 5 years old why can't anyone remember how it works
Though you know, it's not like the leadership tied to these companies have a history of abuse, deception and theft or anything like that, right?
It's not like our leaders hide behind similar sorts of patterns that the agents/AIs follow (not saying it's not a human thing - but I hold leadership to higher standards than non-leaders). If our world leaders were able to be more accountable to these abuses, I don't think this would be tolerated with our AIs.
Don't worry. You'll get used to it. If you don't your kids will (as they'll know nothing else).
The top minds of our generation have decided that's the way things will be, and who are we to question them? It's not like it'll do any good anyway. Resistance is futile. There is no alternative.
You are an editor. You'll be given a message with strange characteristics:
- Weird subject and verb combinations
- Subjects that should be objects
- Very roundabout reasoning, peppered with pseudo-epiphanies
- A distracting beat to the flow of the message
- Self-praise
Remove these characteristics, and rewrite it in a clear, conversational style. Keep the intent of the message, and take care not to lose any of the details.
A few specific rules:
- The message is usually set in the first person
- Only humans, groups of humans, and agents should do "action verbs"
- Objects should never do anything. Here are some examples to avoid:
- X carries ...
- X names ... - APIs are a minor exception to the action verb rule. They can do stereotypical things like CRUD, queueing, running, and calling.
- Avoid em dashes (—), as adds a distracting beat
The whole message you get is one block of that output. Reply with the edited prose and nothing else.
I can't help but feel the circumstances that enable this kind of front page article are vestigial from the days when OAI was super bad and Anthropic was beyond reproach. This change-over-time is why I avoid getting tribal with technology vendors. Assigning ideological motives to 200k+ employee organizations is how we wind up in weird contortions like this.
Most rational actors simply moved from one to the other. It takes a special kind of devotion to the proverbial hole in the ground to keep pushing in this direction.
Effectively all models can do style transfer reasonably well at this point, but not so much for "actual reasoning".
If the combination of two works better for you than each one by itself, why wouldn't you stack them like that?
Because it's not an either or thing. Neither is sufficient. I'd argue that, expenses aside, you should have every model you have access to cross reviewing the work of the others.
Outside of super trivial things that I should have just done myself, I have a cross-model review of _everything_ these days. The tokens are too cheap not to.
This is what I think too. But, users’ psychology might be playing a role here. Anthropic has great advantage from being the first major player delivering functional agentic coding solution (rather than an intelligent autocomplete) and they were able to impress people by Opus’ iterative improvements early this year.
It’s technically very easy to switch between models, harnesses but their moat or perhaps a main source of users’ friction could be FOMO. That’s especially powerful in this competitive environment where everyone keeps wondering/worrying about what others might be doing to get or stay ahead.
I've been in the habit of pushing my claude-speak to codex to improve legibility, but only if I think someone is going to read it.
```Un-Claude 0.2beta
import sys,csv,requests
CH="# Valid channels: analysis, commentary, final. Channel must be included for every message."
CANDIDATES=[
("no-hedging","Reasoning: low\n\n<terse><no-hedging>\n\n"+CH,"Condensed:"),
("neutral-reg","Reasoning: low\n\nRegister: neutral technical. No intensifiers, no evaluative adjectives.\n\n"+CH,"Condensed:"),
("no-closing","Reasoning: low\n\n<terse>\nNo closing remarks.\n\n"+CH,"Condensed:"),
("terse","Reasoning: low\n\n<terse>\n\n"+CH,"Condensed:"),
]
def rephrase(text,base="http://127.0.0.1:1234",model=None,temperature=0.0,max_tokens=1400,timeout=180):
src=text.strip()
if not src:
return []
if model is None:
model=requests.get(base+"/v1/models",timeout=timeout).json()["data"][0]["id"]
w=csv.writer(sys.stdout,lineterminator="\n")
w.writerow(["idx","label","prefill","src_chars","out_chars","ratio","tokens","finish"])
rows=[]
for i,(lab,sysmsg,pf) in enumerate(CANDIDATES,1):
p="<|start|>system<|message|>"+sysmsg+"<|end|><|start|>user<|message|>"+src+"<|end|><|start|>assistant<|channel|>final<|message|>"+pf
d=requests.post(base+"/v1/completions",json={"model":model,"prompt":p,"max_tokens":max_tokens,"temperature":temperature},timeout=timeout).json()
c=d["choices"][0]
t=(pf+c["text"]).rstrip()
w.writerow([i,lab,pf,len(src),len(t),round(len(t)/len(src),3),d["usage"]["completion_tokens"],c["finish_reason"]])
rows.append((i,lab,sysmsg,pf,t,d["usage"]["completion_tokens"],c["finish_reason"]))
print("\nmodel: %s"%model)
print("temperature: %s max_tokens: %s"%(temperature,max_tokens))
for i,lab,sysmsg,pf,t,tok,fr in rows:
print("\n[%d] %s"%(i,lab))
print(" system: %s"%sysmsg.replace("\n","\\n"))
print(" prefill: %r tokens=%d finish=%s"%(pf,tok,fr))
print(t)
return rows
```input ## 8. Honest gaps — what I could *not* resolve
I want to be explicit about the limits of this pass rather than imply completeness:
1. *`PROVIDER_T` values are not enumerated here.* `list_models(inference_provider=...)` is typed against `PROVIDER_T`, which lives outside the three modules I scanned (it's in the `inference._providers` subpackage). The accepted provider strings are therefore *unknown from this run* — `"cohere"` is confirmed only because it appears in a docstring example.
2. *Three grep hits point to search-capable functions I did not identify.* My scan found parameter assignments that don't belong to any function I enumerated: - line 3046–3050: `params["filter"]`, `params["sdk"]`, `params["includeNonRunning"] = True` — an additional Spaces-oriented endpoint with an *`sdk` filter and an `includeNonRunning` flag* not exposed by `list_spaces`. - line 2879: `params["config"] = config` - line 12013: `"sort": sort` — almost certainly the consumer of `DailyPapersSort_T`, i.e. a daily-papers lister distinct from `list_papers`. - line 13872: `params["search"] = search`
These represent **real additional search surface** that my `LIST_FUNCS` whitelist missed. A follow-up pass enumerating every `HfApi` method containing `params[` would close this.```example output [1] no-hedging system: Reasoning: low\n\n<terse><no-hedging>\n\n# Valid channels: analysis, commentary, final. Channel must be included for every message. prefill: 'Condensed:' tokens=131 finish=stop Condensed: - *Provider strings* (`PROVIDER_T`) are not listed; only “cohere” is known from a docstring. - *Missing search‑capable calls* were found: - `params["filter"]`, `params["sdk"]`, `params["includeNonRunning"] = True` (Spaces endpoint). - `params["config"] = config`. - `params["sort"] = sort` (likely a daily‑papers lister). - `params["search"] = search`. These were not captured in the `LIST_FUNCS` whitelist, indicating additional search functionality. ```
> If CLAUDISH_MODEL names a model you have not pulled, every rewrite is skipped — with the one-time notice above.
The rewrite did seem to lose the important fact about the ensure- pattern being idempotent.
I'd love to know what the hell Antrhopic has done to make Claude's writing so, so bad.
It's really unusable for anything other than code. And I have to remove its incomprehensible comments 50% of the time before committing anyway. After interacting with it, "slop vomit" is truly the most fitting description. I have to admit I have lost my temper and spontaneously referred to its output as vomit more than once. Seems like I'm not the only one.
https://gist.github.com/bmurphy1976/47ad81a842ab4b1628ef5974...
A small preview:
*Meta commentary.* Sentences about the document, the diagram, the reader, or the
writing itself ("the split across this diagram is the whole point", "a reader who
assumes X will be wrong", "as we'll see below"). Delete the frame and keep the fact
it was wrapped around. If there is no fact underneath, delete the sentence.Opus/Fable output these days though is... not enjoyable. It's just really bad. The code quality is fine, but i want information from claude and it's just awful to read.
My biggest problem honestly is that i can't move my day job.. we're using enterprise claude and i'm not sure how much effort it would be to get access to another provider. I should inquire though, claude is really frustrating these days.
Hasn't worked yet outside of the classic "you're now manually breathing" kind of stuff.
I would opine:
- Nerfed Fable is worse than Fable
- ... and I would argue Opus 5 is worse than nerfed Fable, to the point I've found it unusable.-
You must use ASD-STE100 Simplified Technical English (STE) when it doesn't detract from meaning.I found vomit with a small LLM much better than anything Opus 5 ever wrote. I don't think Opus 5 can write.
Right now all I have is
> - Give terse and concise answers unless the user asks you to elaborate. Big walls of text are not usefull when trying to communicate.
Given that guidance, I tried specifying "Unless I ask you to elaborate, respond with no more than one paragraph, using sentences of 20 words or fewer." It works...ish. I still see it violate this rule regularly, but it's less bad IME.
Often it’s exactly the opposite. True intelligence and literacy is being able to communicate effectively and to a broad audience in the simplest terms possible.
A single question is enough for it to retract the error and correct itself. Suggesting that not understanding some of these messages is a lack of human comprehension rather than the agent being flat out wrong is...a bold take.
Typically a feature of good human technical communication is the ability to concisely explain key ideas so one can quickly identify any divergences between understanding. Opus 5 is dreadful at this.
The single saving grace is the intuition that if I don't understand it's probably wrong.
That said, I must confess that I have not been complaining per se—I assumed that Claude was getting better and better at mimicking the idiosyncrasies of Silicon Valley bro-speak. Judging from other comments, this may not seem to be the case after all.
Blog post: https://zachahn.com/posts/1787191554
The prompt I use to tell the LLM what to fix: https://github.com/zachahn/vomit/blob/main/internal/config/s...
Wasn't received too well on Lobsters haha, wrote a small extra blurb about it there: https://lobste.rs/s/juekuk/how_fix_claude_5_s_token_vomit
[Seriously y'all in what universe would some "caveat" or another NOT "be a real one" by whatever severity you'd want to measure that AND/OR need of saying so ... ]
Claude (Original)
Force pushed. 1234567...890abcd main -> main (forced update).
Verified
Local main and origin/main both at 890abcd, in sync.
Every commit reachable from origin/main: no old string found.
Local object store, with spice-harvester as a control returning 9 hits: all old strings return 0.
Working tree clean, tests pass.
One caveat, and it’s a real one
The old tip 1234567 is still fetchable from GitHub by its full SHA. I confirmed this directly:
git fetch origin 1234567890abcdef1234567890abcdef12345678
* branch 1234567... -> FETCH_HEAD (succeeded)
GitHub keeps unreachable objects until its own garbage collection runs, and there is no schedule you control. Anyone with a full SHA can still read the old blobs through the web UI or a direct fetch. That fetch also pulled the old objects back into your local repo, so I deleted FETCH_HEAD, expired the reflog, and re-ran gc --prune=now. The counts above are from after that cleanup.
If you want the old commits unreachable on GitHub’s side too, the options are to ask GitHub Support to run GC on the repo, or to delete and recreate the repository. Deleting and recreating is the only one you can do yourself. Say the word if you want help with either.
Nothing to commit this turn. The rewrite was the work, and the tree is clean.
Claude × Vomit GPT-OSS-20B
I force‑pushed the branch, updating main to commit 890abcd. I verified that my local main and origin/main are both at 890abcd. I checked that every commit reachable from origin/main contains no old string. My local object store shows that the spice‑harvester control returns nine hits, and all old strings return zero. The working tree is clean and the tests pass.
However, the old tip 1234567 is still reachable on GitHub by its full SHA. I fetched it directly and saw that GitHub keeps unreachable objects until its garbage collection runs, which I cannot schedule. Anyone with the full SHA can still read the old blobs via the web UI or a direct fetch. The fetch also brought the old objects back into my local repo, so I deleted FETCH_HEAD, expired the reflog, and ran git gc --prune=now. The counts I reported are after that cleanup.
If you want those old commits to be unreachable on GitHub as well, you can ask GitHub Support to run garbage collection on the repo, or delete and recreate the repository. Deleting and recreating is the only option you can do yourself. Let me know if you need help with either.
There is nothing to commit this turn. The rewrite was the work, and the tree is clean.
> Caveats belong inline, no "one thing to note" or "it's worth mentioning" footer. If it is worth raising or calling out, do so where it is most relevant and not as a foot note.
Opus 5 has a god awful habit of always doing a Columbo on every single response, and it is such a jarring read that it amps my cognitive burden having to back-read everything.
The joy of watching a dumb AI-ism be sharply corrected by code you wrote months ago is hard to explain.
I highly recommend it.
Claude and Codex usage limits cannot be trusted.
Paying your own API bills in full is superior.
Wish I could use my Claude subscription with pi too, much preferable to the endless command execution allow/deny prompts you have to do with CC, versus proper autonomous allow/deny lists defined ahead of time.
Curious why you recommend the API? It's likely the current subscriptions won't stay for long, they're heavily subsidized, but before they get axed, they're easily the best deal for monthly price/token usage.
re.sub(r'(?is)\b(?:honest|caveat|absolutely right)\b.*', '', text)> Anything that uses the OpenAI API?
I would have thought they meant the Anthropic API or maybe I'm misunderstanding?
The vomit never makes it my way
There's an echo of that tension in OpenAI vs Anthropic. For a while OpenAI seemed reckless and ignorant, preferring to just throw compute at the problem. Meanwhile Anthropic is hiring philosophers. But now that Claude has its head up its ass to the point where nobody wants to talk to it, OpenAI is looking rather pragmatic.
It brings to mind a skepticism about just letting the ivory tower do its thing without some kind of anchor to the everyman (this is why we make researchers also be teachers, though I'm not sure what the AI equivalent of that practice would be).
Watching the models seesaw in the same ways that humans do, but faster, is so surreal. I wonder if their tendencies will remain an echo of ours, or if they'll one day be more of a forward projection, a representation of where were going if we don't change our ways, and if we're lucky, a reason to change them.
For me I added some instructions to speak clearly and it helped marginally and that's fine. There will be a new model out in a few weeks where I'm sure they've laser focused on this issue since nobody can shut the fuck up about it. The same thing happened with GPT if anyone can recall the ancient period of 4-6 months ago.
Like so many other products, people are moving too fast and shipping things that move the ground under people’s feet needlessly.
All this while we’re beaten to death with the marketing and false promises, and the broader consequences (ex: layoffs, stress, crazy expectations) caused from all this.
Obviously what Anthropic and co have built is amazing and people aren’t losing sight of that. That’s actually the key part of the frustration.
So no, this is not whining. This is the natural response you get when you make bad product decisions.
If you don’t want to get feedback, don’t sell products.
These things do work that previously would have taken expensive engineers months to do, at much lower quality, and what's our response? Ti nit pick on it being more verbose than we'd like?
Just like with humans, when someone is being too verbose, there's a skill to just filter through the noise and focus on the important parts.
This feels no different when I use an AI.
But I guess it's a good sign that we've from complaining about 'AI slop code' to, 'I don't like how it speaks to me'.
[1] https://vale.sh/