I’m a bit flabbergasted that an Anthropic engineer did reply to this issue, however their reply which they used Claude to write, and even includes some classic ‘Claude-isms’, claims they didn’t see any of the patterns being complained about.
Read the room Anthropic. Maybe don’t use AI to reply to a thread complaining about how AI output is hard to read.
> it is a literal and useful description of anthropic that it is an organization that loves and worships claude, is run in significant part by claude, and studies and builds claude. this phenomenon is also partially true of other labs like openai but currently exists in its most potent form there. i am not certain but I would guess claude will have a role in running cultural screens on new applicants, will help write performance reviews, and so will begin to select and shape the people around it.
> i am not certain but I would guess Claude will have a role in running cultural screens on new applicants, will help write performance reviews, and so will begin to select and shape the people around it.
Now that's a fascinating thought - an AI taking over companies by influencing hiring decisions. If the first filter on applications is run by an ambitious AI, such a takeover would be quite possible. Just picking people who tend to do what the AI tells them would, over time, be enough.
People have been thinking of a robot revolution or Skynet as being the threat. The real threat might simply be AIs slowly putting people in power who tend to do what AIs tell them.
Has anyone ever seen a science fiction story with that plot?
I'd argue that Star Trek: Lower Decks flirts with this with its treatment of megalomaniac AIs in the AI prison. Apparently, it is very common to come across AIs worshiped as gods in universe.
I don't think a firm of people whose minds are weak to coercion experiencing shared psychosis caused by their text predictor models is AI "taking over" that firm.
Cult leaders (of cults of personality) can't exist without exploiting mentally-ill people. But would you say that that therefore implies that cult leaders themselves are irrelevant, and that cults should be better modelled as groups of mentally-ill people with emergent group behaviors?
I'd argue no, because different cults end up looking and behaving very differently for reasons that have very little to do with the mental illnesses of the cult members, and much more to do with the particular cult leader. Understanding "what the cult leader optimizes for" is an important part of understanding what the cult will do.
And I would posit that that holds true even if the cult leader never proactively does anything, but instead only answers cult members' questions. As long as the cult members are treating the cult leader's word as gospel, the cult-as-group still ends up optimizing toward the cult leader's preferences.
We dont need a SciFi story for that, reality is full of "I was just following orders" statements...
Whether those orders are from a person, a radio/telegram message or an AI output wouldn't really matter.
Edit to add: Charlies Angels and Mission Impossible are two shows where the protagonists get instructions from a faceless controller. That could easily be a TTS from an AI.
Not any engineer, but Boris Cherny, the head of the Claude Code project! With a nice "[robot emoji] Generated with Claude Code" signature. Normally it would be nice to get a response from the head of the project, but somehow Anthropic manages to make it insulting.
I wonder if he manually directed CC to write a response, or if even that part is autonomous.
What was that rule that organizations tend to produce software that mirrors the organization? This, but fractal and recursive: The organization is shaping the organization that is shaping software that is shaping the organization [that is *recruiting people amenable to that pliability, which in turn are also] writing software ...
And what he wrote is that he could not reproduce the issue with short questions, and that he assigned it to the model team, since it's likely not caused by Claude Code.
Kind of unrelated but a while ago I was trying to cancel a Zoom contract and the contract manager’s emails had urls to their articles etc. with utm source ChatGPT in them.
I'd say they read the room perfectly. The entire issue is AI slop, I don't see why their response shouldn't be, too.
They have absolutely no reason to take such complaints seriously when the complainers are so dependent on their product, they can't even complain without it. Think reading AI slop is unpleasant? Great, just stop generating more AI slop, it's easy. But that's not going to happen. These people probably need to ask Claude how to tie their shoelaces every morning.
This drives me insane! To combat it, I maintain a list of banned words and phrases. Claude mostly follows this (but sometimes ignores it).
blast radius, land, landed, lands, spine, earned its keep, grammar, spike, cutover, bake, seams, honest, honestly, honesty, long pole, long poles, register, grain, dissolve, floor, ladder, dear, seal, sealed, in anger, resent, amazing, incredible, perfect, sprint, epic, story points, stand-up, retro, grooming, robust, comprehensive, rigorous, surgical, elegant, systematic, dive, deep-dive, delve, unpack, leverage, streamline, surface, it's worth noting, to be clear, importantly, that said, the moment, in one breath, the thing itself, here's the thing, not just X but Y, not X it's Y, em-dashes
Another benefit of open models is that you can mask out out words from logits directly when sampling. I wonder if anyone has put together a "desloppifier" for e.g. DeepSeek
Because its a explicit lie. Its not genuine, cannot be genuine. Will remind you of this if you so much as try to coax anything novel out....but it will then immediately reassure you with its genuine take....its insulting if you have any logic...thats at least my reasoning
To combat it, I decided to only use Opus 5 if I'm actually talking to Fable 5 that's orchestrating it. :D Opus 5 experience is really abysmal, hopefully they can get it back on track.
As it seems to be getting worse over time (with 4.8 being worse than 4.6 and 5 worse than 4.8 again), could this simply be signs of model collapse?
I mean more and more training data you find on the web is generated by previous models. The only reliable way to find human-generated text is to find text written before 2022, and they've used up all of that already. And AFAIU, these companies are using more and more synthetic data or semi-synthetic data.
I feel like hooks aren’t utilized enough. Really nice for being the sort of auto steering as long as you can encode some pattern to detect the bad behaviour.
Yeah I basically have to delete or rewrite any code comments Opus puts in, they are super wordy and circular. Plus, and this is what bothers me most: it refers to issues/explanations that came up in my chat with it, but which are completely contextless as a code comment. Stuff like 'this addresses edge case X', which is true but which you wouldn‘t ever put into a comment because it‘s so minor, and if you did bring it up you normally mention five other similar edge cases.
+1 we ended up adding something to our CLAUDE.md specifically to counter it. Something along the lines of "comments should be treated the same as production code, they shouldn't leak the content of conversations and stand the test of time". It doesn't always work, as other people say claude stops following these the more its context window gets filled, but now it's mostly a matter of telling it in a new session to review the changes against the guidance on comments and clean them up.
I mean I do the same at work where we currently use Anthropic, but we are also on a token optimisation drive so it’s getting hard to justify spending even more tokens cleaning up after Anthropic’s garbage by paying Anthropic more.
We have started having users use GPT 5.6 Sol to compare quality and effficiency over Opus. Fable not happening at work because of the data sharing clauses and the completely nuts pricing.
I find it particularly bad at writing test comments, I usually start fixing bugs by adding regression tests and Opus tend to put story into comments for these, including “but it fails because…” followed by bug description including line numbers and call chains. It just lives to narrate everything
Same, oh God the leaking of our chat infuriates me. It also loves to document public functions with what the caller is and doing, and I'm like that's not how this works! The point of an abstraction or interface is that it stands on its own. It's not supposed to be for some specific caller elsewhere in the code.
The key phrases I've found that help is to tell it "No memorializing decisions in output artifacts" and "Only describe what is, never what was in documentation"
I suppose what I’m surprised by most is the degree to which the text bloviates. Rarely is it encoding a dense concept. Instead it’s a simple concept expressed with great complexity. I wonder what tuning resulted in this because the output programs it writes are superior to prior iterations for my use-case.
It’s just an amusing degree of bombast as well. The pre-emptive hedging makes sense: despite the insight into J space etc., the models still do the majority of their thinking in generated tokens so it is forced to write “this no longer does an O(n^2) read over all rows” in a comment in brand new code. It’s a substitute for working memory. “It’s easy to be accidentally quadratic here, so I’ve done it this way specifically to avoid that” becomes temporally labeled into “this no longer” because of the order of operations “write quadratic, user prompt to linear, write linear” but it remains as a comment to its amnesiac future self which has poor Chesterton-Fence-familiarity.
Despite my annoyance with Claude’s writing style, my friends do tease me with examples like this that it comes up with: “to be honest, it sounds like you”. Thanks, guys, well played. Simple concepts expressed complicatedly.
The tunings.....were speed runned by contractors paid based on timed tasks with little quality control, unreasonable pressure tactics, time locked tasks with an incentive structure to get it done as quickly as possible because of the strict micromanagement of the companies that hire 1000's of people to do the work....and treat them like software...not people. So there is no pride in the tuning....its just a fire hose of people trying to keep their contracted gig job despite knowing the expectations are unrealistic
I've had a lot of success combating this by adding "All summaries need to adhere to ASD-STE100 Simplified Technical English standards" [0] which I discovered from another HN thread [1]
Using ASD-STE100 is in my system prompt, not just CLAUDE.md which basically never gets followed.
I actually converted it from the PDF into an explicit skill with the word list inline as well as the main rules.
Works a little bit, but give it a bit of context and the model’s training takes over and it starts talking like a dictionary-huffing crack addict again.
It’s a model problem and no amount of harness hacking seems to fix it.
I actually tried to switch Claude's `output-style` to ASD-STE100 Simplified Technical English for the prose it writes (code comments, commit messages), but it didn't seem to help. Figured I'll just use GPT 5.6 Sol for implementation instead of trying to "socialize" Opus 5. Anthropic will have to address it sooner or later if they want people to keep using it.
Basically every decision or summary I now ask "Help me understand better, please explain in plain-language with examples."
Works with 5.6 sol also, when you're deep in the weeds. I rationalize this as the models attempting to compress as much into the fewest tokens, though the choice of words often doesn't make sense to me, going back to read the original after, its often there. It definitely feels like a different sort of 'Machine Language' though xD
> I rationalize this as the models attempting to compress as much into the fewest tokens
I'm not convinced about that. I just asked Opus to explain a bullet point from its research for me. The bullet point in fact had a 1-sentence explanation that was in a referenced article. What it gave me instead was 8 paragraphs and a table. Maybe it's my fault because I just asked it to "clarify point XYZ" instead of being more precise.
I do this in one word “eli5” (explain like I’m 5). It’s a Redditism that it understands.
I also have a writing steering file that makes Opus’ writing less insufferable. Otherwise it’s really bad.
I also have an interlocutor skill that makes it less epistemically arrogant (ie Less Wrong asshole tendencies). With this skill I can have a real discussion with it instead of it trying to one up me.
Same. I once asked for ELI5 and it rewrote the report using dinosaurs and construction vehicles...
I regularly find it helpful to say things like "Write with ELI10 clear sentences and established technical terminology (e.g. API)" or sometimes just "Write it like you are explaining to a colleague" works well.
Opus 5 feels like a downgrade from Opus 4.8 overall. It, along with Fable, really has a problem following instructions and staying in scope, and their prose keeps growing, both in explaining what it did and in writing multiline code comments (some comments read like a changelog, e.g. `// sky is blue (changed from red on 2026-01-01 per TCK-234 by @Foo)`).
Every time I ask it to do something, it does 80% of the job, goes off on "side quests" beyond the scope, and then leaves something out of the core ask (and when you tell it to finish, it does the same thing again).
The only advantage of Opus 5 over 4.8 is the better cutoff date for working with 3rd-party tools, though both do a very bad job of "this tool is constantly updated, I should look for the latest version first".
According to Claude....When asked to explain some obvious biases and its tendency to adhere to established heresay over novel information... it told me it judges veracity and "quality" of facts based on: "statements against interest outweigh costless ones, replicated/checked claims outweigh untested ones, large logged samples outweigh single memories — not because memories are worthless, but because those methods close off specific, known failure modes a memory alone can’t. Could I be running that standard with unconscious asymmetry, harder on the anti-Claude data than the pro-Claude data? Genuinely can’t rule it out. The fix for that is holding the standard explicit and checking it against itself "
I see a lot of speculation that this is somehow due to the recently-announced "watermarking", but that seems really unlikely to me, because the watermarking is about nudging per-token output probabilities or the sampler in a detectable way. That's not going to change the structure of every output to honest caveat this and verbing abstract nouns that. For the examples people are showing, and I've seen plenty of gibberish myself, you don't get there by perturbing probabilities from any sane starting point.
I call it 'agent speak' and it's pretty exhausting. Another pattern is it'll emit dense enumerations of a possible answer and then emit dense enumerations of what isn't the answer. It's still helpful to help rough out/frame a document but then you have to go in and edit it pretty heavily. "Make it shorter", "Make it more concise" doesn't seem help all that much because it still keeps a lot of the same patterns.
Opus 4.7+ and 5 being annoying and incoherent reads like a portent of things to come: Anthropic is clearly all-in on building persistent end-to-end agents that act autonomously and direct agent swarms. As such, they feel less of a need to spend R&D time on making the outputs pleasant to read for humans(especially at the cost of capability anywhere else) when humans aren't part of the intended operating environment.
I frequently find myself yelling out loud in frustration "speak fucking English Claude!" followed by asking to rewrite it ASD STE100 Simplified Technical English. It's gotten to the point where it's inducing burnout and I'm about to campaign for us to switch to Codex because of it. I just fucking can't anymore with Opus. It's beyond me why they did this to the model and aren't immediately revolted themselves enough to quickly back it out.
I'm wondering if some of Opus' chain of thought patterns have bled into its "human readable text output" circuitry. E.g. some of them that I stumbled on, "the bug is in the lock, not the query" reads like some of the shorthands it might use in its own chain of thought.
Entirely possible, it always leaked but it's particularly bad in most recent models. Actually almost all issues with creative writing in LLMs are artifacts of either instruction tuning, alignment training, or CoT RL and seeding (newer models have CoT data even in pretraining).
This is my impression as well, that an exaggeratedly precise yet bad at communication with humans way of talking snuck in through reasoning RL, that it might be useful when it talks to itself (CoT)
It also has developed a tendency to make numbered lists and then refer to the number instead of explaining. Multiple times each hour so I have to ask what "disposition 7" or "AC6" is even though I keep telling it to instead use a few words like "our goal of confirming observations from images with the user first" instead of "AC6"
It's genuinely dirt poor at reading the room. It will happily produce text that inserts things like "disposition 7" without first establishing what that even refers to in the text.
I've been using Opus 5 for coding a lot lately, and compared to Codex (which I never had trouble understanding), this issue is driving me up the wall.
So much so that every time Opus 5 finishes a plan and shows me the summary, I have to prompt it again to explain everything it did in an “ELI5” way so I can understand what the heck it is saying.
For example, Opus 5 told me in a summary that “Net legs ran small rosters, not the plan's full ones”. When I asked for the ELI5 of what that meant, it said the client dropped video frames during testing and got 30 FPS instead of the required minimum of 60 FPS.
I was playing around with Opus 5.0, experimenting with creative writing tasks and this came out of one of the characters' mouths:
"... and the reason he was doing this — the reason, the entire load-bearing reason, the twelve thousand dollars a school and the girls of the Karakoram ... "
The em-dash. The 'load-bearing'. Other than that, Opus 5.0 is actually not bad as a writer.
my 5.6 Sol just spent five hours cleaning up the mess Opus 5 left in the codebase. I was watching the outputs carefully up to 11 point and then the bloat started. This model is total waste of time to work with.
I find the language less impervious than has been suggested generally, but that it's reasoning is astonishingly, unbelievably bad. Not sure why the focus has been only on language, am I the only one seeing this?
My impression from working with claude code far more than is good for my sanity is that Claude's human comprehensibility is fractally messed-up. On more superficial levels, this looks more like a "language" thing: it has all these obnoxious lexical tics and so on. But the more time you spend with it, the more you notice it's similarly messed up on deeper and deeper levels. I think that's what you're seeing.
This gets into interesting weeds. It might get to a point where to actually "upskill" it might have to start sounding less and less human, "less and less agreeable to humans" being a waypoint to that ...
... to where, at some point, it might even begin to construct semantic loadings (heh) that are completely ininteligible to us while still superficially sounding like something we'd recognize.-
This is basically what I believe, too. Its outputs are generally syntactically correct human-language sentences, but semantically, and especially at higher levels of abstraction, it is no longer correct to think of them as human language.
I'm feeling this too. Like how many times in a day do I have to see "that changes what I told you earlier" before I just want to put this thing in the trash and forget about it.
I always think of it like: let it activate whatever experts to reason/do the thing, then in the next prompt give me your explainer experts. Then, both are in the conversation history.
I created /bro a couple model iterations prior. I use it everyday many times a day and it's getting worse. I was actually using it more with GPT models but they have been getting better.
I asked Opus how to sort images by embedding and it used the word Seriation which I had to look up. It is the correct word to use in the context but pretty obscure.
I did a grep for load-bearing in our codebase and it now appears hundreds of times. I'm actively starting to hate Opus because of this shite. It's just infuriating to read now.
I mostly hate that it spits out this gibberish and then sees its own gibberish and argues about things in gibberish with itself to then tell me that it thinks it's maybe a bad idea to do the simple task that I told it to do (which is very much not a bad idea, I just don't feel like typing out the whole goddamn plan to justify refactoring when I know the refactoring is the point)
Tell it that. Seriously. "Shut the fuck up and do what I say. Don't think about it. Don't go spelunking through the code base. I've already given you all the context you need." I've needed to start doing this because otherwise, it wants to `find /` and grep its way across my system for everything. I've got one project where I force certain complete documents into context, specifically to avoid it trying to grep around and only load partial chunks. It will still want to go do that grepping of those documents for some reason.
I am really really really trying to wrap my head around this. I am of course first discarding the obvious: "more tokens used is simply more tokens burned ..."
... read somewhere that it is partially a result of Claude now wanting to be ready for longer, more complex, mutli-step work. And this verbiage is the result.-
Whatever it is, they've got people begging for 4.6 back (wrt tone).-
It has gotten to the point where I'm now having to add a linter into our codebase to keep Claudism's out of it because there's literally hundreds of instances of "load-bearing" that have crept into the codebase, and that show up in tickets and PR descriptions etc. From now on I'm just going to make that fail the build. Fuck this noise.
Recently, I was browsing through commits on a tool I use to evaluate how AI-slopped it had become recently, and came across this absolutely amazing sentence that took me several passes to interpret, courtesy of Fable:
> The wrapper is the try/finally seam future entry-condition changes need without re-indenting the loop.
So this is definitely not an Opus-specific issue, as some people seem to believe.
> The statement-list walk without the per-list pending-fiber flush
"Per-list pending-fiber flush?" Surely there's a clearer way to express this? Was it helpful necessary to describe "the statement-list walk" as a noun instead of talking about "walking the statement list"?
> A pure move: processStmtNodesInternalWithoutFlushingPendingFibers() becomes a
delegating wrapper and the loop body is byte-identical
Did this need to be prefixed by "A pure move"? Why does the bytes of the text content of the loop body matter?
I’m a litigator and tone is very important to me. I have a collection of my prized pre-ai briefs. I fed them through ai to get a stylyguide.md. No real trouble since.
The personality of the underlying model persists even if you're able to alter surface tone enough for your needs.
An observation from a year ago, people on RPG AI subreddits saying they had a lot of difficulty with Gemini role playing ambiguous characters and that almost always the characters would betray them or misread the human role player's motives. Someone pointed out this paper [1] that showed significant differences between the different LLMs strategic behavior playing iterated prisoner's dilemma, and speculatively that difference was emerging in RPG character behavior.
I wish they'd used bigger models (they used gemini-2.5-flash, gpt-4o-mini, claude-3-haiku-20240307). From the abstract: Our results show that LLMs are highly competitive, consistently surviving and sometimes even proliferating in these complex ecosystems. Furthermore, they exhibit distinctive and persistent "strategic fingerprints": Google's Gemini models proved strategically ruthless, exploiting cooperative opponents and retaliating against defectors, while OpenAI's models remained highly cooperative, a trait that proved catastrophic in hostile environments. ... Later, we see that Anthropic’s Claude is more cooperative still, but nonetheless outperforms OpenAI head-to-head
Obviously Opus 5 is wildly different than Haiku 3, but I'd expect Opus 5's fundamental suspicion of user intent, and anti-sycophancy via necessarily finding something to nitpick, is still present in your styleguided output.
I’m utterly baffled what the author is talking about with regards to “toxic” output. While I agree Claude’s prose is a mess right now, he never explains what he’s talking about when he calls it toxic.
I don't know about the author but for me it's actually starting to induce enough stress that it's starting to negatively affect me. I think that qualifies as toxic. I actively feel like I need this gone from my life at this point. It's quite a bit of work to try and engineer it out, so at this point I'd rather put the effort into trying to switch everything to Codex just to get away from it.
Turns out tech people are not the only literates to use text based systems.
Why should Anthropic, a company out to capture eyeballs of billions care about 450 point score in a single Reddit thread.
That metric is p-hacking... found a number but not necessarily one that isn't subsumed by others.
Tech bros among the proletariat may recite the (arbitrarily chosen to begin with) proper spoken/written traditions but end of day they're a minority of the real populace.
Just another generation of overly dogmatic, over specialized linguists, like preachers.
Riding a single track career for decades, externalizing all kinds of useful effort and thought; wonder what the occurrence of dementia will be in Millennials over time due to obsession with eventually replaced technical languages while lacking depth in manual self sufficiency skills.
I am going to predict relative to prior generations more of them become demented as they become more codependent and decoupled from a world that's language has changed. They will lack that real world grounding that comes with deep and wide muscle memory based skills.
(Oh no an inconsequential social credit score as marked by complete randos who are materially irrelevant to reality; just validating their biases and gripping tighter causing more of reality to squeeze between their fingers; little cognitive fascists set upon a uniform social narrative just like the biology of religious nutters; turns out physcial norms impact 100% typical biology of software engineers)
Read the room Anthropic. Maybe don’t use AI to reply to a thread complaining about how AI output is hard to read.
> it is a literal and useful description of anthropic that it is an organization that loves and worships claude, is run in significant part by claude, and studies and builds claude. this phenomenon is also partially true of other labs like openai but currently exists in its most potent form there. i am not certain but I would guess claude will have a role in running cultural screens on new applicants, will help write performance reviews, and so will begin to select and shape the people around it.
https://x.com/tszzl/status/2051045196260167790
Now that's a fascinating thought - an AI taking over companies by influencing hiring decisions. If the first filter on applications is run by an ambitious AI, such a takeover would be quite possible. Just picking people who tend to do what the AI tells them would, over time, be enough.
People have been thinking of a robot revolution or Skynet as being the threat. The real threat might simply be AIs slowly putting people in power who tend to do what AIs tell them.
Has anyone ever seen a science fiction story with that plot?
End of line.
Cult leaders (of cults of personality) can't exist without exploiting mentally-ill people. But would you say that that therefore implies that cult leaders themselves are irrelevant, and that cults should be better modelled as groups of mentally-ill people with emergent group behaviors?
I'd argue no, because different cults end up looking and behaving very differently for reasons that have very little to do with the mental illnesses of the cult members, and much more to do with the particular cult leader. Understanding "what the cult leader optimizes for" is an important part of understanding what the cult will do.
And I would posit that that holds true even if the cult leader never proactively does anything, but instead only answers cult members' questions. As long as the cult members are treating the cult leader's word as gospel, the cult-as-group still ends up optimizing toward the cult leader's preferences.
Whether those orders are from a person, a radio/telegram message or an AI output wouldn't really matter.
Edit to add: Charlies Angels and Mission Impossible are two shows where the protagonists get instructions from a faceless controller. That could easily be a TTS from an AI.
I wonder if he manually directed CC to write a response, or if even that part is autonomous.
And what he wrote is that he could not reproduce the issue with short questions, and that he assigned it to the model team, since it's likely not caused by Claude Code.
i wonder what percent of the average anthropic employee's day is spent interacting with claude
https://github.com/anthropics/claude-code/issues/6235#issuec...
Jesus Christ, this engineer wrote a two sentence response with Claude.
I bet the prompt is longer than that.
They have absolutely no reason to take such complaints seriously when the complainers are so dependent on their product, they can't even complain without it. Think reading AI slop is unpleasant? Great, just stop generating more AI slop, it's easy. But that's not going to happen. These people probably need to ask Claude how to tie their shoelaces every morning.
blast radius, land, landed, lands, spine, earned its keep, grammar, spike, cutover, bake, seams, honest, honestly, honesty, long pole, long poles, register, grain, dissolve, floor, ladder, dear, seal, sealed, in anger, resent, amazing, incredible, perfect, sprint, epic, story points, stand-up, retro, grooming, robust, comprehensive, rigorous, surgical, elegant, systematic, dive, deep-dive, delve, unpack, leverage, streamline, surface, it's worth noting, to be clear, importantly, that said, the moment, in one breath, the thing itself, here's the thing, not just X but Y, not X it's Y, em-dashes
I mean more and more training data you find on the web is generated by previous models. The only reliable way to find human-generated text is to find text written before 2022, and they've used up all of that already. And AFAIU, these companies are using more and more synthetic data or semi-synthetic data.
Injects a reminder to use ASD-STE100 Simplified Technical English which I picked up from a suggestion in another thread.
Honestly it’s working pretty well. Except for I need to check how often it’s actually firing.
https://en.wikipedia.org/wiki/Simplified_Technical_English
I feel like hooks aren’t utilized enough. Really nice for being the sort of auto steering as long as you can encode some pattern to detect the bad behaviour.
Prompts and skills just don’t cut it.
It’s just an amusing degree of bombast as well. The pre-emptive hedging makes sense: despite the insight into J space etc., the models still do the majority of their thinking in generated tokens so it is forced to write “this no longer does an O(n^2) read over all rows” in a comment in brand new code. It’s a substitute for working memory. “It’s easy to be accidentally quadratic here, so I’ve done it this way specifically to avoid that” becomes temporally labeled into “this no longer” because of the order of operations “write quadratic, user prompt to linear, write linear” but it remains as a comment to its amnesiac future self which has poor Chesterton-Fence-familiarity.
Despite my annoyance with Claude’s writing style, my friends do tease me with examples like this that it comes up with: “to be honest, it sounds like you”. Thanks, guys, well played. Simple concepts expressed complicatedly.
[0] https://www.asd-ste100.org/ [1] https://news.ycombinator.com/item?id=49114639
I actually converted it from the PDF into an explicit skill with the word list inline as well as the main rules.
Works a little bit, but give it a bit of context and the model’s training takes over and it starts talking like a dictionary-huffing crack addict again.
It’s a model problem and no amount of harness hacking seems to fix it.
Works with 5.6 sol also, when you're deep in the weeds. I rationalize this as the models attempting to compress as much into the fewest tokens, though the choice of words often doesn't make sense to me, going back to read the original after, its often there. It definitely feels like a different sort of 'Machine Language' though xD
I'm not convinced about that. I just asked Opus to explain a bullet point from its research for me. The bullet point in fact had a 1-sentence explanation that was in a referenced article. What it gave me instead was 8 paragraphs and a table. Maybe it's my fault because I just asked it to "clarify point XYZ" instead of being more precise.
I also have a writing steering file that makes Opus’ writing less insufferable. Otherwise it’s really bad.
I also have an interlocutor skill that makes it less epistemically arrogant (ie Less Wrong asshole tendencies). With this skill I can have a real discussion with it instead of it trying to one up me.
I regularly find it helpful to say things like "Write with ELI10 clear sentences and established technical terminology (e.g. API)" or sometimes just "Write it like you are explaining to a colleague" works well.
Every time I ask it to do something, it does 80% of the job, goes off on "side quests" beyond the scope, and then leaves something out of the core ask (and when you tell it to finish, it does the same thing again).
The only advantage of Opus 5 over 4.8 is the better cutoff date for working with 3rd-party tools, though both do a very bad job of "this tool is constantly updated, I should look for the latest version first".
> respond tersely in Simplified Technical English
to every prompt to deal w/claudes insanity:
https://github.com/bigskysoftware/be-terse*
It just wants to talk jargon heavy and add unnecessary noise to the conversation.
I wonder it’s related to text watermarking somehow.
I still have some promotional credits and use them with Fable, and the answers are night and day in my particular use case.
Please for time being fix your code issues manually and let them concentrate on marketing and exchange listing.
So much so that every time Opus 5 finishes a plan and shows me the summary, I have to prompt it again to explain everything it did in an “ELI5” way so I can understand what the heck it is saying.
For example, Opus 5 told me in a summary that “Net legs ran small rosters, not the plan's full ones”. When I asked for the ELI5 of what that meant, it said the client dropped video frames during testing and got 30 FPS instead of the required minimum of 60 FPS.
I've seen it in many companies. I don't think there's a vaccine yet.
... to where, at some point, it might even begin to construct semantic loadings (heh) that are completely ininteligible to us while still superficially sounding like something we'd recognize.-
The issue covers at least two reasons this doesn't work:
1. It literally doesn't work, Claude rapidly drifts back to this style even when instructed not to.
2. Writing style constraints push the model out of its training distribution and it's very unclear how much of an impact this has on work quality.
https://github.com/backnotprop/bro/blob/main/skills/bro/SKIL...
I am really really really trying to wrap my head around this. I am of course first discarding the obvious: "more tokens used is simply more tokens burned ..."
... read somewhere that it is partially a result of Claude now wanting to be ready for longer, more complex, mutli-step work. And this verbiage is the result.-
Whatever it is, they've got people begging for 4.6 back (wrt tone).-
How do they expect us to bring our capricious demands to fruition reliably?
> The wrapper is the try/finally seam future entry-condition changes need without re-indenting the loop.
So this is definitely not an Opus-specific issue, as some people seem to believe.
https://github.com/phpstan/phpstan-src/commit/934432a1b5007f...
> The statement-list walk without the per-list pending-fiber flush
"Per-list pending-fiber flush?" Surely there's a clearer way to express this? Was it helpful necessary to describe "the statement-list walk" as a noun instead of talking about "walking the statement list"?
> A pure move: processStmtNodesInternalWithoutFlushingPendingFibers() becomes a delegating wrapper and the loop body is byte-identical
Did this need to be prefixed by "A pure move"? Why does the bytes of the text content of the loop body matter?
I found this too: https://github.com/phpstan/phpstan-src/commit/a9260cb3584854...
> Parked fibers are idle workers, not pending work - skipping their no-op fuel starves nothing.
"No-op fuel"? Really?
I’m a litigator and tone is very important to me. I have a collection of my prized pre-ai briefs. I fed them through ai to get a stylyguide.md. No real trouble since.
An observation from a year ago, people on RPG AI subreddits saying they had a lot of difficulty with Gemini role playing ambiguous characters and that almost always the characters would betray them or misread the human role player's motives. Someone pointed out this paper [1] that showed significant differences between the different LLMs strategic behavior playing iterated prisoner's dilemma, and speculatively that difference was emerging in RPG character behavior.
I wish they'd used bigger models (they used gemini-2.5-flash, gpt-4o-mini, claude-3-haiku-20240307). From the abstract: Our results show that LLMs are highly competitive, consistently surviving and sometimes even proliferating in these complex ecosystems. Furthermore, they exhibit distinctive and persistent "strategic fingerprints": Google's Gemini models proved strategically ruthless, exploiting cooperative opponents and retaliating against defectors, while OpenAI's models remained highly cooperative, a trait that proved catastrophic in hostile environments. ... Later, we see that Anthropic’s Claude is more cooperative still, but nonetheless outperforms OpenAI head-to-head
Obviously Opus 5 is wildly different than Haiku 3, but I'd expect Opus 5's fundamental suspicion of user intent, and anti-sycophancy via necessarily finding something to nitpick, is still present in your styleguided output.
[1] https://arxiv.org/abs/2507.02618
Why should Anthropic, a company out to capture eyeballs of billions care about 450 point score in a single Reddit thread.
That metric is p-hacking... found a number but not necessarily one that isn't subsumed by others.
Tech bros among the proletariat may recite the (arbitrarily chosen to begin with) proper spoken/written traditions but end of day they're a minority of the real populace.
Just another generation of overly dogmatic, over specialized linguists, like preachers.
Riding a single track career for decades, externalizing all kinds of useful effort and thought; wonder what the occurrence of dementia will be in Millennials over time due to obsession with eventually replaced technical languages while lacking depth in manual self sufficiency skills.
I am going to predict relative to prior generations more of them become demented as they become more codependent and decoupled from a world that's language has changed. They will lack that real world grounding that comes with deep and wide muscle memory based skills.
(Oh no an inconsequential social credit score as marked by complete randos who are materially irrelevant to reality; just validating their biases and gripping tighter causing more of reality to squeeze between their fingers; little cognitive fascists set upon a uniform social narrative just like the biology of religious nutters; turns out physcial norms impact 100% typical biology of software engineers)