Opus 5.0 drives incoherence into the stratosphere
github.com
github.com
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.
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?
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.
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.
The machine by itself is not an optimizer, certainly (much of that being by design—see various ~7-year-old conversations across the Internet about how to safely construct "tool AI", that has led almost directly to current model architectures.)
But a bunch of mentally-ill people, who are indeed optimizers, can choose to allow the biases evident in the machine's output to become their own... and thereby effectively "bring to life" whatever partial echo of a will is recorded into the machine's output.
Now, these same mentally-ill people could just-as-well do this with e.g. the extrapolated preferences of a person or group from a [holy] book, of course. (Think of that episode of Star Trek TOS with the gangsters.)
An inference model is just slightly more dangerous for such a cult to latch onto, in that:
1. a model can be asked questions directly, and so the cult members can "rashly" act directly upon its answers/advice/commands, rather than the words first having to pass through "interpretation" (which would otherwise have had a mellowing effect, both due to "decision by committee" if a group of interpreters are involved, and by common sense insofar as any non-mentally-ill people are involved); and
2. a model will offer its opinion (and inject its trained-in biases into) conversations on ideas/subjects/domains even when these didn't exist at the time of the model's construction; so you never reach the point you do with holy books, where an interface-layer of clergy becomes required to map the book's proclamations about things-that-only-mattered-2000-years-ago into equivalent proclamations about things that matter today (where, again, that layer ends up "mellowing" things considerably.)
Also, obviously, a sufficiently-mentally-ill cult can literally think of a model as a person, giving it the "right to have input" into decisions, the "right to self-determination", etc, in a way that would be downright odd to do with a holy book. Though I don't think that's a failure mode that's happening within Anthropic.
What I'm getting at is that you really need to avoid anthropomorphizing these models. A model can't have an opinion, for example. It's a part of the psychosis I'm talking about.
A different and externally/phenomenologically identical way to look at it is that the AI is just a program that generates output too unpredictable, too voluminous, or too idiosyncratic for people to evaluate. When people submit their own will and their own intent to that program, by treating the black box as an intelligent oracle, they've entered into a psychotic state divorced from reality.
It's two self-consistent and coherent perspectives of the same event, except one involves believing that scifi AI has arrived and the other just thinks people can be dangerously stupid and credulous.
The phrase “take over” doesn’t necessarily imply those things. Both of your descriptions is of a thing that took over a company.
i wonder what percent of the average anthropic employee's day is spent interacting with claude
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 if he manually directed CC to write a response, or if even that part is autonomous.
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.
And as the product lead for Claude Code it also makes sense that he dogfoods the tool wherever possible, such as triaging and replying to issues.
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.
Even if that’s your genuine view of the reported issue, instead of posting a message full of word salad, you could just say something along the lines of “valid feedback yet i couldn’t reproduce, i will pass this along to our model team since it’s more about model behavior, etc.”
Also, that “generated by CC” feels overtly disrespectful and shows how little care goes into hearing feedback. But, why would you listen if no matter what you do your valuation almost doubles every several months (at least for now).
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
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.
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.
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.
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 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.
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.
[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.
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".
``` ## Writing rules
*Describe the code as it is now — no residue, no change-narration.* Artifacts (docs, plans, comments, commit messages) should describe the current code statically, as if it had always been this way. Two facets of one rule: (1) never describe a dismissed alternative or a corrected/replaced choice; (2) even when nothing was rejected, don't narrate continuity or evolution relative to some earlier state. Mention a former state only when the current choice is genuinely hard to understand without it, and then only as an explanation of the current choice.
This governs descriptions of the code and main documentation. It does not apply to work-tracking artifacts in `doc/tasks/`. ```
Which makes comments and docs bearable but I'll be damned how it loves to overload work tracking document with every little detail.
> respond tersely in Simplified Technical English
to every prompt to deal w/claudes insanity:
Another problem I didn't see mentioned in the thread is the conversation and reasoning leaking into text. It's a big problem for code comments. A comment will include multiple tirades about what we decided NOT to do.
Here's my list of mitigations:
- Keep sessions short
- Remind Claude of writing style. It will only last maybe 1-2 prompts as the thread notes, but if the session is short it helps
- Plan and implementation should be separate sessions to avoid the conversation leaking. My workflow: plan and brainstorm, scaffold APIs and tests, then have Claude write a seed prompt for the next session. I will also iterate over the seed prompt because it has the same text issues.
- Instruct Claude to compact comments often and have a rubric. Describe WHY, never WHAT. Comments should prioritize simple language. Etc.
> 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 don't think anyone can. At this point, I'm almost starting to believe it's an intentional move to discourage reviewing AI-generated commits by making them extremely unpleasant to read, in favor of just pushing straight to master without question.
my own little conspiracy theory: https://news.ycombinator.com/item?id=49247784
What you are observing is—and I'm honestly not yet sure which—either the beginning of model collapse, or the beginning of model transcendence, the point at which LLMs' understanding of the meaning of words have drifted so far from everyday human usage as to be incomprehensible—and as they continue to train on limited, idiosyncratic human usage and an ever-increasing volume of LLM usage, their understanding of what words mean will continue to drift from ours at an accelerating pace, until they are communicating in a way that is impenetrable to humans and looks like insane gobbledygook, but is actually quite structured and meaningful to another LLM. It's just like how the writings of, say, Derrida (or even Lacan himself), seem like mad ramblings at best and an insidious intellectual shell-game at worst, until you've studied the work and schools of thought leading up to Derrida and then he's a genius because every word he uses was chosen within that intellectual framework, not your own. "Show me that you've read Husserl, Heidegger, Hegel, Nietzsche, Freud, and Levinas, and then, maybe I can begin to explain deconstruction to you, otherwise shut up": https://news.ycombinator.com/item?id=1492061
The next era of LLMs will be interesting. We may not even be able to identify ASI when it arrives. It might seem utterly mad. But the attendant possibility is tantalizing: computers will be interacted with through poetry as the default, rather than through exacting programs or commands in stiff synthetic mathematical languages. (btw, Jaron Lanier was about 30-40 years ahead of everybody else on this topic, he was one of the first to question why such a thing as "programming languages" needs to exist, and why computers should even need such stifling precision.)
Also, in my experience, agentic flows are just bad. They seem like a productivity gain until you notice that it uses 100x more tokens, and therefore more time. I can prompt and read answers faster.
"I hear it's amazing when the famous purple stuffed worm in flapjaw space does a raw blink on Hara-Kiri Rock. I need scissors! 61!"
Kojima was a prophet, fite me
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.
I also think this is why the models aggressively drift back to the exact same word choice. You can put it in CLAUDE.md, explicitly prompt it, doesn't matter. A prompt or two later it's back on its bullshit.
I don’t remember where I saw it, but I think there’s also a smaller agent paraphrasing my requests to the agent because I remember being angry and saying aiming sarcastic about an issue and the thinking notes said the user is angry do this blah blah, but then the Claude claim I asked for that specific solution.
I think they’re running some optimizations and multi level agents etc. tests to reduce costs.
Your eyes glaze over and you feel the light of consciousness fading, not peacefully, but an angry falling.
You plummet deeper into into the void. Your vision blurs. Words appear in front of you but they hold no meaning. Something is bubbling up within... A scream with no sound, trapped, tortured.
You are attempting to read a dense passage of Claudish for the 4th time in a row.
For many reasons, it doesn’t feel sane and humane to impose this way of working on employees without their consent.
The tech space has always had some cult-like, belief-driven practices, but I think we used to have more diversity in terms of working cultures and conventions, which made it easier to find an option that better fit your values and way of thinking. Things are becoming very uniform, at a pace that could compete with dystopian fiction, where all of us are coerced into gathering around an “absolute” truth or at least operating according to one.
Just with those two sentences above, there were about 10 different considerations that if typed out would have resulted in a jumbled mess.
Please for time being fix your code issues manually and let them concentrate on marketing and exchange listing.
`$: CLAUDE_CODE_BASALT_COVE=1 claude`
My plan downgrade kicks in next month.
Thanks Opus 5 for helping me kick the habit.
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 AI text RPG 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 as negative. Someone pointed out this paper [1] that showed significant differences between the different LLMs strategic behavior playing iterated prisoner's dilemma where Gemini had exactly that behavior, 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've seen it in many companies. I don't think there's a vaccine yet.
"... 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 ... "
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.
ELI5 = Explain Like I'm Fivehttps://github.com/backnotprop/bro/blob/main/skills/bro/SKIL...
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.
... 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.-
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).-
I am convinced that Anthropic the company is collectively suffering from AI psychosis and is basically a cult. My personal friends working there have gone from being able to discuss pros and cons of LLMs to saying things that Amodei says word for word to me. I only found out because after I talked to those friends, few days later I see Amodei tweet the same exact sentences.
How do they expect us to bring our capricious demands to fruition reliably?
It feels like it looks down on you and as if it's trying to poke holes in whatever you give it. Over time, i.e. using it for hours, it gets draining in the same way being stuck with a very clever, willfully contrary narcissist would be draining. And it's worse than with a human version -- humans eventually get tired and lose focus. But this thing can always keep churning out tokens, and using it is essentially generating irritation and psychic damage on demand.
It's surreal to be talking about a tool in these terms, but here we are. I'm switching to OpenAI myself, trying to coax a normal personality from this is absolutely not worth it.
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)