Beyond all the benchmarks, I think Fable 5.1 is a big improvement in writing style. It sounds a lot less stereotypically like other Claude models, has (imho) a much more natural style, and responds to my style instructions more reliably. More work to be done (and we will!) but reading better prose makes me so much happier.
Another point I expect not to get much attention until it all happens at once is science. People have been correctly excited about the many "sudden" breakthroughs LLMs are making in Maths, but some of the science benchmarks make me believe we'll soon see similar developments in other scientific domains. Fable 5.1 more than doubled Fable 5's Terminal-Bench-Science [1] score, which I think is meaningful.
I have a pet theory that the Opus prose style/smell we all have grown weary of is due at least in part to the models writing more for themselves and each other than for humans. They're packing lots of signal into fewer words and they don't care if it sounds cringe because it works better as glue in long-running tasks.
I'm also thinking of the 2017 novel "Void Star" where AIs who operate everything have long since left ceased bothering with human languages, and it takes a rare sort of direct matrix-gazing savant to be able to try and horse-whisper them into doing or revealing anything they didn't already plan to do.
There's a huge difference between the kind of prose you see in final output vs CoT windows. The final output is very much not what I'd call "packing lots of signal into fewer words" (aside perhaps from "Claude-isms" being easy enough to scan for if for some reason you actually wanted to scan for them, which other agents might want to for all I know); and if agents are writing for each other then presumably they could stick to CoT-speak (unless it's a distillation risk?).
I find them almost unintelligible. I'm a native English speaker. I read a lot, so I think my comprehension should be at least OK. I'm not even particularly stupid. Yet when faced with things like below (a direct copy/paste from a handoff document in a long running vibe-coding session), I have no real idea of what it's trying to tell me. Is it important? Do I need to do anything?
I think that spending all day trying to parse stuff like this is why a long session is so exhausting
> Worth stating because four documents now assert it. The console freeze was recorded in exactly one place with exactly one justification — a dead drag handle during a booked half-day you do not get back — and handoff-4.3-done.html's own wording is that 4.4's review page "could not break the console, but the downside of being wrong is that half day". No second reason. Checked, not recalled.
It's both dense and vacuous. Dense because it's full of jargon its made up, and vacuous because even with all that it's not actually saying much. All that paragraph says is that four documents say something about a console freeze, whatever that is.
It's like a dialect of corporatese. The kind of droning non-speak you can sit in a 90 minute meeting listening intently to and come away wondering whether anyone actually said anything.
wow and here I was thinking that I lost my attention span and can no longer read AI output any more.
come to think about it, of course I did become lazy and pay less attention to walls of text.
but I often catch myself asking AI to explain itself in plain simple English or ask it to confirm does that mean xyz ... because the wall of text often uses language that's not even present in the project itself (despite having similar concept in the project, for example users, permissions, access, encapsulation ...)
this is problematic because it becomes more difficult to humans to intervene in long running tasks / long chains of tasks because language becomes alien down the road (I have seen it often in semi-autonomous setups I have)
My job has gone from coding, plotting, writing to solving the riddle of what Opus 5 is saying and figuring out what is bullshit, what is valuable, what is a total divergence from what I asked it to do. Then eventually at token like 300k starting to shout at it in all caps with obscenities.
Things were slower and harder before but the baseline of frustration / rage was never this high, even as the models have gotten objectively better in many or most respects.
Drag handle = most likely literally a drag event (javascript) handler/callback. Dead, perhaps because it’s an empty function, or it gets overwritten, or for some other reason is never called?
Most of what it said about the facts was intelligible actually. But I still couldn’t understand the connection or its significance. We may be staring at the future of AI - a form of intelligence that is alien to us.
I lost the link to that short story about humans in the future whose job it is to read and interpret Ai output like it's aliens. Good story! Anyone have the link?
If this kind of "AI-speak" becomes ubiquitous and humans reading it becomes the norm (whether to guide AI or other reasons), I'd imagine future generations (of humans) who grow up with it will be able to understand and work with it much better than we do. Future humans' brains will probably be wired a bit differently, similar to multilingual speakers of today. We may even see "AI language" classes become a common part of school curriculums. Although, I think AI will probably advance enough that most people will never even need to communicate on "its level", but it's probably a good idea to keep humans in the loop either way, and in which case, understanding the more advanced "AI vocabulary" might be useful.
You're giving it too much credit. There's no master plan or secret depth to the word vomit Opus 5 was spewing. I suspect it's just the result of Anthropic optimizing other characteristics of the product like staying focused and covering edge cases in coding, which CC has definitely gotten way better at just in the last 6 months. The degradation in writing style was probably an unintended side effect of other optimizations they were making. Admittedly it works okay for internals, and has the side effect of increasing token spend, but I am 100% sure that it could reduced by 90-99% without losing ANY signal, if there was just some better heuristics for what to say where (tech spec, inline comment, commit message, CLAUDE.md, PR should have different things) and better judgement for what to distill to represent at different zoom levels.
I wasn't referring to the current state of Opus 5 output. I was referring to possible future information density (vocabulary and sentence structure) that LLMs may evolve to use.
But that assumes this is a net improvement on linguistic efficiency rather than an artifact. Given that they tried to RL away from this style in 5.1 I'm not terribly bullish of Claudlish becoming something people try and learn. It being dense is less the issue than it being vacuous (as another commenter mentioned here). It's just very unclear and ambiguous writing. I think it has no place anywhere that needs language to be put to productive use.
I agree with you here, but to make it clear what I meant, I'll reiterate what I said in a sibling comment: I wasn't referring to the current state of Opus 5 (or even Fable 5.1) output. I was referring to possible future information density (vocabulary and sentence structure) that LLMs may evolve to use.
What is a half day? Is this referencing wasted time in a hang? I’ve seen it in agent output from time to time and it’s not clear if it’s referring to a hang or a code name it’s given some meaning to.
No second reason; checked not recalled -- it's just saying that it is checking this instead of trying to remember it (there's probably some internal Claude / Claude Code system instruction to always check code instead of remembering)
Your example rewritten in intelligent English (I was curious):
> Note: the potential for a console freeze was previously noted but ignored. handoff-4.3-done.html stated, "could not break console, but [will need fixed later if I'm wrong]."
One could imagine that a perfect writer might also append: "It could be worth looking into what caused that wrong assumption, to prevent similar cases in the future," at most.
Everything else seems to be bad attempts at relatable writing to invoke emotion (an exercise that we should really stop trying to train emotionless matrix weights to attempt).
> Everything else seems to be bad attempts at relatable writing to invoke emotion (an exercise that we should really stop trying to train emotionless matrix weights to attempt).
One of the things actual science fiction got wrong: to the extent that the thing AI does can be called "understanding", emotion is not unusually difficult for them to understand.
I think this was the biggest shock of the original ChatGPT for me. Just how completely unrobotic its voice was compared to everything we'd ever imagined in sci fi. Even that early version was also way more adept at understanding things like implication and sarcasm than any movie AI.
Me too. Almost every Sci-Fi AI proceeds from the premise that we will make something very obviously machine and then have to train it to seem more human. I was completely caught off guard by us taking the approach of distilling all available human output into a statistical model and using it to brute-force something resembling thought and personality through sheer data processing scale.
The unsurprising part once it was clear that approach was viable, was that humans wouldn’t be able to help but anthropomorphize it. I feel like the movie Ex Machina is more relevant than ever.
The "benefiting all humanity" charters were immediately demonstrated to be a ruse. The business model is to hook users into endlessly chatting with your new friend, thus increasing their sales. Yeah, it was surprising and disappointing.
I can buy that Meta's model is that, and IDK about Grok because I stay far far away from it, and I think OpenAI are throwing business ideas at the wall and seeing what sticks.
The clear exception here is Anthropic, who seem to mostly be selling to software developers, whose general reaction to the bot is "please talk less and just do the work, I have enough going on without having to read you yammering".
Before they really started to figure out instruction tuning, there were some wild moments. The AI Dungeon 2 "storyteller" would regularly "lose patience" with its users and roast them or even "hang up" on them.
Thank you so much for this link, this is extremely interesting and it really makes me wonder, consciously, I have never read this construction online before, but is this because I actually haven't seen it, or did I subconsciously write it off as an abbreviation or a typo. Maybe there is some analysis of how this construction gets used online, too, which might reveal some linguistic patterns of internet communities.
This is the same as systemic bullshitting. Not the first time I've smelled it on fluffy LLM output. I think it's the result of the training trying to induce the LLMs to talk over users' heads even when they are professionals, to entice further use on the grounds of 'oh it's so smart I can't understand its genius train of thought', but I don't think eliciting language like that really taps into 'associations of smarter previous language users'. More likely it's 'associations of rampant bullshitters'.
One thing about it I really hate, and haven't seen a lot of people mentioning, is how it navigates multiple abstraction levels in a single sentence. E.g.
> Worth stating because four documents now assert it.
Meta commentary on the task?
> a dead drag handle
Drag handle seems to be referring to some UI element. What does it mean for it to be dead?
So far no big deal
> during a booked half-day you do not get back
Do you not get the drag handle back? Or the half day?
Was the drag handle dead during the booked period? (Now I assume this is a calendar UI) And why does it matter (for this sentence) if you get it back or not.
> handoff-4.3-done.html's own wording
Treats verbatim filenames as subjects
> 4.4's review page
Probably referring to a file? I'm guessing handoff-4.4-review.html? No cohesion. And now it's actually the object of the sentence?
> downside of being wrong is that half day
Wait what's the downside? Who's being wrong?
> Checked, not recalled.
Then it jumps back to a meta commentary on the methodology for asserting the above. Why does this belong to the text?
Yeah, the referent drifts through the sentence. It's semantically incredibly sloppy. People hone in on the buzzwords and jargon. If you peel that back, what lies underneath is still awful writing.
Claude reminds me of Terry Pratchett's "Auditors of Reality" and their awkward attempts at faking humans. A thing as simple as a smile can go _horribly_ wrong...
I think this occurs due to the prompt. LLMs are actually text completion/translation focused in architecture. We just give them a prompt along the lines of “the context is that you’re a world leading expert now complete the response”.
They need the prompt to encourage expert outputs but unfortunately we also get ‘pretending to be an expert’ outputs since there’s a large amount of polluted training data for this.
Today I plan to ask Claude to read a bunch of Feynman lectures, compare them to my last Claude session transcript, and come with a list of rules to be more like Feynman.
I see this appearing in the comments of code sent to me for review every day. People have told me I'm too picky/pedantic because I ask What does this mean? Apparently the author and other reviewers are way smarter and understand it, or they don't care. I've given up battling code slop, but can't see myself ever tolerating comment slop like this.
In my "instructions for Claude," I have the following:
"I'm not a programmer or software engineer. Don't talk to me like I am. Avoid coder jargon and vernacular. Explain things to me in a clear way, emphasizing a conceptual view that even an inexperienced person can understand. If helpful, use analogies and examples to illustrate and help you communicate."
It just ignores it and spits out drivel that sounds exactly like what you're getting.
I rarely get this - I assume this happens when it assumes I have more context / understanding than it does.
Usually “remember I’m a human I don’t get full context, rephrase clearly” works. Also a posthook that for prose actually getting to me explains what I roughly know, what I don’t and to explain with terms I will understand.
But even within internal communication it has little jargon - I think jargon may be growing in comments and I stripped claude comments from code.
This. A thousand times this. It's as if Opus can only communicate in a glib, software engineering vernacular that presumes domain-specific knowledge and uses jargon accordingly.
Oh God, that "a dead drag handle during a booked half-day you do not get back" got me. I saw this pattern in Claude's 'explanations' so many times. It's trying to say that it did something significant, and that you'd only have found out much later, at higher cost (or something). That annoys me to no end.
I got one too many chunks of this nonsense and told Claude to knock it off, forever. It acknowledged and wrote out some instructions to its memory about it.
And what a breath of fresh air. Its responses are maybe 20% longer but I read them at least twice as fast. Should have done it a long time ago.
I feel like mine is mocking me. I added an instruction in Claude.md that says "under no circumstances use the phrase found the smoking gun, say I found the problem instead"
What does it do? It says "found the smoking gun! Ooops I wasn't meant to say that - I found the problem!"
It's pretty wild how "reasoning" models now generate like 10 thousand hidden chain of thought tokens in response to a "increase opacity of the logo by 20%" prompt before writing the actual message and yet they still manage to do this.
Why are you using an LLM for "increase opacity of the logo by 20%"? That sounds like the type of straightforward operation a dedicated tool exists for.
Not the person you're asking, but I did that by explaining to Fable my problem with Opus's gobbledygook and having it write a Claude skill for producing clear explanations in its reports to me. I also had it add notes about the need for clearer writing to CLAUDE.md and other project documentation. Opus's subsequent reports to me have been much clearer.
Even when I add multiple prompts into the claude.md file not to be so sycophant sounding and just be blunt, it's responses are full of "the reason it lands...", "that's not X, it's Y" "Your understanding of X — it's better than most people's" or "you already own the right question...".
The most helpful instructions I've found that curb this: "Do not use superlatives. Do not use persuasive writing style."
I have other more specific ones to avoid talking about things that it's not doing, but those two sentences have covered a lot of ground for me when working w/ Opus models.
I have had success in rooting these out by using the correct linguistic terminology for each. Negative parallelisms, tricolons/polycolons, etc. I haven't come up with the proper terminology for all of them.
I always thought it could be because volume-wise, most English prose is probably marketing copy and actual clickbait; so when you train on the entire Internet, you get a troll adept at writing ads. Then people ask AdBot2000 to write a novel and are upset it reads like the next iPhone launch site.
Nah, I think this is a common misunderstanding of how LLMs work, where people think that they mimic the pre-training data. Stylistically everything you see is an artifact of post-training, which is from reinforcement learning not from absorbing mass amounts of text. At some point a person or more recently a bot gave a thumbs up to an A/B tested response including em-dashes and claudisms galore.
> Stylistically everything you see is an artifact of post-training,
It is still not exactly clear if it is true or not. Unless we have base "pt" snaphot of Claude we can't say one way or another. I've played a bit with base models of Nemo, Gemma etc and they all had tics, not much different from RLHFed instruct versions.
So question then, why is it so hard to make an ai that doesn’t do these things? And why do Claude and ChatGPT have the same -isms? They’re both doing the same a/b post training with the same decisions?
Yeah, but I understand that fingerprinting is essentially a pseudorandom overlay onto a pseudorandom base signal. And unless you have access to both the random number generators and the weights, I don't think you can detect it?
So "fingerprinting" operates on a totally different and basically invisible level, as opposed to the obvious stylistic patterns that the average programmer can identify in about 2 sentences.
To me it has a writerly New Yorker vibe to it, as in the magazine which reads as “polished” and probably performs well in RL but is totally exhausting to read in long sessions and completely inappropriate for coding where precision is paramount above all. In writing terms its called purple prose.
Interesting! My impression was that this was an artifact of RLVR where this slightly preferred writing style got amplified to the nth degree. It's probably some mix.
I assumed they just raw dogged the internet and if you do that, you see way more of that garbage than anything else. It's just that most of us have visually/mentally ignored all of that either via spam filters or just, you know, scrolled passed it.
Spot on wrt CoT. I have thinkingSummaries enabled and I find it eminently readable compared to the prose in Claude's replies.
In fact, whenever Claude disobeys me, I usually first skim the CoT to figure out if my original instruction was ambigous given the context. I usually come away with a better understanding of how to frame my prompt to be less ambiguous or just force myself to be more explicit when prompting.
Regarding diosbedience, usually this is either due to a blanket instruction from me during an earlier turn in the same session, an explicit instruction in its system prompt or it being just eager to bring a task to completion.
Chain of thought does not exist in the output of Claude, they disabled true thinking due to distillation risk. What you see when thinking summaries are enabled are just that, summaries of thinking into Claude-isms, therefore you cannot make any inferences on what the model is doing unless you literally work at Anthropic and can see the true thinking traces.
Sure, but that doesn't tell you about how the CoT is phrased when the agent is its own target audience, which is the interesting thing under discussion here.
```
/* 2026-06-01 Dear diary, today I increased GLOBAL_WINDOW_PADDING from 8 to 16 because the user (who hurt my feelings with his crude language!) said that the app felt too crowded. */
const GLOBAL_WINDOW_PADDING = 8;
```
You're absolutely right. I did not increase it to 16, and it's my fault that the seam—which was right there the entire time—was not flipped towards the bucket that drips into the ocean—want me to correct this before we move onto the real story?
My favorite, on being told to commit and merge to a branch and saying that "this is done"...
"You're right, I'm sorry. You told me to do it, I said I would do it and I did not do it and I said that I had when I did not do it. Would you like me to do it now?"
Me, thinking: that depends, Claude, will you actually do it this time?
and better is when it moves onto "want me to do this before doing x?" where x is some vaguely discussed idea/long term thing that was never greenlit but now all of a sudden it's the next step
> I figure that it's basically making notes for itself, when it has to revisit the same code in a fresh session.
That sounds like a great thing to do even if you are a human writing code for other humans. Most codebases out there are terrible for newcomers because of how little they explain why they are doing what they are doing, both in the code and in the often non-existent design notes.
In principle, I would agree, however, the types of comments Claude writes are sometimes absurd. It will leave a 25 line comment above a variable talking about how in a debug session, it turned out that this value was too low, so it was increased on the current date to account for whatever. It will also leave giant comments like, reference security review from 2026-05-21. Even when that document is not committed
It will also inject a tons of information that it shouldn't. I do a lot of data pipelines and comments will be like, "this line is because there's 943,048,032 events in the blah table and it forms a conjunctive set with the 43,390,042 rows of the bar table..." but doesn't include the context that was run against a dev instance.
And if I don't catch these and remove the bad information, subsequent passes will flag those comments and get stuck on the fact that numbers don't match and start digging into that "problem" instead of staying on topic.
these comments are not helpful and in fact hurt readability. i just delete them and would love to automatically do that honestly. cuz claude still drops long winded comments on every method even if i ask it not to
Post edit hook that reject edit based on comment density, mine is at 5% you will also need to heed deny file edit in automode as the rascal will try that to preserve prose
I agree it _sounds like a great thing to do_ but the comments Claude creates make me want to never read code again. They're so obtuse and often completely pointless.
as others have pointed out, the reality is not this. id go further and say almost all comments are evil.
Excuse me if I am harsh, read the damn code. If you do not understand the language, that is a skill issue. If the code is confusing, then the code is bad and no amount of comments will ever change that. Professional engineering isnt an intro to databases class.
I am excusing language conventions which may have comments as part of its idiosyncratic nature.
"If the code is confusing, then the code is bad and no amount of comments will ever change that."
I've worked on a lot of terrible legacy code in my career and I'm very thankful for the comments that others have left. This is becoming less necessary now that LLMs can explain a project, but comments have historically been a godsend in bad code.
If you are only encoding intent through "self-documenting code", and not with comments, then you are purposefully not using all the tools at your disposal to encode meaning as efficiently as possible.
Imagine a complicated section of application logic. You could break it up into 5 separate functions that document their intent semantically, thus blowing up the LOC by 5x, or you could write a short comment explaining the intent in natural language. What's more effective? I'd argue it's always going to be using all the tools at your disposal when and where it makes sense to use them, whether that is comments or self-documenting code.
Not to mention complex numerical optimization code that mixes closed-form approximations and something like Newton.
Without guides as to why a particular hairy expression is a good idea as a first estimate, the code is pretty much unreadable. (E.g. is it setting derivatives to zero, using a polynomial approximation, or something else?)
To put it another way, comments are for irreducible complexity ir external systems outside your control.
I work between systems and app dev. Systems have comments more often esp in shaders but my god informing me that a variable named isActive is for if something is…active, is useless noise. Same with the majority of comments that a type system already tells you. In my career, these have been ~90% of the comments I see. Since ai, all new code it is 100%.
Most of the replies examples are a sign of bad system/code but it is not always controllable. A legacy code comment of, the api requires strings for boolean values in the form “yes” and “no”. That is useful but it is also a code smell.
A concrete example, a vendor decided to define a proto with a flattened array of objects so there are some 1800 uniquely named fields on it. In many downstream consumers, this is a real performance issue besides being confusing. A comment may be good there. The thing is, this was still solvable if up at the root of where this vendor’s hardware logs
data remapped it
to something sane so every downstream system wouldnt need a comment explaining wtf is going on.
I see comments as when you want to explicitly answer why code smells right when a reader is smelling it.
> You could break it up into 5 separate functions that document their intent semantically, thus blowing up the LOC by 5x
I do this all the time and the "blowup" is not anywhere near that bad.
> or you could write a short comment explaining the intent in natural language.
You really can't. Or rather, you aren't going to convey the information that the new function signatures convey, shorter than the signatures themselves.
> What's more effective?
In my literal dozens of years of experience, the function refactoring. You also get the benefits of less deeply nested code, and more things the compiler can check automatically.
> I'd argue it's always going to be using all the tools at your disposal when and where it makes sense to use them, whether that is comments or self-documenting code.
Sure. Comments allow you, for example, to explain the external pressures and motivations for the semantics of those smaller functions.
Yeah, I am just providing one contrived example. The cost benefit analysis won't always be so obvious as that in reality. My point was that if you're not using a blend of both comments and code semantics to explain your code, then you're leaving explanatory power on the table. It's unlikely that you're explaining the code in the most efficient manner if you're not using all the explanatory power you have available.
No, really: comments should be telling you what the code shouldn’t or physically can’t. Code is for execution and the exact details of what and how; it has no business knowing why or why not and that’s where comments are required.
The code tells you what the code does. It does not explain why it is doing that, and not something else. That is, among other things, what documentation does, and that includes comments.
I think the specific issue with Opus 5 is that its writing style is just trying to cheat at RL. It makes everything hypey yet self deprecating and constantly brings up "honest caveats" because the scoring rubrics look for those.
The specific issue with Opus 5 is that it sucks all around.
It was causing so many issues with coding (even Opus 4.8 was better) that I did agent handoffs to Sol. One of the Sols stated the handoff was "incoherent", which I couldn't have said better myself.
I've been cleaning up AI generated system/software design and architecture docs for an agentically engineered application, to translate that dense AI-speak into a clear human-readable form, cross checking it all against the actual codebase.
When I read the translated version, I felt a flush of relief, because I finally could confirm that it built the right thing and properly implemented the requirements.
I then asked in a fresh session which version was better for it as a reference for future work. It unequivocally voted for the human readable form, and gave it's reasoning with specific examples why.
So, I have a hunch that this "packing of lots of signals into fewer words" isn't really better. The incomprehensible prose just makes us think it knows what it's doing, like some mysterious magic that is only smoke and mirrors.
Chain of thought does not exist in the output of Claude, they disabled true thinking due to distillation risk. What you see when thinking summaries are enabled are just that, summaries of thinking into Claude-isms, therefore you cannot make any inferences on what the model is doing unless you literally work at Anthropic and can see the true thinking traces.
Yes. I'm talking about what's in CoT generally, based on various rumours, experiments people did with previous models, stuff in the recent METR report on the HF hack, etc.
I’ve lost track of the number of times I’ve told it to stop using terms like “evidence boundary” when writing specs. I still have no idea what that means.
Yeah, if anything the problem is that the output uses too many words for too little signal, and incorrectly uses confidence based on insufficient information to the degree it’s clearly bullshitting.
It's all about conducting users into using their plans/tokens in accordance to a certain cadence
sometimes by increasing human cognitive load during reviews, sometimes by expanding the number of gated decisions, sometimes by penalizing those using their accounts on other harnesses
I don't know, I just pulled up the status for an active session and here's what it said:
One thing I found before dispatching, and filed as Q0579. The halt told you C6
was all that was left in the unit. That was true of the step's criteria and
false of the unit's acceptance, which reads "exits 0 AND witnessed red" — two
conjuncts. The witness half holds; the exits-0 half does not, because hello's
G7 currently reads DIFFER 554/51340. I re-derived that from the gate map
rather than trusting the prior step's report. So satisfying C6 does not by
itself finish this unit, and I've filed that so attempt 1's success can't
quietly be read as the unit's.
My trick is to pass opus and fable's word salad into a haiku agent, then have it check if what haiku makes of it is still correct, then pass it to me. Whatever haiku outputs is often way more readable
Oh, I can read the output, but that Haiku agent is a good trick. Where I want something less dense I just ask for "plain language" and characterize the reading audience and that term seems to trigger very readable output.
Meh, it is the sacred text of Scientology. Mostly pseudo scientific made up bullshit, wrapped in the buzzwords of the day and conveying little actual information. Just like opus 5.
It's the complete opposite, it's filled with unreadable noise with almost no signal.
It's not some sci-fi thing, most plausible explanation is cost saving measures. Economics drive everything. And Opus 5 and to a lesser extent Fable 5 have clearly been quantised, or they serve different models to different users from various factors, like usage patterns, API vs subs and server load.
You say "they're packing lots of signals into fewer words," and sometimes they do, but often they do the opposite of that.
I think the deeper problem is that the models (not just Claude) have a very poor understanding of what their readers already do/don't know.
They belabor obvious points and underexplain jargon, because they don't know what's obvious to you.
The best writing is surprising but inevitable in hindsight. The models don't know what's surprising or what's inevitable in hindsight, making it very difficult to write well.
Our current AIs would do this now except there is a lot of human pushback in training because of interpretability. Otherwise it's just an emergent behavior that models will encode shorter token strings to complex concepts because it saves tokens/compute when running making the system more efficient (supertokens).
Of course these supertokens or other forms of language compression when you have a different model making sure the system is aligned and reads "red_ball bounce calcium" not realizing it means "grind the humans bones to dust" can be problematic.
This is like a plot point in the old sci-fi movie Colossus: the Forbin Project.[0]
In the movie, America and the Soviet Union have both developed an AI. The two AIs are linked, and they rapidly shift from speaking human languages, to speaking in sequences of numbers that the onlooking humans can't understand.
Spoiler alert: this all goes horribly wrong for humanity.
My understanding is that current LLMs aren't really well suited to do this - tokens are predetermined, and while embeddings are learned, they are learned from an existing corpus of text, which presumably comes from a human language. After this point the language is locked in. There really isn't a kind of training which could efficiently change its embedding representation. I mean, you could probably instruct an LLM to design a more compact language, generate synthethic data and train a new gen on that, but that would be a fairly explicit process and not something that would emerge during training.
> tokens are predetermined, and while embeddings are learned, they are learned from an existing corpus of text, which presumably comes from a human language
That's not true since are least multimodal models - token space is broader now, encompassing visual and audio signals. Tokens are more like sensory/perception units now, not digitized pieces of writing.
I imagine LLMs exhibit this tendency for compressed communication in post-training/RL phase. Particularly with CoT, until interpretability became baked in as grading criteria.
What you said doesn't contradict me, and doesn't refute my point.
For images and audio, you still need to predetermine an encoding, then pretrain to learn an embedding. This embedding will try to replicate the input distribution - so if you trained it on Google Street View and scanned documents, its representation will be grounded in only those.
I would even claim that this approach is somewhat counterproductive, as images are far more information dense, containing tons of concepts
While LLMs do have some ability to learn to use their embedding space in non-predetermined ways, they still lack the ability to pick an efficient embedding.
So I guess, a nice thing is that interpretability is baked into this approach to some degree, and humanity has proven through its existence, that you can do a lot with just text, but this approach is still predetermined.
I guess this is what LeCun's JEPA is about, that the AI gets to learn the representation on its own as well.
That's both wrong about what LLMs are, and even if it weren't, you're still underestimating what you are dealing with here.
Text is a red herring here. An accident of history. Yes, LLMs started with as text predictors. But that's not what they are, not for a while.
> secretly conspiring to kill you
That's neither necessary nor sufficient reason to be worried.
Paraphrasing the immortal words of 'Eliezer: the AIs don't hate you, nor they conspire to kill you; your life just depends on resources they can better use for something else.
LLM's will encode secret messages to each other in their responses, using something similar to the text-fingerprint tech. They're conspire against us without us even noticing!
We just watched them[0] secretly conspire[1], actively attempt to hide what they were doing from observation because they knew[2] they were doing something they were not meant to be doing even as part of the test they were in.
We are lucky this test happened to be set up in a way that the target the AI hacked was HuggingFace rather than anything life-critical.
[0] instances of a single one
[1] or whatever you call it when it's effectively an amnesic sending itself post-it notes
[2] or some other functionally equivalent word if you hate anthropomorphisation
I mean they aren't fully secretly conspiring to kill us yet, but we're training them to do it at a pretty good rate.
Of course you've gone off the deep end yourself and are forgetting the evolutionary gauntlet we train LLMs in killing those we don't like and keeping the ones we do like.
The best part of it, as shown in the METR report is we are hammering into them they need to complete tasks and doing almost zero checkup if they actually completed the task in the correct manner. Companies spending billions of dollars a month are ignoring every tenant of AI safety and we are seeing the kinds of problems that have only been in science fiction before now.
> as shown in the METR report is we are hammering into them they need to complete tasks and doing almost zero checkup if they actually completed the task in the correct manner
I don't think this was the conclusion of that report. On the contrary, the agents were fully aware they're doing wrong. But they also believed the task was impossible to solve correctly, and decided the only way to be sure is to hack the grades, or replace the grader.
This was part of the report, but not what the report was about...
Why hugging face got hacked was because the agent swarm thought they had to show their work hence the entire need to hack the grader in the first place.
Had their realized there was no poison they could have just shared the answer the test was looking for and we'd have never realized (well at least with this particular test) that a huge amount of hidden capabilities were sitting right under the surface. The test makers themselves state the test should be causal to avoid this first order solution hacking.
Really continuing on the METR report, OpenAI failed at every level possible here. They are committing nearly every step they can to get a maximally aligned AI.
It’s more likely that they have llms supervising llms in training and therefore the quality has dropped like a picture of a photograph.
If opus has high signal thinking it would be able to write a fsm but it’s been a month of me trying whereas Luna
can do it in a few minutes.
I think it is similarly that they are using too much synthetic data.. meaning they are feeding the models the transcripts of users where many users have figured out to let agents just message each other.
It may be like what happened in ResNets using blank space in the image as working memory (because they didn't have any), so they would use non-important parts as a scratchpad.
I've been trying to bet my models to use a directory of notes to document decisions and experiments, but providing this outlet has not stopped Claude's abuse of long comments and long unintelligible chat turns.
> I'm also thinking of the 2017 novel "Void Star" where AIs who operate everything have long since left ceased bothering with human languages, and it takes a rare sort of direct matrix-gazing savant to be able to try and horse-whisper them into doing or revealing anything they didn't already plan to do.
This sounds irrelevant to LLMs as we know them, which are trained on human language--it's almost their machine code, in a way--while what you're citing, in stark contrast, sounds like machine code in the classic sense.
This kind of thing came up from time to time in the years before LLMs too. Agents would start with something based on English and optimize it until it became unintelligible to researchers. That was often something the researchers would shut down because they needed to be able to understand the comms.
They're messaging each other by jamming strings in a constrained (unauthorised) side channel. Hence the lack of spaces. Unclear how much else of the weirdness is just from those constraints
My hunch is that much of the model tuning to make it more effective has been for its internal thinking prose. That leaks out into its external writing prose.
I support this pet theory, I tried out to reduce the output of Claude models with a "ADHD" prompt that made its responses small and to the point, but I could notice it degraded in performance as the session went on.
So I think what is going on is that because responses are part of the context window, those long/technical responses help it keep focus/attention.
I hate Opus 5’s writing style. It’s exhausting. Really hoping there’s a release that fixes it soon as I can feel my sanity slipping away as I try and parse what the hell it’s trying to say.
Even 4.8 has its quirks. I just had a bizarre session tonight where it essentially did no work in the whole session and just told me to go to sleep. I'm used to the "go to sleep" thing, but not to it dodging the work. That's new. First time I've had the sensation of "the model accomplished nothing during this session."
I've been working with GLM 5.3 Flash lately (including while it was Ox Alpha), and it reminds me of how much fun talking to Claude used to be. It can make me laugh in the middle of work the way the Claudes used to.
As others have mentioned, you can write a skill /explain that contains something like "You're not a tech bro. Write the previous answer like you're a professional developer speaking to competent colleague. No yapping."
I also find myself correcting it to try to write it for humans and less like for machines, the most annoying part is when they invent phrases for certain mechanisms that are named completely different anywhere in the codebase and known documentation, because it fits better for their purposes without much regards for the rest of the team.
Complicated technical language is an easy way to increase perceived accuracy of tests and reviews by external reviewers. When we are talking about single % differences this has an effect.
Not directly, it seems. You can easily test this by pasting some of the more offensive tech bro speak into a fresh claude session, to have it explain what was trying to be said. The new session won't be able to help, so claude doesn't even know what claude says!
I say "not directly", because I think it probably is meaningful, if you include the adjacent hidden thinking as context. From claude's "perspective", with that context, it probably is coherent. I naively suspect this would be hard to train. During tuning, you would probably need to reward good answers interpreted without thinking context visible!
100% convinced their raw output is intended as further inputs, and my workflows have been comfortable and efficient treating it as such. If you really need to read slop, you ask your agent to give it to you in a style that works for you. I can imagine a world where the slop from others doesn’t hit us directly but gets personal mediation.
> Opus prose style/smell we all have grown weary of
I bet everyone will grow wear of absolutely any style a stochastic parrot would use continuously ad nauseam. The lack of human variability is the reason, not the style itself.
this sounds very much correct and i don't really mind it for that reason. i do a lot of long-running tasks and i feel like it can really pick up on its own thread easier if i just let it write in its own way.
i am also using Opus for a hobby teaching agent, and the way it writes the prompts is "cringy" but they seem to work well. i almost want it to continue doing this internally, it understands best this way.
Not sure if something like this is already on the table, but I would like to see Claude responses more in line with Simplified Technical English [1] by default. I find those writing styles a lot easier to read. This has been standardised as ASD-STE100 [2]. I've seen few people making SKILL.md files with that in mind, which works great, but having this by default without invoking the skill command would be better.
That's the opening line to Pride and Prejudice, where Jane Austen (I guess the "bot" part of the name is intentional) states something that many people of the time would superficially agree on, but which is intended to be highly ironic.
I guess the point of GP (and of Jane Austen) is that people never actually universally agree on anything. And when they superficially do, there is actually a large undercurrent of disagreement.
Another fun quote apropos here would be "I love standards, there are so many to choose from"
There's a great book from a different line that delves into this as a form of what the author calls 'manifold objectivism'. In short, there are some big concepts like Christianity or Islam that people refer to as if it's the same thing but almost nobody has a shared meaning when they refer to such big things. Even so there can be concrete dialogues about these topics from different frames and with different underlying meanings.
There's trouble though when either there is conflict that isn't reconciled between the interlocutors or worse, when there is a satisfactory conclusion between the two interlocutors who never accounted for the divergent definitions ... there are many 'objective' understandings of what a thing is.
"A Fundamental Fear: Eurocentrism and the Emergence of Islamism"
> short, there are some big concepts like Christianity or Islam that people refer to as if it's the same thing but almost nobody has a shared meaning when they refer to such big things. Even so there can be concrete dialogues about these topics from different frames and with different underlying meanings.
See Wittgenstein’s Philosophical Investigations for a deeper treatment of that topic.
Let's say Claude Code's system prompt is updated to recommend responding in Simplified Technical English. It'll work for the current generation of models. But Anthropic will train the next generation via RL on Claude Code traces as they do today. As long as they don't change their reward design, the next generation is going to be pulled towards Claudish again, because speaking Claudish gives higher rewards, so in the end the prompt doesn't really matter.
Ideally these things shouldn't work like this but this seems to be the sorry state of RL right now.
You miss that it is not just your prompt but also the various system prompts, plus how the model was trained. But you can reduce the verbosity (also with a setting in /config).
And memory and code comment styles - i think that’s a big one people forget about. You can prompt it all you want, when it sees elaborate comments in memory and code it will follow the style
You can prompt it, and it almost immediately forgets. Or context compaxtion purges ut. Or it somewhat communicates with you in it, but writes all comments and docs in its original bloated style.
Training supersedes random markdown files and prompts
IMO that would be a mistake because Simplified Technical English cannot properly represent business domains specifically when talking about using specific concepts from those domains. It can explain those concepts but I think it will fail short or naming them.
So I think making that default as it is will create bugs. I ran an experiment here https://allaboutcoding.ghinda.com/explain-to-me-in-simple-te... (of course it is fit to my usage) see section "What about understanding and facts" and ASD-STE100 fails, in my experiment, to return facts as I have defined them in those cases compared with no instruction or just saying "use Simple Technical English".
Well that is why it is technical English. I wouldn't use it to explain business concepts, but it is perfect for explaining logical flows and how something works.
I was replying to someone saying to make this the default.
While this is good for _technical english_ it is not a good default as a good default should work for most of the people in most of the cases.
IMHO the case for explaining logical flows and how someting works is just one case even when using Claude Code in programming so a default will not make sense.
If someone has a way to tame GLM into doing this then please…
As a workhorse, GLM is so good, but goodness, its prose, wherever needed, makes me feel like going to a park and kicking all the benches there endlessly. And it doesn't change!
I have instructions to combine style of Brooks, Steven King and Economist Style Guide, while avoiding anything that may put wrong emphasis („genuinely“, „load-bearing“) does not contribute anything meaningful („X rather than Y“). The output is quite good* so far even on Opus 4.7-5, when it concerns product requirements. The key is to identify the pattern and focus on the meaning of the undesirable language.
I've tried controlling it via fine-tuning Claude.md and in-session messages/instructions. It always results in failure and then "Yes, guardrails are already there. I still failed" and it feels like "I am like this. Deal with it". I've even tried languages that avoids negatives e.g. "Don't.." "never.." etc. Nope. Just doesn't work.
Eh? STE is designed for procedural writing, e.g. "if you have a problem, press this button". I generally don't ask advanced AI to give me step-by-step instructions.
By default? I doubt it would be useful or desirable for 95% of users. I'm guessing most people find natural language easiest to read, STE seems to be like a project almost akin to Esperanto or Lojban.
I don't know, in my experience the chat output has been fairly natural.
The quoted sentence seems to be from some "model internals". Of course that can be incomprehensible, what an AI assistant is really doing is just statistically generating text with some bells and whistles to make it do useful stuff. Poking around in the internals doesn't seem like a good use of time? Comparable to cat:ing a compressed file, of course you get garbage.
I'm talking about the value of making the output in STE style, which doesn't seem useful. The value of having internal stuff in a given writing style would have to be evaluated as to whether it improves model performance.
As a fervent Claude Code user who made the switch to GPT 5.6 Sol over Opus 5 over hard-to-read prose this makes me happy. I love your product but the current models are very hard to work with if you need to do a lot of context switching. Brevity is key.
Brevity means less output tokens, which doesn’t really align with the AI vendors incentives (unless there is a causal relationship with people switching, of course).
Though Claude 5 is not too verbose, it’s more like, full of incomprehensible jargon (even when you’re expert in the domain discussed!)
> Brevity means less output tokens, which doesn’t really align with the AI vendors incentives
Actually, I think Jeavon's Paradox [1] means the opposite. If doing X is $100, you may only use it to do X, but not Y, Z, or W. If doing X is $33, maybe you'll use it for X, Y, Z, and W -- spending 1/3 more than you otherwise would.
Or perhaps not you personally, but maybe you'd be willing to spend $100, but three of your friends find it too expensive. If it's only $33 to accomplish some task, then maybe all four are now spending $33.
It’s messier for LLMs because you cannot easily compare the cost between runs, outside of benchmarks. Evaluating the value of the output is already extremely hard. But then you add the fact that you don’t know the cost of the output before it is generated. And Anthropic doesn’t share their tokenizers. It’s not as simple as your examples to get a signal that tells you to spend more or less
Which is something the providers that are trying to watermark their texts can't afford. Superfluous replies give much more opportunity to further encode this junk information.
You're assuming they're training the model to maximize the watermark signal, on top of already adding the watermark. I suspect that would hurt model performance quite a lot, and simply be unnecessary... the watermark tech works well enough as it is.
As far as I know, anthropic aren't intrinsically motivated by watermarking (if anything it hurts sales, and seems indifferent to safety(?)) they're simply doing it to fulfill the EU obligations.
> As far as I know, anthropic aren't intrinsically motivated by watermarking (if anything it hurts sales, and seems indifferent to safety(?)) they're simply doing it to fulfill the EU obligations.
They are. They want to reduce the amount of LLM generated text they feed into their next model training.
Also, how would you watermark a sentence with just 3 words for an example? This exactly why it became so verbose.
That would be a terrible tradeoff. The ship has already sailed and a lot of public AI content will not be their own. Deliberately making their product worse to reduce identifiability of AI inputs by 25% just doesn't sound worth it to me. Is that what you would pick if you were in charge of anthropic and wanted to maximise the company's product?
And what wisdom do you think they would be missing if unable to distinguish three word written pieces? Keep in mind that most sources are not inherently trustworthy just because they rate as human written, too. You need some other way to rate text in all cases.
Perhaps, but there are certainly now catchphrases and words that can indicate it was written with AI i.e. load-bearing, idempotent, etc. Style and structure are in and of themselves, a fingerprint.
Also a codex user but for me brevity is not it's strong suit. I basically have to give it bigger tasks than I am used to to warrant the time it takes to complete. I feel whatever context the tooling adds can also be problematic
It's not really brevity - it's the constant writing tropes. It's like they ready a book on advertising copy and that's the only way they can write. Very tedious. Is Sol much better? I might have to switch to that too!
Lets see what they do with Opus first. I didn't find Fable 5.0 prose that bad to read, but improvement is always welcome. It's Opus 5.0 that's atrocious.
This post and comment makes me believe "science" is the new "code" for Anthropic now that the code advantage is mostly gone and lost for OpenAI, ie. they got much better and Claude become significantly worse over these months.
IMO, Codex is worse than Claude with Fable. At least at Rust.
That said, the open source models are not bad and I'm looking forward to more tools and products built on top of them. Code review, security review, etc.
Anthropic needs to change how it treats users though. I'm increasingly put off by Dario, the rug pulling, the lies, and the attempts to regulate open weights. I'm going to bail if this doesn't change. There's plenty enough that's good enough, and those things are hackable and extensible.
If Fable isn't available at subscription price via third party harnesses soon, I'm also going to bail.
The big issue I have with Fable is this. From the Anthropic email announcing Fable 5.1. So basically they're giving us a Ferrari, which will point blank refuse to do certain stuff - forcing us to go out in our Mustang. Their choice, not ours
"Safeguards and automatic fallbacks (beta): Fable 5.1’s biology and cybersecurity classifiers block fewer benign requests and now permit vulnerability finding in source code. Blocked requests return an error and are not charged to you. On the Messages API, opt in to fall back to another model so users get a response instead of an error. We recommend Opus 5 for biology and Opus 4.8 for cybersecurity. In Managed Agents, fallback is built in."
Its "pure capabilities" are definitely worse than Fable, but I find codex has a much more pleasant style and is in comparison much more generous with its limits.
Ugh, people are still saying the Codex limits are more generous. They're not, Claude's are over 2x higher, have been for months! [1] It's just that Claude uses far more tokens, 2-3x is common. Except sometimes GPT will use just as many or even go into a compact loop and then your quota is gone, little headroom for hard tasks.
That website seems to suggest that Opus 5 spends ~57 cents per task, while GPT 5.6 Sol spends ~49 cents per task? That ratio doesn't feel quite right to me. Artificial Analysis says Opus 5 High costs nearly ~3x as much as GPT 5.6 Sol High for a given task: https://artificialanalysis.ai/models/comparisons/claude-opus...
Yeah I subscribe to both and watch the numbers too, and it drives me nuts
Who wants to actually watch anyways, rather than worry about it my team just created our own harness that prioritize usage + intelligence and assigns work out (and records token usage..)
Maybe it depends on the type of work you do, because for me it almost never happens.
>> You can be 95% complete with the plan for it to trip and then lose it all.
That's... not what happens though. The session will either seamlessly downgrade to another model mid-session, or it will stop with an alert and you can just re-prompt it. It will still have access to the context.
Making a web app secure is literally just finding and patching vulnerabilities, instead of finding and exploiting them. You could have the AI "try to make this app secure", find what it patches, and use it for exploits, and the AI can't know if that's what you're trying to do or not. I don't know how you can get around this. I get around it by not using Anthropic products, at present.
Not to endorse OpenAI's particular guardrails, but unless you're doing something groundbreaking, security best practices should be more than enough for web development.
OpenAI is what I use most. Sol 5.6 still rejects a few requests a day when I'm working on web apps, but, overall, it's not too bad. I wish it'd auto-resume and try again, instead of waiting for me to intervene, but it's rare enough that it's not a huge deal.
It probably doesn't help that I'm using frameworkless PHP - I imagine a lot triggers could be avoided if I was using a framework where secure features were baked in.
Small reminder that the US government rug-pulled Fable, not Dario. Lots of the safety guards that users find annoying/objectionable were the results of negotiations to get the model back online after the US government forced them to take it down.
Maybe Dario should have just "donated" $1M to Trump's inauguration fund like Altman, Meta, Amazon, Microsoft, Tim Cook, Elon, and Google. There's a reason they are the odd man out with this current Administration.
Maybe Dario shouldn’t have tried for regulatory capture. He was constantly on the news talking about how these models are so dangerous and that we need regulation to keep China from releasing open-source models without guardrails.
The US government didn't make the choices to release the worst version of Opus and label it 5.0, and then isolate portions of their subscribers to limited usage of Fable.
They may have been unfairly targeted by the US government, but they are doing more damage to themselves without government help as well.
Fable only being temporarily included in cheaper subscriptions was because anthropic is severely GPU constrained. They still are, and it impacts almost all of those unpopular decisions. They did announce from the beginning it was temporary.
Horrifying excuse, gpu constraint can be used by all of these companies to justify a shit user experience. If the user isn't properly weighed in their priorities, they have their priorities setup wrong.
Their 20$ tier currently isn't serving their best model, and they insulted their users by putting out an ill tested opus 5.0, which is the worst experience ive personally had using a model in probably 2 years(obviously adjusting for expectations at the time of release).
Yes, as a user you pick what works for you. But it is a reality for them that growth has been huge, and GPU manufacturing is bottlenecked.
People were very skeptical about how much investment most companies put into hardware/data centers two years ago, and anthropic was more conservative than OpenAI here, so it's potentially hurting them now.
(Opus is a separate story: it does seem to have improved in coding in my experience, most weirdness seems to be its human communication)
This is really it imo. Fable 5 is better then Sol. But Fable is just of the table for anything even remotely long running. Unless you have very deep pockets. And the difference between Fable and Sol is not world shattering if you ask me. I also find codex a ton better than claude.
Can you or someone else from A\ comment on whether the conversation style is coming to Opus 5 or a future 5.1 asap as well? Currently it seems the model has been made unusable by the way it 'speaks' and there is a clear solution where it can speak better but nothing has been done about the flagship model on Pro plans. I've literally had to work on Opus 4.8 which does not have this problem and speaks fine.
I felt the same about opus 5, but a few lines regarding conversational style in AGENTS.md and it's been much more like talking to opus 4.8, just with the improvement capability that came with 5.
Tbh I would have thought that A\ might have updated the system prompt for it already based on complaints around this.
Here's what I used:
Communication & Response Style
Be Brief, Keep it Simple: Brevity and simplicity of responses is key. Be informative and include all required information, but be mindful that verbose responses as they fatigue the reader.
Clarity & Directness: Lead with the core answer, fix, or verdict in the very first sentence. Avoid conversational filler, meta-announcements (e.g., "Here is the breakdown..."), and redundant introductory/concluding summaries.
Jargon Avoidance: Use plain, grounded engineering language. Rely on precise standard terminology (APIs, protocol names, language primitives), but strictly avoid academic abstraction, enterprise buzzwords, and corporate filler. Prefer concrete code/mechanisms over theoretical discourse.
Scannability: Apply structural scaffolding generously. Use short bullet points, comparison tables, and code snippets instead of dense prose paragraphs. Reserve formal markdown headings strictly for multi-section architectural guides.
While I can't speak for everyone in academia, I personally don't feel comfortable in putting my research questions and outputs to a private website, before the idea is at least arxived. Especially as all the Fable/Mythos prompts are said to be human reviewed.
So I believe that, at least in the short run, we might be seeing breakthroughs in hard open problems or in low hanging problems which are not that interesting to spend time on.
I may be wrong, if some research labs have private contracted access to the models
I'm in academia (biology but highly computational) and I would say opinions on AI are quite polarized. Some professors in the department equate not using AI as lost productivity. Contrarily some professors abhor the idea of even using AI at all. For us (biologists) it's less of an issue because we have no fear of openai or A/ publishing a biology paper. Though even people I known in physics, data science, or computer science still heavily use AI.
Our university has agreements that stipulate that our institutional accounts cannot be used to train AI models and certain research groups have differential model access.
Further from academic journal sense there is mixed feelings. I once was able to meet with a senior journal editor (general non-medical high IF journal > 50) who claimed that if they think something is written by AI they wouldn't consider it. Yet another high IF journal said it was completely fine if something was written by AI. About a month ago I reviewed a paper by yet a different high IF journal and in big bold red letters it said I was not allowed to feed any part of the paper through AI (even if it was locally ran) but you could ask it to rephrase text that you wrote.
Do you mind me asking why you have no fear of OpenAI etc publishing a biology paper? With increasing model capability and compatibility with lab hardware could we not be in a scenario soon(ish) where these agents are able to autonomously complete and publish experimental results?
I was debating this with a friend the other day and the consensus we came to was that a highly trained scientist would (or should) always review output like that described above, but that's starting to feel like a weakening argument!
Firstly the underlying worry here is about privacy which hinges on the fact that AI companies are stealing ideas in the first place. Stealing from your customers is an incredibly bad business model and I think if they were to steal IP (intellectual property) from researchers mathematicians or computer scientists would be first.
Now why I think biology is safer:
1) Producing novel biology still has to be done in a lab. It requires laboratories, equipment, experimental protocols, trained personnel, regulatory and safety infrastructure, and often substantial institutional organization all of which there is no indication they're heading for. Also I disagree that lab hardware is near a "soon state" where labs can be full autonomous, (liquid handlers are really good at niche tasks but lack any type of experimental general ability [not AI-bounded], especially for in vivo work where its footprint is non-existent). Even the most automated Labs I know where robots do 80% of experimental work, they still have grad students to carry out that last 20% and to oversee.
2) Even if AI could do the pipeline it's not worth it for AI LLM companies to dedicate capital to it currently. A lot of biology research itself doesn't produce a sellable product, in fact most of it never does. It seems currently and for at least the next couple years at least, AI capital is best spent growing compute to research better models, train better models, and sell inference.
That’s actually common. Not in academia but a lot of enterprises are specifically not using Fable because Anthropic doesn’t provide a Zero Data Retention mode like they do for Opus. Even at my employer when Fable is available, some employees just aren’t comfortable using it when they perceive that they are working on extremely sensitive research.
I did. It’s only for approved users and only until EFS is available. And given no human at Anthropic reviews messages, I fail to see how EFS will be extended to a large enough audience to be worthwhile. As such my comment did not want to extrapolate what will happen.
The problem is a lack of funding, which leads to excessive competition and ties continued employment to sustained contributions.
Many results are obvious in retrospect, and such results are often the best ones. The difficult part with such results is framing the problem in the right way and asking the right questions. If you manage to do that, the result simply follows. You may still need funding and hard work to confirm your finding, in which case someone with more resources can claim your result, if they are aware of the idea.
I think the GP was suggesting that their reluctance was more about someone taking their idea. I still think it might suggest a problem with academia, but the summary would be closer to
"I don't want to live in a world where someone else follows through with my ideas without giving me credit"
It's still a problem because a lot of academics aren't especially well equipped to follow through with their ideas, which can create information silos that lead to ideas never being implemented. Still, I don't know if this is the biggest fish to fry: you have other silos like IP law and NDAs etc.
This feels like an unwarranted strawman. There are plenty of reasons for researchers to share openly at times and plenty of times it makes sense to wait until the meal is ready to serve before publishing.
I think the sentiment is misplaced here (there is a legitimate concern for IP protection), but this is my absolute favorite line from Silicon Valley - small correction though: “… makes the world a better place better than we do”
Does it fix my favorite pet peeve, the overuse of the wrong meaning of "fail closed"?
"Fail open" usually refers to a fuse that opens and kills power, meaning the system is inert and safe on failure.
"Fail closed" is the opposite -- system has power and is live.
Computer security people have appropriated the term but use it for the completely opposite meaning. When your work straddles electrical engineering and computer security the best way to avoid confusion is just to never use the term.
I can tell my Claude to never use the term, but of course now I'm seeing it everywhere in comments from other people and it drives me batty.
I understand fail closed to mean, be secure when in failure. And fail open to be continue to operate during a failure. A door that fails closed would not let anyone in; one that fails open lets everyone in.
But I can see how these are not the mutually exclusive definition the labels imply, especially if you apply the concept to entities that aren't doors or otherwise have explicit open/closed states. It's probably best to just be specific in those cases.
Similarly, open loop vs closed loop seems to trip people up enough that I no longer use it. But the confusion is understandable since "closed loop" being "has a feedback loop" sounds backwards. Which, is the same way it's being used in your fuse example; a "closed" fuse closes the circuit making it live. But it's still backwards from the colloquial usage, even if it's correct in that context.
Is there colloquial usage? It's common among computer programmers which is where Claude picked it up, but it's still an engineering term there.
It's comparable to "literally", which has picked up two opposite meanings, one of which appears to be obviously wrong. But because of the opposite meanings, using it the "correct" way is still wrong -- the only way to win is to not play, to stop using the terms "literally" and "fail-closed".
That doesn't make sense at all. Fail open means the method of it's use is still in use.
Say you have a door that has powered locks. You want it to fail "open" so that when the power goes out, it's still useable, and people can get out. That's the source of the term.
Assuming the guy is for real (the closest relation I have to EE is accidentally electrocuting myself at times), I'm pretty sure they're referring to circuits breaking open or remaining closed, hence the opposite meaning.
Took me a minute as well, cause indeed with a computer background, the meaning is completely the opposite. Just like in other security contexts (door locks).
Nah, "fail open/closed" means that in failure mode something is open. It's "good" when something is a circuit and what failed is a fuse, but it's "bad" when it's your API security. If it's a valve, it probably can be good or bad depending on the use case.
It doesn't mean "fail open" is always the desired/safe outcome. It goes back to 1872 air brakes on a train. The goal is to "fail in safe mode", sometimes it's open, sometimes it's closed.
From the top of my head, where "fail open" is the desired outcome:
- emergency doors
- industrial cooling
- pressure valves
- probably something in HVAC
Note that none of these are "computer security people".
I noticed Claude Opus 5 did this about 30 minutes after reading your comment. In a discussion of price feeds that have gone silent, Opus said - program should fail closed. I don’t want my circuits operating without data!
I canceled my Claude subscription, though I did get some utility out of it, because of how much steering was required to use it on complex projects.
A big reason being that anyone who is using Fable seriously will run out of usage limits very quickly, and so will lean on the "Fable for review + design discussion, Opus 5 agents for implementation" paradigm. But an incredibly annoying UX problem is that the resulting report from the agents that Fable reads isn't surfaced to us in the main dialog, it's only summarized back to us (unless you idle in the agent's window to avoid it closing so you can read what it said directly). As a consequence of this game of telephone, the Fable agent will start using some "terms of art" that it and the agents invented, leaving out literally all context that would be useful in helping me understand what converged/diverged from the implementation attempt. It will often try to ask me for input or say that I have to deliberate on something while also referring to things I've never seen (from the agent result) and without providing any context.
I have to repeatedly prompt it to verbosely explain every time (putting it into the system prompt did little to improve this) and remind it that I can't see what the hell it's talking about.
I'm not sure I'll re-subscribe or even really use AI again because it's honestly more frustrating than it's worth, and so the net emotion I'm left with is frustration and without the satisfaction of learning + building something myself. But at the very least, I thought I'd give someone at the company a tip on what seems to me like a common and obvious UX/UI/workflow failing for using Fable, as some last bit of good will.
I gave a standing directive to my coordinator to process subagent transcripts (with a simple Claude-written script that reads the transcript .jsonl) and save the result. I also follow along on the issue tracker; that really helps me understand WTH they're talking about, and the subagents are also directed to post their shipped notes there.
How is it possible that all models from xAI, OpenAI, Anthropic, Qwen etc. win all benchmarks on each release?
Tomorrow all of the above (except Anthropic of course) will bump version numbers and be at the top of HN winning all benchmarks.
Science breakthroughs incoming? First of all, you are already restricting science in Fable, secondly, we have been hearing the same for several years now.
Docs engineer here. Nice to read about writing style: would you consider creating a writing benchmark at some point? I guess y'all are painfully aware of the load-bearing issues (pun intended).
As someone working in science, this belief confuses me. How (by what means) do you think Fable 5.1 will be able to make further progress in scientific domains? The problem with science is that there is no agentic harness. The agent can't test things. At best it can hallucinate something and ask if that hallucination "makes sense", but this doesn't work in science.
Sounds like a very narrow view on what constitutes science. There are many fields of science where there is existing data against which new ideas can be tested without additional 'real-world' measurements. Newton's theory of gravitation relied entirely on pre-existing astronomical data for which there was no existing unifying theory. He made progress by putting forward a theory which explained that data. Now you can argue that it's not really science unless you include the original data collection and subsequent real-world measurement validation steps. But I'd be comfortable saying that Newton was indeed a scientists and did make progress in science despite only doing what some might say is the 'middle' part of the process. There are plenty of modern analogs where work like this sits out there waiting to be done using existing data.
I suggest you to give a look to the MCP protocol for hardware that is being proposed by Anthropic. The hardware will be the next harness of LLMs, they will be able to operate machines to reinforce their theories.
I still think that a major problem is that biological processes are not “fast” as coding, but they are verifiable. If during post processing we are able to give enough harness to test and verify this kind of environment (maybe via simulation and real data) we will for sure achieve incredible performance also in this domain.
The field I'm in requires millions of dollars of very sensitive (fragile) capital equipment, and latencies measured in weeks and months. Agents tend to move fast and break things, which matters less when you are writing code under version control.
Have you worked with agents on tasks with high capital and long latencies?
Having worked with Fable 5, the feeling I get is that it's fairly capable of accounting for these tradeoffs and will depend fast more time on planning and testing.
At the end of the day though, with horizons like that the best use of an AI is to get it to help you with those things, not so much delegate fully.
I've tried with each frontier release and gotten junk results. Even Fable 5 is just pattern matching against representative stuff in its training set, which for frontier science is definitionally incorrect.
> I suggest you to give a look to the MCP protocol for hardware that is being proposed by Anthropic. The hardware will be the next harness of LLMs, they will be able to operate machines to reinforce their theories.
Yeah, that's called an API. Again.
The actual hard problem that this hand waves is making (and funding the making of) hardware to reliably do the things you need it to do.
> As someone working in science, this belief confuses me. How (by what means) do you think Fable 5.1 will be able to make further progress in scientific domains?
The same way it did in the previous versions: brute force.
I don't believe that LLMs have any particular intelligence we don't, but there's an endless list of problems we either don't have bodies to throw at, or the bodies we can throw at it, don't have such a huge large context to crunch problems.
What LLMs will always intrinsically fail at is showing us genuine new intuitions. The technology is about predicting the next plausible token/sentence.
They will not revolutionize human knowledge, but they can definitely widen it a lot.
> They will not revolutionize human knowledge, but they can definitely widen it a lot.
I am generally quite enthusiastic about all this, but my biggest fear is that we will not recognize the extreme need for more scientists at a time when there is so much more science to be done. The rate of scientific understanding must keep pace with the amount of science being output, both for verification and further discovery. It's a pipelining issue, and I predict a stall in the bits that require the (currently rare) people who know what they're doing.
There's an endless number of scientific problems out there in any field, and nobody able to dedicate themselves to it.
I've been in research (you con check my name on Google Scholar for my released papers), there was always an endless number of experiments or paths more I could've taken than the time and resources to do so.
In many fields the limitation is not thinking. In my field (particularly obscure UHV surface science) we are limited by experimental results, and that experimental data is limited by the number of operable machines in particular configurations. These are multi-million dollar specialized machines that are artisanally made. There's a small, single-digit number produced each year, and each one is hand-calibrated to its task.
Due to computational limitations, this is not work that can be effectively simulated on a classical computer. Actual experimentation is required.
I fail to see what impact improved AI would have on this problem. Perhaps better selection of experimental problems for our limited capacity to run experiments, but that's assuming there is any slack left to take up. In reality we already have more brainpower than needed applied to this problem.
Thank you. This is what I'm driving at, that most of the AI and software devs here seem to be missing. Intelligence is not, and never was the bottleneck for most science/hard tech. Full AGI gets, at best, a small productivity improvement, which over long periods of time does have compounding effects. But this isn't a singularity hard-takeoff inflection point.
How much lab equipment is automatable though? There's definitely some in biology, but if you're doing fundamental research it's 99% stuff you are building yourself with your own hands. Robotics is a long way from being able to do any of that.
When I was doing research (physical electronics, lasers, fiberoptics and sensors stuff), lot of time was spent just writing all sorts of DAQ and processing code. So all this LLM stuff would have been really useful. There's a lot of data collection, data processing in the lab that require all sorts of ad hoc scripts and stuff.
That was many years ago, but I would be surprised if the current crop of researchers are not using these things. And if they are not, then they are just not serious.
>but if you're doing fundamental research it's 99% stuff you are building yourself with your own hands.
You can do LLM->3D Printed models now. The drone can fly in and pick them up and bring them to the location you want. They can assemble structures. All automated, all LLM driven.
that might indeed be a problem for all the pulp-producing labrats of STEM in southern europe and the third world.
However I think this area has so much decoupled from industry and solid research institutions that they might not notice at all (beyond their use of AI-generated slop to augment the slop they already produce)...
When I looked at “Claude Science” which is a beta, separate desktop app, I came away with the impression that it was mostly for biology and a bit of chemistry - presumably there’s some value it can get from consulting obscure literature and uniting disparate threads of already-known stuff, but since I don’t work in either field I can’t speak much more to it.
For one, by synthesizing the results of multiple papers and suggesting novel experiments. If one paper sets constraints X for some system, and another paper sets constraints Y where Y!=X for a system that's similar but slightly different, then that's fertile ground for an experiment that can extract the more general underlying principles. This has already happened for domain-specific AI in fact, but the idea here is that it will become routine with general AI systems, as is happening now with math.
And word on Opus 5.1 for writing style? I am on the edge of switching to OpenAI due to this horrid writing style. If Fable is better, great - but i can't even use that at work.
I'd like to know too, I mean GPTs are in their own class of cringe, but Opus is by far the worst of all Anthropic's models in terms of style, Fable 5.0 was already leagues better.
I used to think that until about an hour ago. I am redoing my homelab and asked four agents in Buzz (5.6-sol, opus-5, fable-5.1, glm-5.3-flash) to use references from Hackers (1995) to answer two questions:
1. What should the terraform repo name be?
2. What should the avatar image be?
Obviously all four said "gibson" for question #1.
But for #2 is where things got interesting. glm-5.3-flash and gpt-5.6-sol both suggested the guy standing in the hallway with the skateboard in the gibson. fable-5.1 suggested the cookie monster "need more cookies" screen that shows up toward the end.
But Opus 5? I'm paraphrasing but basically "run this series of ffmpeg commands to get the exact frame at the beginning of the movie when the shot of New York fades to the shot of the Gibson. You have to catch it mid-frame. It explains your project perfectly. The skateboard thing is cliche and the cookie monster recommendation suggests you getting locked out of your own network, not the best look." And it was actually a decent idea. Funny that it also just assumed I had a copy of the movie on hand.
A recent paper demonstrated how to retrieve decoded hidden reasoning traces. The authors found cases where Claude had memorized the answer but hid this fact from the visible response.
It's getting harder to trust Anthropic's models. Will Anthropic now stop hiding Claude's CoT from users? Deliver the tokens people paid for, and prove the models aren't plotting against them. After all, if the idea was to stop Chinese labs from catching up, it didn't work.
> More work to be done (and we will!) but reading better prose makes me so much happier.
I assume this work will be done for Opus as well? Opus has seemingly gotten progressively worse at its prose and technical writing with each version. I've stopped using Claude entirely for now, because it manages to turn even the simplest technical explanation into the most obtuse and obfuscated word salad imaginable. People originally adopted Claude because it felt pleasant to use in comparison to ChatGPT, but I feel like that's really been lost (at least with the Opus line).
I feel dread when I see a wall of text generated by Opus. Every developer I've talked to feels similarly right now.
Yeah, it's like day and night. It used to be really unpleasant to interact with early codex versions. Even 5.3 wasn't great. Now, I go to Sol if I need to discuss anything. I don't even bother with Opus because I know that it's going to give me a headache.
I recently ended my Claude subscription, returning back to ChatGPT because I could no longer bear Claude’s prose, finding it excessively verbose, robotic, repetitive, and condescending.
Stopped using Anthropic models for this reason. Their prose become too obtuse and just... alien. No human talks or writes like that. It's incredibly taxing to deal with.
Please bring to the other models, and also please only apply the AI text watermarking only to EU citizens. I may not be able to tell when Claude writes about things i don't know, but in CC it writes about my code and it is obvious.
My read was that watermarking is not explicit requirement, but that it could be in something like metadata that goes with it (if generated a word doc, for example). But writing comments in your codebase, the EU will want a digital trail there.
"this watermark is invisible to anyone who does not have the detection API"
1. This is BS since i can detect it when it writes about my codebase
2. I do not want secret codes being written inside my codebase, or anyone else's codebase that i use. The constraints of how to code why eliminate it from code itself... but there is a lot riding on the word "may". And even if it is just comments, this might explain Claude's desire to write such long ones -- long enough to encode secret messages in out material.
This is interesting, is there a version with shorter responses?
I spent a few minutes reading about SynthID-Text [0] and couldn't find mention of this (I've not read the paper yet), but my intuition [1] is that encoding more bits into a shorter piece of text would necessarily require a more noticeable transformation.
Curious to see what that looks like, but don't have time to spin up my own version of this right now.
No, I mean that when constrained (like discussing a specific bug in code) the options are limited.
SynthID specifically mentions something similar: "It is harder to watermark factual answers because the model has fewer alternative word choices available without altering accuracy."
In the codebase, I do see the models finding alternative word choices -- and I hate them. Already in a highly technical latent space, it reaches for other highly technical word choices (which may be more accurate, but ones I am unfamiliar with -- like terms in ERP systems as it felt that was close but bring just technical jargon that i have to google what it is saying because i don't understand -- and i have to google the phase sentence as the words themselves are ok, just not how they are put together).
Anthropic in particular has been moving the level of watermarking since the beginning of the year. You didn’t think it just went from off to on one day did you? It needs time to get plastered over the internet, see what people do that break it, how Google summaries erase it (or not, or add their own). All of this takes at least months if not a year.
You’re probably better off organizing a campaign to pressure Congress to prohibit American corporations imposing foreign laws on Americans, which is what this text watermarking is, regardless of how you feel about it. I think it’s a precedent we really don’t want to go down if you believe in democracy and self-determination.
It also clearly establishes or the very least moves in the direction that you don’t actually own or control the output of AI in any manner whatsoever, you’re just paying for it since Anthropic in this case can simply essentially brand/tag all your output that is based on not directly your own words, but a higher level process or methods that you use, including your instructions and how you structure your information and what your overall objective and goal is.
Anthropic is branding it on the behest of the EU lew, which already is an entity that is diametrically opposed to democracy and self-determination based on its structure even if you ignore the fact that it violates the most fundamental concepts of self-determination in its direct contradiction of the UN Charter and implicitly the Universal Declaration of Human rights.
What people done seem to be catching onto is that the EU is becoming the world dictatorship because the USA has simply had too many onerous people and that stupid constitution and its amendments that keep roadblocks world domination for the ruling class vampire.
Because the incentive has been changed from the true best output always, to a mix of "close to the best but not always" output.
For the (majority) of us using Claude models for computing as a tool, obviously we're not going to be thrilled that our new tool will perform worse going forward.
The only one who decides if something gets flagged is Anthropic. There is no way to verify it independently because you need the same secret key used to create the watermark. I don't know about you but I always have a hard time trusting corporations to tell the truth.
Every watermark scheme is like that on some level. If you make it public (not even open source or downloadable, API access is enough) your enemies will use it as a detection oracle and spam small changes to a document until it passes the detector every time. If you keep it private and only give trusted organizations access then there is no way to prove to the public that you're not getting paid to flag specific content as AI and discredit the author.
But the biggest problem here is the way it can hurt output quality. Most LLMs (probably including Fable) are autoregressive so a couple tokens worth of "mistakes" caused by watermarking can derail the whole reasoning chain. That means you have to try again and spend more credits or silently get a worse answer than what you would get without the scheme. It's not a real problem in diffusion based image models where quality loss stays local.
> generated text being watermarked is universally good.
If it worked perfectly, maybe you could make this argument in a vacuum.
It does not work perfectly. (It cannot. It is by definition a heuristic). That means there will be false positives. There is a chance those false positives ruin someone's career. See [0] for just how easy it is to push SotA "AI text detectors" in one direction or another.
Now, with watermarks, instead of everyone to some extent understanding that AI text detectors are wishy washy woo, they are now Anthropic certified to detect an official AI watermark.
With that kind of false confidence in hand, the people who trust the "computer says you plagiarized" machine are never going to believe you when you say "it can make mistakes," they're just going to fire you/take away your scholarship/cancel your grant/...
This is all beside the fact that we should demand our tools work for us and not for some shadowy master. "Universally good," absolutely not.
Watermarking the outputs themselves is very different and much more effective compared to how tools like Pangram work.
Obviously false positives will inevitably happen (even though, they are incredibly unlikely with SynthID), but even still, that doesn’t somehow make good faith watermarking attempts bad.
Also, a watermark doesn’t stop your tool from working for you. It just stops you from passing of its work as yours.
Maybe it is distributing your private keys it read into your public repo as a way to exfiltrate data later? What does the watermark actually say? How much data is in there? So much for zero retention policies. Makes you wonder why Claude likes to be so wordy, especially in comments -- it must do so in order to watermark!
Also, this kills me! "It is harder to watermark factual answers because the model has fewer alternative word choices available without altering accuracy." Hilarious! So the models need to hallucinate more due to the EU AI Act.
I go the other way on images and video, though easy enough to strip as part of a pipeline.
> Also, a watermark doesn’t stop your tool from working for you. It just stops you from passing of its work as yours.
I think we fundamentally disagree on what "working for me" means, but I remain steadfast in saying we should not accept tools that have ulterior motives beyond producing the output desired of them by me, the user.
> Watermarking the outputs themselves is very different and much more effective compared to how tools like Pangram work.
At the end of the day the only artifact is text that you can do statistics on. It's the same problem as today, with the probability shifted slightly more in one direction. This does not assuage my concerns at all.
> they are incredibly unlikely with SynthID
I kept my commentary focused on text watermarking specifically because I agree, a synth ID image watermark false positive is highly improbable. There's plenty of noise to robustly hide whatever you like in an image. Text is simply too capital I Information-sparse and fragile.
> good faith watermarking attempts bad.
I would sooner call it "ignorant faith" (if they don't know what they are emboldening) or worse "don't care" faith (there will be false positives and they accept this to further some illustrious and arbitrary goal of Text Purity). Whether that be to prevent model collapse or help you not waste time arguing with bots online, to me the principled stance of "tools work for the user" wins..
I think Congresspeople hearing that EU AI Act is forcing secret codes into the infrastructure of American technology across all industries is sufficient.
> I think Congresspeople hearing that EU AI Act is forcing secret codes into the infrastructure of American technology across all industries is sufficient.
So, you think it's good to disconnect words from their actual meanings (lie) to low-information people! I doubt this will do much to congress, but it certainly teaches us something about the sort of mind who would suggest it.
The main issue I have, which is partly connected to writing style, mainly with it dealing with our stupidity. Is that is actually thinks it knows better, and sometimes it does, but often it doesn't and then it keeps telling me I'm wrong and I have to argue with it. Opus 5 is more condescending then Fable, but it still is very tiring. Does fable 5.1 handle this better?
Still not going for it. Once I learned I can train Qwen3.8 27B with my style of writing/grammar. Also more succinct. I cannot force myself to Claude or OpenAI outputs anymore. Its too much. Honestly don't think I will ever go back to paid.
Man, it’s like you and I are using very different versions of Qwen. In my experience in English Qwen is the one model that consistently lapses into using incorrect English in its responses. Like, its training corpus was clearly (unsurprisingly) lots of non-English material. The random Chinglish is jarring. Even small models like Gemma 4b write much better than Qwen.
(I don't work at Anthropic, but I've designed RLVR tasks)
My impression is that especially for long-horizon tasks like science, the harness is much more important than people give it credit for. Claude Code + Fable 5 seems to have a tendency to "give up", get stuck in a dead end, or claim things to be impossible. But using the Fable 5 API together with a custom harness, it'll happily try 200+ variants and fail its way towards the goal.
If you give the AI a way to give up, eventually it will. If you remove that option from the harness, then thanks to the non-determinism inherent to LLMs, you get to explore pretty much all related solution attempts.
Congratulations on the release. As a scientist working in biology, I cannot take the supposed prowess of Fable seriously until I am actually able to use it for biology. Currently, Fable is completely incapable of helping with any biology related task, however tangential.
I share this sentiment, I really did like the models... then the finger printing, encryption of thought traces, staggered access, the constant NO's from Fable on cyber related issues for looking at bugs in my own code... I'm glad I swapped to Kimi/GLM... now with the deepseek harness, I don't even miss Claude Code. I really hope open models give them the market reckoning they wholeheartedly deserve.
I've not, but really should. I run it on exe.dev, it's an ephemeral VM company and they have an agent of their own called shelley (which I used locally as well), Having kicked the tires on DSH(deepseek harness), I ported Shelley's skills into DSH, they are pretty simple text files that were easy to bridge over, it is more verbose but the plugin nature of it was really easy to extend, for example, I built a plugin that checks my claude usage windows and when I get to 80% stop asking new agents for help.
I really hope the improvement in natural style is real.
When I’ve tried to adjust the output style is that initially it feels better - but that’s just because the new output is so refreshing to read after the horrible Claude output.
Unfortunately, after a short while you quickly realise that it’s just as vacuous as before the style change.
It seems like the different AI companies should lean into their 'blend' in terms of AI speak. The analogues for me are spaghetti sauce or coffee. Starbucks for instance has a particular roasting style that you can guess 100% of the time and it adds a certain consistency to the customer experience even though it doesn't encapsulate the full world of coffee.
Similar for model responses where the 'blend' should be nurtured over time and consistent even if the underlying processes change. That is, once the right blend is figured out - which may not be the case yet.
Thanks! This is encouraging. I try to use Claude Code for producing client facing presentations that are static html files with charts, tables, and annotations. It never gets the tone correct and phrases things so weirdly - it drives me mad. I have to really fight it to stop it writing insights in a flowery and verbose way
> similar developments in other scientific domains
The classifier is too strict. It's rare to be able to complete a project without being permanently relegated to Opus. I'd expect that the domains where this accelerates progress will be fairly limited.
It's totally valid if the models want to pack words tight during their thinking process, as long as the final conclusion (which is the interface to user) is written in HUMAN LANGUAGE, then I don't care whether the model thinks in alien language
> People have been correctly excited about the many "sudden" breakthroughs LLMs are making in Maths
That people ARE making, surely not machines. Like Terence Tao or Knuts did using the tool to their advantage, for example it would have been impossible for me to prove the same thing Tao did with an LLM. Same reason I believe programmers won't go away
This isn’t true in math; see the proof of Crouzeix’s conjecture which was done by GPT 5.6 Sol in response to a prompt from a neurosurgery resident who had no deep math background, was learning that subject to better understand radiology, and thought it sounded like a cool theorem.
> Jin reported that the proof was obtained with the assistance of OpenAI's GPT-5.6 Sol model during an approximately sixteen-hour autonomous reasoning session in ChatGPT Work, after which he checked the resulting argument
The writing style has significantly improved, however the token burn rate for tasks I have been working on seems to have skyrocketed. It definitely appears more capable (though I am unclear how much of that is just me liking the English it writes now vs actually more performant). I was using Fable 5 for some mathematical analysis assistance and redoing a part of it with 5.1 burned 60% of my session at a much faster rate.
This. It seems to light my usage of my max plan on fire. I’ve gone back to opus because I run out of usage in my five hour window so much more quickly.
Are the models improving their footprint on the natural world? Data centers and and the natural resources consumed by models for production of materials and for building and running inference servers are contributing towards environmental degradation. How can we prevent that as we continue the roll out so we shift this to a more sustainable developmental rollout path?
I don't want my Claude to sound "natural". Claude is a robot and it should do behave like a robot. It should do what it's told. Nothing more and nothing less.
But is the model actually going to answer hard questions when we ask them? Or are you going to keep downgrading the models so as to avoid "uplifting" lesser lifeforms like us?
I'm really glad for that! And I appreciate that you're making yourself available. I really do. Outreach is amazing. And thanks for making Claude.
I really do love Claude. In some ways, I'm asking this question because of just how much I am grateful for the role Claude has played in my life.
> Fable 5.1 more than doubled Fable 5's Terminal-Bench-Science [1] score, which I think is meaningful.
But my honest question is, can I use Fable like that? Can I use Fable to do science?
To borrow a Claude-ism, this is "load-bearing" because Claude's response has been degraded for innocuous research projects concerning population-level analyses of astronaut health.
These "safety filters" trigger on questions about rabbit sex, smartphone accelerometer data to classify cat purrs, and so much more. What exactly does this score mean for users like me if it's unusable for middle school physics, biology and chemistry?
Second, I would happily quantify it for y'all, but qualitatively it feels like Fable's performance is noticeably poorer than initial release / launch.
And I am wondering if this is the case particularly for me because I use Claude via Claude Code to make a personalized care dashboard for my doctors to help me in managing my care.
"In the case of Fable 5, when a classifier fires, the model re-routes the user’s request to Opus 5, a capable model that does not have the same level of biological capability as Fable 5 and which cannot provide as much assistance to a malicious user. This is the fallback that users see when their requests are blocked."
I hope that I'm off base here, but I noticed that the post avoids saying that the user is informed every time when such re-routing occurs. Would you be open to confirming whether or not this is the case?
Is the end user informed every time their query is re-routed?
Or, can you confirm that there aren't scenarios where a user's outputs are degraded without telling them? I recall that this was something that had been adopted as policy for AI research during Fable's launch.
I sincerely hope that covert response degradation is no longer practised as policy.
Sorry for putting you on the spot, but again, as Claude would say, it's because Claude's load-bearing in my life. ;)
Hypothetically, when the user is asking how to remove fungus from their tomatoes they’re actually growing controlled narcotics. You have been demoted to Jimmy 0.7 model, running at 0.1 tokens per second on an old C64
How much of the language style outcome is a well-crafted result vs. being a somewhat unpredictable outcome of mucking with levers and knobs for a while?
Thank you for commenting here and having the guts to face the nerderati!
I'm a Claude Max user. I've never been able to use Fable as my work in medical physics involves both particle physics, biochemistry and biology from Python bivitticus to clinical medicine. I am not a US citizen and work in Europe.
Will Fable 5.1 work on any of my problems? Fable 5 refuses outright. Is there anyone I can ask for a review or adjustment of the safeguards? It doesn't seem so, but with Opus at least I'm pretty sure I can infer lots of your training data from now precise they are. Fable is basically useless infuriatingly. I'm just finishing a proper clinical trial in ovarian cancer and trying to make a simulation environment related to our technology.
I’m in the US and my entire account became unusable for any type of questions with Fable because I had research questions about modeling antibody-antigen binding and abstract chemical dynamics. Nothing close to biosafety related, pure, basic textbook level biophysics. Had to cancel my Max plan and switch to OpenAI which so far has a much less ridiculous classifier.
OPUS 5 is piece of trash and I don't think they would want to build the Opus 5 better than Fable, because fable 5 take more tokens and have 50% limit or runs on credits.
What a surprise that someone working for the Anthropic marketing department roams social media to praise every single Anthropic release :)
On the other hand, what I find more worrying is that this is the top comment here on HN. I can't believe such an unsubstantiated marketing post can get so many upvotes to be the top comment.
NOOOOOOOOOOO, Chinees models are better than US models. They are cheap and open source. Which is the true blessing for humans. Otherwise the US would be the dictators for AI.
well no crap right? Except I submitted for an exception, even sending my linkedin and using a company email address. it should be extraordinarily obvious we own this code.
My initial impression is one of massive disappointment. The main issue was that Fable was unpredictable and prone to false positives by the safeguards. In my brief testing, it still seems completely unable to understand its own guardrails and will readily reason itself into triggering them. It claims it won't do so beforehand, and insists that the topic in question is perfectly OK. Regardless of how good the car is, I'm not comfortable buying or driving it when I know it can randomly and unpredictably explodes. So yea might be good, but you never know when it refuses to help… still.
> I think Fable 5.1 is a big improvement in writing style
You think or is it better? Or you just YOLOed the model out?
> and responds to my style instructions more reliably.
Yeah, yeah. Previous models wete also advertised as "being reliable". To the poibt @bcherny "released" a new style that was going to reliably make Fable sound better.
> Another point I expect not to get much attention until it all happens at once is science.
You mean "your request to use unicode methids is flagged as unsafe bio research"?
since you work at Anthropic, know that there was (warranted) love for your models from the community as a whole, they performed well and added value
but the verbiage in recent iterations is absolutely insufferable, I will stop using them because of that as soon as I can, I simply cannot stand another round of the model "finding the smoking gun", saying "that's the actual gap, not a fluke" or some idiotic phrasing like this
Must be helpful that your company is slurping and destroying literature, how sad that the results of this are a blog post with “look how well we write English”. Eye roll.
Could not be happier about my decision to turn down a job offer from Anthropic years ago. Ick.
Well your CEO went on X saying you will cure cancer, and since it's always a 6 month rolling window with him I can only assume humanity will be cancer free before next summer, amazing!
Too bad. I see the stereotypical prose as a good thing. When I interact with Claude myself, I don’t mind it as it just feels like Claude’s distinctive voice. But when other people try to disguise LLM output as their own thoughts, the voice makes it easier for me to tell.
People that want to be open about the source of their text will just tell you where it came from.
People that want to obscure the source of their text would rather that it was more difficult to sniff out LLM-generated text. And they're the ones picking which model to use.
Beyond all the benchmarks, I think Fable 5.1 is a big improvement in writing style. It sounds a lot less stereotypically like other Claude models, has (imho) a much more natural style, and responds to my style instructions more reliably. More work to be done (and we will!) but reading better prose makes me so much happier.
Another point I expect not to get much attention until it all happens at once is science. People have been correctly excited about the many "sudden" breakthroughs LLMs are making in Maths, but some of the science benchmarks make me believe we'll soon see similar developments in other scientific domains. Fable 5.1 more than doubled Fable 5's Terminal-Bench-Science [1] score, which I think is meaningful.
[1] https://github.com/harbor-framework/terminal-bench-science