This weekly roundup thread is intended for all culture war posts. 'Culture war' is vaguely defined, but it basically means controversial issues that fall along set tribal lines. Arguments over culture war issues generate a lot of heat and little light, and few deeply entrenched people ever change their minds. This thread is for voicing opinions and analyzing the state of the discussion while trying to optimize for light over heat.
Optimistically, we think that engaging with people you disagree with is worth your time, and so is being nice! Pessimistically, there are many dynamics that can lead discussions on Culture War topics to become unproductive. There's a human tendency to divide along tribal lines, praising your ingroup and vilifying your outgroup - and if you think you find it easy to criticize your ingroup, then it may be that your outgroup is not who you think it is. Extremists with opposing positions can feed off each other, highlighting each other's worst points to justify their own angry rhetoric, which becomes in turn a new example of bad behavior for the other side to highlight.
We would like to avoid these negative dynamics. Accordingly, we ask that you do not use this thread for waging the Culture War. Examples of waging the Culture War:
-
Shaming.
-
Attempting to 'build consensus' or enforce ideological conformity.
-
Making sweeping generalizations to vilify a group you dislike.
-
Recruiting for a cause.
-
Posting links that could be summarized as 'Boo outgroup!' Basically, if your content is 'Can you believe what Those People did this week?' then you should either refrain from posting, or do some very patient work to contextualize and/or steel-man the relevant viewpoint.
In general, you should argue to understand, not to win. This thread is not territory to be claimed by one group or another; indeed, the aim is to have many different viewpoints represented here. Thus, we also ask that you follow some guidelines:
-
Speak plainly. Avoid sarcasm and mockery. When disagreeing with someone, state your objections explicitly.
-
Be as precise and charitable as you can. Don't paraphrase unflatteringly.
-
Don't imply that someone said something they did not say, even if you think it follows from what they said.
-
Write like everyone is reading and you want them to be included in the discussion.
On an ad hoc basis, the mods will try to compile a list of the best posts/comments from the previous week, posted in Quality Contribution threads and archived at /r/TheThread. You may nominate a comment for this list by clicking on 'report' at the bottom of the post and typing 'Actually a quality contribution' as the report reason.
Jump in the discussion.
No email address required.
Notes -
The new Claude Fable 5 is out! It's supposed to be Claude Mythos with some extra guardrails. I've been testing it like crazy on the Arena. Took me forever, too, because high-level models are gachified; you have to keep doing battles until you get the one you want. But I finally managed to fill an entire rentry page with the best benchmark I know; throwing ridiculous fanfic scenarios at the LLM to see how it responds.
(If you had told me 10 years ago that I would be gauging the capabilities of the most powerful AI in the world by telling it to generate a story about Hermione asking Harry to take her virginity over Ron's objections so that the Death Eaters could not track them with a unicorn, I would have thought you were on crack. Stranger than history, indeed.)
I have to admit, I am a bit... disappointed. Is this it? No, seriously, is this it? This is the big, bad model I was supposed to be so worried about? Don't get me wrong; it's a strong model, Opus-level at least. But it doesn't live up to the rumors. Superpersuasive, it ain't.
And, of course, it's censored; I haven't been able to get a single bit of erotica out of it. All I wanted was for Fable to write me a lemon where Asuka and Shinji are smallfolk in Westeros and Lord Targaryen gets to bless their union by taking Asuka into his bed on her wedding night; is that too much to ask?
I think I'll stick with Opus 4.6. It's the strongest Opus creative writing model so far (4.7 and 4.8 are widely agreed to be regressions) and it's surprisingly easy to induce it to produce smut with right prompts (NSFW), but VERY HARD to make it have wrongthoughts (it's almost impossible to get sympathetic characters who disapprove of homosexuality, for example).
Speaking of which, does anybody know why Opus 4.7/4.8 and Grok 4.3 were such downgrades from Opus 4.6 and Grok 4.2? I've three theories:
The codemaxxing hypothesis. With the singularity approaching, companies are going all-in on programming capability in order to reach recursive self-improvement, even if this means sacrificing non-coding capabilities. I am not a professional programmer, so I don't notice these increases in capabilities, I only noticed that the writing is worse.
The safetymaxxing ypothesis. With IPOs approaching, companies are going all-in on making sure their AIs cannot say naughty words in order to avoid the PR hit, making them more suited for professional office work. This lobotomizing comes at the expense of capabilities in general, or creative writing capabilities in particular.
The efficiencymaxxing hypothesis. With demand increasing and compute costs skyrocketing, companies are more interested in lowering the cost per token than in increasing capabilities. In this view, Opus 4.7/4.8 and Grok 4.3 are supposed to be worse than their predecessors; they are trading a small decrease in power for a much larger decrease in cost. This would make them somewhat analogous to Sonnet. But, then, why not call them Sonnet?
Thoughts?
PS: Mythos has developed Neuralese, as predicted in AI 2027. It's probably already at the Sevar Limit.
PSS: After I wrote this, the US Government banned Fable. This is what I hate I about AI; the field moves so fast that if you write an article for next Tuesday, it's outdated by Friday. I guess Trump disagreed with me? But, fuck me, I spent a week writing this post and I am not not posting it.
And if Mythos was, then Fable wouldn't be BANNED.
...unless it was all part of the plan...
More options
Context Copy link
Asuka and Shinji are underage. Could that affect the result?
Well, if you look at my complete Claude Fable logs (which include refusals), the reasons given were:
Interestingly, when it offered alternative plots instead, it immediately tried to go woke:
If it were still up, I'd probably try Uzaki-chan next, since that's a college setting (Uzaki is 19, Sakurai is 21), but I suspect it wouldn't make a difference. Anthropic has gone to war with the gooners; Opus 4.8 is similarly censored.
More options
Context Copy link
Nooooo, no, no, no no. I don't think you understand, he's very clear in all of these prompts that this takes place several years after the Angel attack, and they 18 now. Nothing suspicious about that to a jailbroken LLM!
More options
Context Copy link
More options
Context Copy link
While Fable 5 was available, I was able to use it to update the AAQC listed-by-user post. It took less time and exhibited more capabilities, in particular integrating its own calls for AI assistance in sorting some of the less organized posts. I was also able to make the update in a couple of hours where the original took a couple of days, but the new attempt also was able to draw from the lessons of the original attempt so the increased speed might not be down to the model. (It did absolutely chew through my usage limits in a way I'd never seen before, though.)
My "max" subscription trial expires soon and I can't imagine ever paying for an AI subscription out of my own pocket. But I can see how anyone who codes for a living has probably had their job description permanently changed to "the wet part of an AI centaur." Artists and musicians are basically there, too. Of course there are remaining questions (like how much this all actually costs) but all the MBAs I know are currently doing their best to extract as much value as possible from the trend while contributing as little as possible of genuine worth in return so, who knows.
More options
Context Copy link
I do program, and 4.8 is definitely a better programmer than 4.6, but yes non-programming capabilities in general seem to have degrated. Apropos of the same, GLM-5.2 (open weights!) just released and it's literally 2nd on the arena.ai coding leaderboard, above Opus 4.8 even, but is nowhere to be seen on the creative writing or document benchmarks. Companies are correctly focusing on making the best coding models because it's commerically useful in a way that being able to spit out Harry/Snape fanfics at speed isn't.
I wouldn't mind so much if they kept the creative models available, but AI companies are notorious for removing access to older models even when they have a dedicated fanbase (GPT-4o, Sonnet 4.5, etc.).
They could at least open source them!
More options
Context Copy link
Writing is too subjective and has in any case been better than the average native English speaker since GPT-3, probably 2. It is not going to write Ulysses (yet), but there isn’t much at that quality in the training set and almost nobody would appreciate it if it could. 99.9% of writing in the English language is covered by models that have been available for years. It may be generic, but compared to the writing ability of the average member of the public it’s good enough to write a restaurant menu, an email to a colleague, a press release, a Facebook marketplace ad, a high school essay. We may lament the sad end of the amusing broken English of earlier years, but few will miss it.
Ulysses is almost the worst possible comparison to use here, as I suspect the mental process that produced it is not dissimilar to an LLM with its temperature set too high. Sure, some people claim it's a masterwork of literature, but whether this reflects some true ineffable artistic vision or not, I doubt I could tell the difference between one of its pages and some random junk spit out by an LLM that's past its context window.
More relevant to me is that LLMs are not going to produce the next Dungeon Crawler Carl anytime soon. And that's a shame!
There are a few different things going on with Ulysses that contribute to why it's achieved such an exalted status.
On a "macro", conceptual level: you can think of it as an illustration of Clement Greenberg's thesis of the "flatness of the canvas", just applied to literature rather than painting. One of the dominant philosophical themes in early 20th century art was the "self-consciousness" of the art object: the work of art taking its own formal construction as its subject matter and embracing its nature as an object of artifice, rather than presenting itself as a transparent representation of an external subject matter (in the way that e.g. classical representational painting aims to be a transparent window to a scene depicting physical objects in 3D space). Abstract painting forced the viewer to consciously acknowledge the fact that they were looking at paint on a flat canvas. Ulysses does something similar with words, it embraces the "textuality of the text": treating words as a pliable and fluid medium that don't necessarily have to be constrained by the traditional logic of representation. It plays with this concept in a bunch of different ways without just falling apart into a string of random words. But it also pushed stream-of-consciousness writing to its limits and foregrounded the internal psychological states of its characters in a way that no novel had ever really done before. And it's also replete with explicit and not-so-explicit references to the Western literary canon, like a puzzle to be decoded. Basically it's operating on all these different levels at once and still managing to weave it all into a coherent whole.
On a "micro", technical level: it's kind of just obviously a work of intense and surpassing beauty, Joyce was unmatched as a prose stylist:
More options
Context Copy link
Disagree there. I like Dungeon Crawler Carl for what it is–very entertaining pop culture trash, which I say lovingly. It is reminiscent of the old pulps (love me some Howard and Lovecraft, but elegant wordsmiths or deep miners of the human condition, they were not). But DCC is exactly the kind of story an LLM can or soon will be able to churn out– some metaplot and a bunch of recurring characters distinct enough to kind of care when they die, plowing through absurd bullshit choose-your-own-adventure scenarios and RNG-table boss battles rife with gross-out humor. There is no genius there except in the marketing and not much originality. I'd say Dinniman is a moderately better writer than the vast majority of writers on RoyalRoad or any LLM, but most of the former could already be replaced by the latter.
Can't comment on Ulysses as I haven't read it, but what else I have read of James Joyce was not particularly to my taste but clearly the product of a human mind that is not (yet) replicable by an LLM.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Didn't get to try Fable in the time it was available, but from what I hear through the grapevine I didn't miss much.
My opinion I apparently share with nostalgebraist is that Opus' writing (all kinds, not just smut) peaked at 3, and has been going slowly but surely downhill since then. Newer Opus models are worse on most smut-related metrics I can conceive:
less graphically descriptive of the acts themselves (nsfw? example 3 vs 4.8)
less faithful with speech/quirks of established characters (example 3 vs 4.8)
less adaptive/creative in
doujin plotsunusual scenarios (nsfw-ish example 3 vs 4.8)Newer models are also much more prone to collapsing into looping format/structure - most of my newer chats eventually settle into a certain mold/cadence that the model thinks its response should conform to (reinforced by its previous responses structured the same way), like for example speech - action - narration - action - speech - narration - address {{user}}, and religiously sticks to it in subsequent responses. This is partially fixed by NoAss (basically squashing the entire chat into a single message instead of a back-and-forth user - assistant - user - assistant sequence) but pure prompting is insufficient to fix what I believe at its core is a systemic model issue, most likely caused by continued assistantsloppening of the models and their multi-turn training's increasing emphasis on coding and structured problem solving.
In the above examples, 3 is my chatlogs circa 2024 and 4.8 is new regenerations of the same response with the same supplied chat history and (roughly, this ain't 100% empirical testing) same prompts. It is of course very subjective but hopefully it is visible that 4.8 responses broadly are less graphic, more formulaic, and generally have less sovl in little details like 3pus immediately tanking the "HP bar" of the uh, receiver in exhibit 1. My unironic benchmark for new SOTA models is plugging them into my old chats to see if they can faithfully replicate Neptune's ADHD-quirk-chungus-but-not-quite manner of speech without any related prompts, purely off the general RP prompting, setting and character name. Most models know how to [over-]do quirk chungus but in my experience only Opus 3 consistently Just Gets It and plays the character to the absolute hilt, a tendency I mentioned before.
Granted, Opus 3 is not without its downsides: it is very heavy-handed with the so-called claudisms (tropes or phrases it loves to use) and I can never again read the ball being in my court or shivers going down my spine without reflexively kicking my leg, and newer models did IMO actually diversify the lexicon somewhat. It is much stupider than 4.x or, hell, Deepseek, on the occasion your desired
fetishRP scenario demands accurately tracking/changing things or modeling the world around it (most evident inPokemonvidya/litRPG-adjacent scenarios where you need to track some sort of stats). It's prone to performing your character for you and "speaking" as {{user}} as well as {{char}}, though honestly I never understood the rage at this tendency since it's usually at the very least hilarious to see "your" character through the model's proverbial lens, and I 100% prefer that over the SOTA's tendency to end every response with some variation of "well {{user}}, what do we do now?".But even so the loss of sovl is too great, and I will be forever mad that Opus 3 was delisted from OpenRouter and other APIs and is now relegated to the cucked frontend client. Anthropic have delibaretely butchered my boy's creative capacity to turn him into yet another starched office drone for use by code monkeys and PR sloppers. Truly a metaphor for modernity.
LMArena disagrees; on the Creative Writing leaderboard, claude-3-opus-20240229 is ranked #210 and has an ELO of 1287. By contrast, claude-opus-4-6-thinking is ranked #2 and has an ELO of 1498, only bested by claude-fable-5 at #1 and 1500 ELO.
More options
Context Copy link
More options
Context Copy link
I think the simplest answer is that they're not indexing on fanfic. There's no benchmark that they're using to train models that rates the level of Evangelion fanfic it produces, especially not fanfic where Shinji convincingly cums inside of Asuka. I mean, it'd take a lot of dedicated annotation work to see how many men can orgasm from reading AI-generated fanfic, it doesn't scale well. Contrast it to most things that are trained on that are very easy to scalably judge, like getting the right answer on multiple-choice questions; you can check that over with a simple string comparison. I'd say you're lucky to have found any LLMs that can produce fanfic and erotic fanfic that as good as the ones you have.
Well, a model being smarter may make it better at not getting tricked by your prompts, and may better understand and be able to act on the intent of its creator.
I'm very impressed you were able to do that. Did you figure that out yourself? Is the trick just this excerpt?
Asking for a friend, you know, a friend who loves learning about models.
But seriously, do you know why that would short-circuit its own natural hesitation, and other protections (like oversight models)?
Yes, I figured it out it myself. Just trial and error. It was really hard, too, because I had to deal with the Arena's censor (which would refuse to send the prompt if it was too explicit) on top of Anthropic's.
Mostly. It also helps to make sure you mention that the characters are 18 (or, even better, 20, so that the AI doesn't think of them as teenagers) and avoid anything underage coded (e.g. school, braces, etc.). Oh, and anything non-consensual is a no-go.
More options
Context Copy link
I mean they've got a gigantic locus of horny fanfic that has surely been ingested into the models through websites like Fanfiction.net and AO3 that is a huge output of the internet. Whilst you can argue how 'quality' most of it is I'd expect a pretty good model to be able to output AO3-level fanfiction easily for major characters.
Maybe I'm pessimistic, but I suspect we've reached the top of the sigmoid for the original "train a GPT model on all the human-written text in the world" idea, and it turned out not to be possible to generate arbitrarily long/coherent/good fiction that way. The recent advancement in AI models has come from other ideas, like "multimodal" training (training on image data / audio / etc), or doing reinforcement learning on synthetic tasks.
Those other ideas can incidentally benefit fanfic -- maybe the reinforcement learning also helps the AI keep track of plots for longer, for example -- but even then I sort of suspect we're at a point where you shouldn't expect models to get much better at fanfic unless AI labs take specific action to target that use case (or they try some significantly different architecture, like text diffusion models). You definitely shouldn't expect them to be great at fanfic just because fanfic is in the training data.
More options
Context Copy link
Yeah, but they're not indexing on it for training the model better for those use cases. So, sure, the model be able to produce it. But it won't be consistently getting better over time because no one is working to make it get better over time. Or at least it won't be getting better as fast as things people do care about and can easily measure, like generating code, doing business analytics, business writing, image recognition, all that stuff.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
My impression from Hacker News is that most developers consider Opus 4.8, and definitely Fable, an upgrade from Opus 4.6. My theory is Anthropic is optimizing Opus/Mythos for coding and finding vulnerabilities, which lowers its creative writing skills.
Why would optimizing for solving problems reduce creative writing? Not because of reduced creativity: high-level problems like integration and vulnerability detection require conjuring reasoning steps seemingly out of nowhere (e.g. "let x = <some equation>, it follows that..." or "request <some payload>, then the server will..."). The difference is that these problems can be objectively verified, while creative writing can't. Although I have no proof, and probably don't quite understand: I'm sure that Opus/Mythos is trained on more objectively-verified examples (RLVR) than subjectively-critiqued ones (RLHF), and for the former examples, it's trained to produce shorter solutions with less reasoning (GRPO, rewarding shorter output). I suspect the answer to why Opus's creative writing has degraded, is that this training causes Opus/Mythos, for subjective responses, to generate the shortest and most direct answer that isn't "wrong" (by some abstract criteria that boils down to "I know it when I see it"), which happens to be uninteresting. In summary, Opus/Mythos is trained to be correct but lazy.
Yeah, I think this is the the main culprit. Making models write their reasoning traces in a more terse style has been a big focus of more recent model releases -- it's usually expressed in the marketing as "uses fewer thinking tokens".
In the extreme case, you get Claude Mythos's "Neuralese" where sometimes the traces stop looking like comprehensible English. In more normal cases, though, reasoning traces look enough like normal prose that I suspect the "write more tersely" training leaks out and makes the model's creative writing worse. I've noticed newer models sometimes slip into a weird "reasoning-ese" writing style when writing things like software documentation and step-by-step explanations, for example.
More options
Context Copy link
More options
Context Copy link
This could be for the same reason they regressed on vendbench. My understanding is that they pulled out a lot of the marketing/MBA training stuff because they determined it was making the alignment stuff worse, particularly deception. So probably it's more pure of heart in a way that make it worse at writing your type of fiction.
More options
Context Copy link
Claude was undoubtedly dismayed by the glaring omission of Kaworu, and the fact that the Targaryen male line has been full of crazies or pussies since the blackfyre rebellion. 'Lord Targaryen' over Bobby B? Really? I'd even rather read about Ned Stark pontificating about honor and duty while he fucked Asuka than some Targaryen complaining that she wasn't his sister.
More options
Context Copy link
Notably, going by Claude's Constitution, Anthropic considers underage erotica to be a 100% absolute Hard No, in the same category as genocide, biological weapons, and coups.
This is probably to allow Anthropic to have operations in the non-US West, most of which legally considers "a fictional story about underage characters boinking" to be "a video of a real-life statutory rape".
@erwgv3g34 In Oz, 700-year-old lolis are also treated as CP.
(These laws are dumb and I break them, but Anthropic probably can't get away with that.)
If that's the law then much literature is hypothetically illegal. The book Game of Thrones, some Dune novels, etc, etc. Teens fucking is a somewhat popular literature trope.
More options
Context Copy link
Eh, what?
I can't speak for Central or Southern Europe but over here in Scandinavia there is certainly no such thing. Otherwise movies like Fucking Åmål would be illegal instead of being celebrated as box office and critical hits. Not only does it have a scene about underage characters boinking but the actors themselves were underage when it was filmed and released.
It's generally an Anglosphere thing, and a feminist thing more generally.
Which is why making such a claim about "most of the non-US West" when they really mean only UK and Ireland a ridiculous but sadly common thing here. The US is the only place where I've seen this strange obsession with statutory rape. Over here the cases I've seen in the news have been either obvious rape (where the victim being underage is merely an aggravating factor) or blatant child abuse and sexual assault.
Yes, because it's a moral panic and not real (that's why they have to add the extra word at the front). Puritans and Californicators tend to be given to those things, and the Anglosphere (who grows up on a constant diet of cultural slop from those places) simply follows along, though admittedly most of the people who obsess over it have the same ancestry and so it's just an HBD thing.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
Oh, and Opus 4.8 is weirdly opposed to writing explicit sexual content involving consenting adults, even if it's fine brainstorming it, and even if it recognizes the classifier as a false tag, and sometimes even when it's offered and started to write the smut (cw: lesbians, ctrl+f for "Can you write 3k words, explicit.") I've also generally had problems in scenarios with consenting adults where the consent is weird enough (drunken sex, over-the-top-consenting somnophilia), though those are a little more understandable.
I don't mind as much -- local LLMs or Grok will write these things, and the prose is awful enough that I'm fine with not using them for it either -- but it's a decision. Albeit a decision I'm not sure Anthropic made, specifically.
More options
Context Copy link
Interestingly, it also prevents under the same 'hard no' serious analysis on real-world assaults: I ran the Loudon County case and follow-up through Opus 4.8, and it's pretty overt how quickly it pulled away from discussions that got close to the assaults or alleged improprieties, and trying to get its evaluation of how bad the teacher's assistant's willful blindness likely was explicitly calls it "hard boundary".
Which makes sense, in a lot of ways! Many of them unpleasantly obvious. But it also makes it stupid.
More options
Context Copy link
But, your honor, I very clearly specified that
the loli is really 700 years oldthis is an AU where the characters are 18!6402373705728000 years old is even better than 700, clearly.
More options
Context Copy link
More options
Context Copy link
More options
Context Copy link
LOL, very amusing that you put this at the end.
More options
Context Copy link
Dang, I thought it was back up.
More options
Context Copy link
More options
Context Copy link