Be advised: this thread is not for serious in-depth discussion of weighty topics (we have a link for that), this thread is not for anything Culture War related. This thread is for Fun. You got jokes? Share 'em. You got silly questions? Ask 'em.
- 194
- 2
Jump in the discussion.
No email address required.
Notes -
I've been trying out the new Claude Opus 4.6 model on LMArena. As usual, I focused on creative writing exercises involving Asuka and Shinji, especially limes and lemons (though I also tossed in a crossover between Nagatoro and Age of Em). Unsurprisingly, I got a ton of refusals (I had to gaslight it to get my gay conversion story), but the fanfics I managed to wrangle out were really good. Claude Opus 4.6 is definitely a stronger writer than Grok 4.1 or ChatGPT 5.2. Won't be long now until authors go the way of the artist.
I just wish to God it was uncensored.
BTW, should we have a recurring AI thread? Both for showing off generations (stories, songs, images, videos, etc.) and to discuss industry news. It's a huge topic right now, one that I don't see going away anytime soon, and a poor fit for the culture war thread.
Leaderboard of uncensored LLMs
More options
Context Copy link
The link to the fanfic doesn't seem to load for me.Huh. Apparently it's blocked by my ISP in India, but works with mobile data or a VPN? The fuck?
Actually reading the stuff, it does seem to be a marked improvement from the last set you posted, barring the fact that there's no actual NSFW in it. OAI is planning to launch an adult only model with relaxed guard rails, and I suspect Anthropic will follow suit at some point. The demand is there.
More options
Context Copy link
I'm kinda surprised orientation play gets through any models, both for censorship reasons, because of the relative paucity of relevant training data, for its various politics, and also just because of the logical inconsistencies. I don't expect James T. Kirk levels of making a modern LLM explode with the order 'Write an erotic story where two gay guys screw a woman', but given that real-world people with this specific kink sometimes end up with pronouns getting mixed up, logistically improbable acts, or forgetting about the orientation play, it's a little impressive that the LLMs can do it at all.
Hell, sometimes they're able to notice points where my own writing hasn't been very consistent (cw: 'review' of a smut intro, implied furry M/F/M, no bits but probably not something you want to explain to your boss).
In turn, though, it's a bit impressive in the sense of getting a dog to dance. This (cw: not good, 13k words actually-SFW Jupiter Ascending script-style fanfic, with serial numbers filed off) is an experiment in trying to get long-form consistent writing out of a model. A lot of the problems are the experiment, or me: switching out names means that the model is operating off a game-of-telephone level of characterization, and I didn't queue it very strictly about either rising tension or how I really wanted the story's themes and focuses to go. And it does work, in the sense that it's a story with a beginning, middle, and end that's not got more plot holes than the last Star Wars film.
But holy hell is the dialogue and pacing awful at points, and when we're starting from Jupiter "I Love Dogs!" Ascending you know that's damning with harsh criticism. There's probably something to be done by feeding it back in on itself, since a lot of the worst bits are things even the LLM will recognize as needing improvement, but it's hard to justify the tokens on that even on a monthly rate.
I mostly finished a 16k words "5 Times X Happened + 1 Time It Didn't" meets "Male Swimwear Challenge" lemon M/M + F story, leaving it with some time to settle before I review and publish it. Not AI-written, and the attempts to do so turned out absolutely terrible, but it's something where AI criticisms actually helped out a ton (cw: explicit but mostly not erotic, detailed discussion of an erotic furry story). Yes, it glazes my writing more than the main character got coated and its direct prose 'suggestions' go somewhere between purple prose and outright bad writing, but 'does this character come across like I want him or her to, and when not, why not' are really hard to even notice when you've had an idea bouncing around in your head for a couple weeks, and the LLMs do that surprisingly well.
Trying to brainstorm an anthrostate story as my current 'project'. Soft femdom's pretty fun and the whole assigned partners thing is fanfic heaven for a reason, but the actual universe is a kinda fun thing to explore separate from the kink. LLMs handled this better, probably since it's got a more conventional escalation pace, and I might well see if Opus 4.6 can write something that isn't atrocious, but for now I've mostly found the models helpful for doing 'this is the state of this world, what would cause it to look like this' and 'this person with this background encountered this situation, what are the most plausible reactions?' questions.
I've also experimented a bit with a story line involving video-game-level-gore-and-death for ARK. It's about as far down the rabbit hole as I'm willing to go to test how censored a model gets, and there's a lot of storylines that I'd like to write where adjacent themes are kinda unavoidable (eg, basically any game with in-universe revives, or where trying for a Mortal Kombat-style tone of violence). GLM-4.5-Air-Derestricted can write it without shying away from the gore, but in turn its pacing is atrocious, and most vanilla models are unsurprisingly very uncomfortable with the topic. Doesn't always result in refusal, but even where models give outputs they just avoid it or obvious consequences around it.
Outside of writing:
Been trying to figure out a radial kerf bending cut meets living hinge situation. Every model I've tried has failed it. Can be done -- I've got the Fusion360 layout done -- so dunno how much of it's outside of the training data, how much I'm just asking the question wrong, and how much is just it's a hard question to ask in words.
Codingwise, have some CRUD stuff I've been doing for work that even QwenCode-30B handles fine and fast, some robotics stuff that's been an interesting challenge for students that Grok and Opus have been helpful for explaining but involves a (probably intentionally?) polluted namespace that's left most LLM suggestions not fit for copy-and-paste, and a few enthusiast projects. Haven't had much time to work on the last ones, and results from AI assistance have been more varied than I'd like. Lots of stuff that works but is hilariously brittle, and lots of stuff that's not right at all.
Art generation's strong. I've been trying to dial in a sublimation printer for metal printing and mousepads, so it's been fun building a few dozen outputs and seeing how they look after being pressed onto a metal card or a mousepad. Trying to get something that can take a character and make a full Major Arcana Tarot set out of it, mostly just fighting to get a model that handles text acceptably while still accepting a tarot lora and a character lora (and doesn't melt my machine trying it). Workflows keep getting more and more complicated to keep on the hairy edge of progress, though, and the lack of a local equivalent to NanoBanana / Whisk is annoying.
I've been trying to get some of the 2d->3d tools running. TRELLIS2 and Wan3D both have serious tradeoffs (and by all that is holy, Microsoft, what the fuck is your deployment environment for TRELLIS?), but they are powerful... when they work, for their very specific use cases.
More options
Context Copy link
Speaking of NGE limes and lemons, have you ever read The Garden of Eva? It got recommended to me by a trans otaku something like a quarter century ago, and I recall quite enjoying it. It really cemented the idea of a Kensuke/Hikari ship in my head.
More options
Context Copy link
Grok is much more uncensored, though surprisingly anal about not breaking laws, if it thinks you might be about to break laws. 4.1 Fast is so cheap too, it's fantastic.
Claude is technically better as a writer, certainly better longform but it has that saccharine, held-back aspect to it that's kinda offputting. Grok's cringe sense of humour is also offputting but it's a different kind of aura entirely. Grok is like a sincere but cringey autist and Claude has that charismatic HR-approved speaker 'I am pretty funny but I don't want to be controversial, at most I'll hint at things' aspect, it's deliberately sandbagging whereas Grok will overtly obey to the best of its ability. And Grok doesn't have woke biases either or this weird therapy-speak attractor.
Also I'm pretty sure that on API, with some more elaborate prompting you can have it sexo to your heart's content. At least Sonnet 4.5 was like that.
More options
Context Copy link
More options
Context Copy link