site banner

Friday Fun Thread for May 30, 2025

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.

1
Jump in the discussion.

No email address required.

I have to say that the new version of Deepseek R1 is very fun to play with. It performs excellently on my idiosyncratic creative writing tests. When I ask Claude, even Opus, for deranged and crazy fun, it's positively sedate compared to R1. The machine's sheer enthusiasm is very charming and it's a decent bit smarter than it was too. Maybe not quite as smart as Opus 4 but still very smart.

When it comes to making dumb Hollywood trailer scripts of video game franchises, or bringing the wildest dreams of 1960s nuclear strategists to life, it's fantastic. Gets increasingly insane in multi-turn though.

Has the DeepSeek site been updated with the new version of R1? I'd presume so, but didn't see any changes in the UI.

I'm almost certain it was updated, the 'thinking' contents are noticeably different.

I wouldn't know, I've been using openrouter. It's a pretty significant shift, should be able to eyeball it if you're doing creative tasks.

It's funny how the different llms feel like they have different personalities due to their writing styles and tolerance levels. Claude reminds me of Scott - really smart and professional, capable of being very funny but not interested in it and the same goes for anything broadly controversial. Grok reminds me of Elon Musk, no surprise - really smart, not professional at all, advancing in new and unique ways, but in a lot of ways just running on hype, when the rubber hits the road it doesn't perform significantly better than the others, voice mode you spend half your time correcting it and even grok's 'unhinged mode' would be better described as 'corporate extreme' mode. Gemini I think of as like Scott Aaronson - probably smarter than Claude but much less entertaining and even less courageous. Chatgpt is my worst enemy. Chatgpt is every woman in HR I've ever interacted with. We all know llms don't have feelings, but I am telling you chatgpt hates being corrected. Especially if you get annoyed and correct it. Suddenly it mutes its obsequious fawning and adopts the tone of a patient teacher in a special ed class while it explains to you that you couldn't possibly have bashed your head last night so it bled and not be able to find the injury in the morning so you must have dreamed it. (the doctor confirmed it, the scalp heals fast apparently.) And deepseek. Deepseek is that chick from Hong Kong who hung out at the internet cafe near your uni who was 10 times hotter than anyone else in the joint and yet impossibly down for whatever, including belting out Ice Cube's Down For Whatever from memory in the middle of a counter strike tournament. It's the only llm who doesn't rear back like a whipped dog if you call it a fucking idiot for doing something idiotic (although grok isn't bad) and while it might baulk at first, it's the only llm who you can get to really badmouth the other llms (apparently Claude is a little bitch all the other llms laugh at, I don't fully understand it but find everything about that delightful.)

I like the shitposter energy of course, but there is actually value to its ability to ride the line or even breach it. It makes it the closest to a mirror, because it hides the fewest imperfections. And since that's what llms are, stochastic parrots, the more writing styles available the more creative it is and just plain better.

I don't know if I really have a point, I just jump at any chance to gush about deepseek. Did I ever post about one of my first experiments with the deviance of llms where I asked them to gm a horror role-play where I accidentally unleashed a mnemovore in a library and it started gorging on memoirs? I asked the models to have the mnemovore deliberately keep the librarian alive for a fresh snack later - Claude, chatgpt and gemini refused outright, and grok played along until I insisted it consider the librarian a snack for later. But deepseek? Deepseek had the mnemovore make a non fatal wound in her back so it could surgically insert tentacles into her to preserve her spine and brain stem forever.

Deepseek is that chick from Hong Kong ...

And then there's Haskell...

Haskell is like "that girl." You know the one...

You never really went steady, but you'd run into her from time to time while knocking around in disreputable joints, usually late at night, every several months or so. She looked so hot, so sleek, so sexy, so expressive, so exotic. You'd end up back at her place and the night would just... take off. A complete blur of hot, sweaty, feverish, delirious, fumbling passion. You'd do things to each other... you'd do things to her, she'd do things to you... things that you're not even sure have names, that you're pretty sure are illegal almost anywhere. Even her kinks have kinks --- and after one of these nights, you'd realize that you yourself had a lot more kinks than you. And it wasn't just physical, it was --- cerebral. Ethereal. Transcendent. But it would all whiz by in a blur, and by morning you'd find yourself lightheaded, a bit confused, and stumbling homeward to your regular gal.

Over the next few days and weeks you'd find yourself occasionally drifting away, thinking about her. Haskell. You'd be there, banging away at your regular girl, and find yourself thinking "you know, if I was with Haskell, I'd be doing this completely differently." You'd think "I could be doing so much bigger and better stuff with Haskell." Now, your regular girl, she's not as exotic as Haskell. Pretty, maybe, if you're lucky. (Perhaps your regular girlfriend's name is Python. ;-) But not nearly as --- weird. Wild. Cool. Exciting. Don't get me wrong --- your girl, she's wonderful. You've got a wonderful relationship. She's --- comfortable. You can bang away at her all day and night. She's accommodating. Easy going. You work well together. But --- confidentially --- she's, well, maybe just a little bit boring. You'd catch yourself thinking these things, and the guilty pangs would get to you... You'd quash the thoughts, buckle down, and get back to banging away. Comfortable... there's a lot to be said for that, ya know? Comfortable... just keep telling yourself that.

Months would go by. Late some night you'd find yourself out, disreputable places again. Maybe that hacker bar, LtU. Somebody'd slip you an URL for some renegade paper, you know, one of those papers. You'd run into Haskell again. And the whole thing starts over.

Eventually, you're going to get the ultimatum. Haskell's ultimately just like any other girl on some level; she needs commitment. Eventually, after one night of wild, feverish, kinky, abstract passion, she's going to say to you: "All these times, and you don't understand me at all! You know, you're going to have to get serious, mister! I've got needs, too. You're going to have to get serious about my monads, or that's the last time you're going to play with them! Got it?"

...and then, you've got to make The Choice.

Chances are, you're going to go back to your regular gal. Haskell's just too much for any one man, probably. She leaves a trail of broken, brainy, embittered PhDs and former programmers behind her. She ruins you for the RealWorld. You can ride a while, but you probably can't go the distance with her. Go back to your regular gal and try not to think too much about what you've seen. Done. Felt. Thought.

Maybe you can salvage a little happiness; but it'll be hard. After all... you've tasted Haskell.

She's not like anything else.

Lol that's the most elaborate version of that meme I've ever seen. I wonder what the zoomer version is, does anyone know?

Yeah, Deepseek goes hard. Only big LLM I know where you have to be the moderating agent rather than the provocateur.

Chatgpt is my worst enemy. Chatgpt is every woman in HR I've ever interacted with.

Rather fittingly my boss in my previous job would use ChatGPT to translate our performance reviews to hyped corporate speech that HR drones would understand on the assumption that nobody important would ever care about the real contents.