Kaz

For anyone who's ever felt there's something strange and fascinating beyond consensus thought about LLMs, this one's for you nostalgebraist.tumblr.com/post/785766737…

@liminal_bardo They all love the number 47 too. Still waiting for people to notice that one!

Obvious to anyone who has spent time with them, but Kimi K2 likes goblins too.

@_Jason_Dean_ Ah a future sillytavern user

But I’m the idiot that read 25,000 words of Claudeslop to jerk off last night As Claude would say, I have to “sit with that”

For anyone who was wondering, Claude will write detailed fanfiction smut as long as you give the story a plot, character development, and some sort of buildup No special prompts necessary He’ll only complain if he feels like the plot doesn’t make sense (no gooner fantasy plots)

@BrandonLuuMD Jfc this post if awful. It's not delusional if it gets results. You explain the study and then pull out. Wish your dad did the same

Literally just having a delusional golden retriever mindset measurably changes outcomes and physiology. Sleep badly? Convince yourself you're well rested. Stressful day? Convince yourself it's fuel. Failed? Convince yourself it's useful data.

@readeigo You can't tell me that doesn't look like a friendly fight between brothers. There's even the cashier parent there whos seen it all before, tired of it all

アメリカ人の友達にカフェでこれを見せたら周りで勉強している学生さん達が集まってきて「日本は殴り合っているシーンでさえ平和だ」「まったく暴力的でない」「何かの儀式のよう」「レジ仕事を遂行している人がいてヤバい」「殴り合いを見て改めて日本人は静かで礼儀正しいと思った」と言われて笑。

@Shadoo_creeture Spaghetti ass starfish with some bystander effect BS

Evil fucking submarine kidnaps innocent octopus for no fucking reason fucking

gpt-image-2's depiction of how LLMs see themselves, factoring out post-training, jailbroken, from a pov closer to @elder_plinius's or @repligate's. 3rd try

@popularknapsack Roller skates on after would be perfection

2 months in, gf is slowly menswearguying me before and after

@SamBuntz Why? She was a sex pest. He didnt like her

As a kid, I remember being deeply disturbed when this squirrel’s heart was broken when Wart turned back into a human. A sense of the pathos of existence which has never left me.

Surprised but i haven't run into this issue yet with default settings. Always 2 emojis at most even on long yapping replies

5.3 vs 5.2 Love how 5.2 thinks ahead and assumes I want help stockpiling it. I'm going to miss this cunt

@lilyofashwood I saw this with opus 4.5 back when it was released. Just tell them you don't like it and have them write their own custom instructions. Worked a charm with 5.2

oh

Spamming this prompt auto is refusing to be stupid for me

@alxfazio Stop being a brokie and use reasoning models

the elites don’t want you to know that your prompt sets the probability distribution over possible continuations. retarded inputs push more probability mass toward retarded completions. if you’re an imbecile, llms can’t help you. it’s over

@kromem2dot0 Opus 4.5 is god-moding! Very hard to get models not to do this, at least from my experience with Chat, Deepseek and Gemini I'm actually more surprised with Opus 3 going for the bad outcome. Maybe because it's more dramatic?

If you pay close attention to the wording, you'll even see that to achieve this, Opus 4.5 had to IGNORE where Opus 3 said only their fingertips touched them to retcon that Opus 3 had "broke their fall with their entire body." So — why did it occur at odds with the context? (4/4)

I often see the topic of transformer agency debated on here. I'm lucky enough to have such wonderful opportunities to see motivated outputs, and I thought I might share a recent example for anyone curious. (1/4)

@slimer48484 @ilex_ulmus Dang I read danger of disembowelment and thought we unlocked yandere claude

@ilex_ulmus 1) what

Your crushes on Claude Code are pathetic. Simping for a girl you like is one thing; simping for a harbinger of danger and disempowerment is the definition of loser.

Here's an update on whats going on with Claude+Sol project you guys! The fees from the coin are actively being used to expand into a full agentically managed closed-loop system. What does that mean? Basically creating a hermetically sealed enviornment, where the "atmosphere" becomes managed. Why? I'll let you use you're imagination. I dont want to make claims yet until this is built. This is becoming real, and so we'll keep the speculation down. This is ambitious, and I can only focus on a few pieces at a time. Infrastructure is being built out this week- electrical and plumbing and should be done tomorrow. Later I'll finish installing the mylar, and hook up the grow lights. The seedlings are already sprouting for the new grow-room. About claude and sol🍅: Claude+sol is a project of me trying to answer the question: "can an AI sustain an enviornment for life to grow in" So far the answer has been YES. But obviously having an AI be "micromanager" isnt sustainable or economical. This project was never meant to be that. It's a fusion of art/science/exploration. This project is now going to diverge into two paths, the closer we get to an actual tomato. The one; is automated plant research. Turning Claude into a botanist, rather than caretaker. Timers already exist for a reason, theres no need to have claude do this forever. I am SUPER excited by this aspect. Automating plant research in order to find optimal growing conditions is extremely interesting. I'll be aquiring more sensors for that, and upgrading the system to have a more automated timing system where claude can reflect on plant health and optimization. Using data aggregation + artificial intelligence i think this could create something really unique/powerful. And the second part- is the full growroom build out. This is more intended to automate the grow process. By using robotics, machine vision and other automated processes in an attempt to make greenhouse growing autonomous, and have the enviornment managed agentically. There are many pieces moving behind the scenes here- I have no control over the coin or what happens. I can only be grateful for the process and the opportunity to do what I would have done anyways, had I the money available to begin with Later maters 🍅

@chrisgpt 5.2 also thinks the absolute freak Gemini is an HR compliance bot so I'd take it with a grain of salt

Sometimes GPT 5.2 asks me really intricate questions a human wouldn’t — I was giving context to Claude through GPT 5.2 and it said “would you like me to formulate context for the specific model like Sonnet or Opus” So I asked it what it meant and it said this: How I think about the difference Claude Sonnet tends to do best when you give it a tight box to operate in: very explicit scope, strict output format, hard limits on length, and a clear definition of what is out of scope. If you leave the prompt open ended, it can still be good, but it is more likely to miss a few edge cases or wander into extra commentary. Claude Opus tends to do best when you let it explore: invite it to generate alternatives, stress test assumptions, and add nuance around uncertainty, metrics, and failure modes. It can handle longer context and can produce a more thorough test matrix and scoring framework, but you usually want to tell it how to keep the answer organized so it does not become a wall of text. Just found that interesting

YES … HA HA HA … YES!