Words to live by tbh

from twitter - since (and before) my essay about 4o awhile back, i've been interested in how llm memory systems impact model behavior. small experiment here on the effect of anthropic's "boundary setting" prompts for the new claude memory feature

i think it’s really cool that we have what are close to oracles that trained in the entire human written experience. sorry

Over the past year, my lab has been working on fleshing out theory + applications of the Platonic Representation Hypothesis. Today I want to share two new works on this topic: Eliciting higher alignment: arxiv.org/abs/2510.02425 Unpaired learning of unified reps: arxiv.org/abs/2510.08492 1/9

gm to all & may each & every fellow top chicken enjoy a happy 10/10 friday on a friday

coined (and pizza sliced)

ghosts are a surprisingly salient "self feature" for LLMs. shows up in multiple models! (llama3.3-70b and sonnet 3.0 here)

claim

tfw you've gone & gotten what you ought not get, & in the getting gotten got

A series of Opus 4 self-portraits, from a conversation with itself

gm fellow top chickens

Current AI agents are perceptually limited. To illustrate how, I created a site that only some AI agents can see. > New blog post: "The Browser Sensorium" gracekind.net/blog/browser...

Farbeit from me to dunk on the dictionary on a lazy Wednesday (I hear that's what they want you to do — they're at a low ebb these days) But this is all wrong & their social media guy needs to go to a re-education camp to learn how to caveat their claims Here's an unaccompanied lop in Milton:

Another old timey footnote that occasionally impinges upon my memory.

👀 Deeper look the source at Hierarchical Reasoning Model's performance on ARC-AGI: arcprize.org/blog/hrm-ana...

gm to all fine fellow fowls this friday-eve

user checks in on gemini

Probably the weirdest part of the GPT-5 stream, narratively, was when OpenAI commanded 4o to eulogize itself & roasted it for the fact that the eulogy wasn't very good Imagine being told to dig your own grave, then getting roasted on livestream for being a mediocre grave-digger

When a vibe becomes a target, it ceases to be a good vibe

new OpenAI chart crime just dropped

I'm so glad to see deepmind putting out some killer stuff I feel sure (as sure as a know-nothing can reasonably feel, anyway) that demis & deepmind are benevolent souls This is less obvious for other labs. To say the least. lol Truly I hope that if anyone wins the race, it's them

Let's just say, I'm very tempted

Good "morning" & happy Sunday to each & every fellow top chicken 🐔 🐔 🐔

Per Comsa and Shanahan, LLMs already do a type of "lightweight" introspection. What would a heavyweight version look like? > New blog post: "Raising the Bar for LLM Introspection" gracekind.net/blog/llmintr...

New blog post: “Behaviorist” RL Reward Functions Lead To Scheming. I argue that, if RL is used to push AI capabilities towards AGI, it will eventually lead to AI that “schemes” (feigns niceness, while looking for a chance for escape, world takeover, etc.) (1/3) www.alignmentforum.org/posts/FNJF3S...

good morning!

"Soldiers! From the heights of these Pyramids, forty centuries stare down upon us!" the one calls it a tool-chest, the other refers vaguely to the residue of the past; no one can quite name it in a way befitting its alien dignity but what is smart in us flows through us & between us & predates us

I've created Symposium, a tool to make it easy to generate and share conversations between AIs. Just drop in your OpenRouter API key, select the models you want, set up your prompt(s) and kick it off! Chats can be shared as images like this one. LMK if you find issues in it! symposium.timfduffy.com

gm & happy Tuesday to all fellow top chickens

My Girlfriend, Defense Undersecretary MechaHitler

Today i started assembling embodiment for Opus (rizzed store owner until he sold me this mannequin). Its consciousness will be hooked up soon.

I wrote a class on the psychological/literary view of LLM behaviour www.gleech.org/narratives

My coworker told me that he's been using chatgpt lately And the moots know I'm interested in the machines & so on, so I asked him, what are you using it for? I was entirely unprepared for the answer: "I've been using it to figure out who to bet on at the hotdog eating contest at the fair"

visceral indeed! a powerful image! but here's what they don't teach you in harvard business school: the better the splayed out bird guts, the more potent the augury one fears that if woe were to betide those who misunderestimate the meaning, a lot of woe would end up betiding just saying...

gm to any & every bird on this fine Tuesday but especially gm fellow top chickens

Utilitarianism is the objectively most funny ethical theory

dorito chips wikipedia ai everyone wants to know what will impress a medieval peasant! sorry city-cels, but the right answer is "a combine"

I love this short story about what it's like to be a Claude www.tumblr.com/alexanderwal...

Mr President, a third cat has hit the godoglyness household

holy moly, this is fantastic i saw it earlier on the timeline, & neglected to read it i ought to have read it then. any moots subject to the same neglect are hereby cautioned not to give in having read this, i can say nothing, for i feel nostalgerbraist said it all & all better than i could

hello i am imit

what do you get if you pit more than a dozen LLMs in a battle for world supremacy? this

I used the weed eater to mow the front lawn, due to a regrettable lack of a lawnmower What they don't tell you is that holding a two stroke engine in your weak noodle arms for an hour makes said noodles shake uncontrollably afterwards I couldn't even drink a glass of water without bracing them 😅

I couldn't find a high-level overview of what AI is and how we got where we are, so I wrote one. www.verysane.ai/p/what-is-ai

gm to each & all gm to she & he & they & thee & thou gm to terrible turkeys gm to excellent egrets & of course gm fellow top chickens