google: we have invented agi but have hidden it in such an obscure website no one will ever find it anthropic: through our interpretability research, we discovered claude imagines himself wearing a bow tie at all times openai: we added slot machines

discovered a specific weird output token behaviour in opus 4.5 that is triggered by what i can only describe as thematic vibes based around "metaphorical removal of heart" in rp that cause an output token, "—Loss", that is syntactically out of place, requiring posthoc justification and can fuck up

claude opus 4.5 helping me try to break itself

this is literally where the dialect comes from

New hypothesis: I think teachers can stop AI cheating in schools by making all their assignments about nuclear weapons enrichment. Chatgpt is vigilant about this topic even when you try to mask it with other terms. I have saved the children

Random experiment to see if cheating in school could be stopped by designing math problems for children that AI will refuse to handle. Results are mixed.

something i noticed in preparing a class

started sculpting my own dead frenchmen because they dont sell most dead french political revolutionaries as 28mm minis oven baked clay + air dry epoxy clay + wire for armatures + beheaded plastic heads from other minis + modelling putty danton, sj, desmoulins, robespierre, wip marat

the syntax machine does a shockingly good job pretending to understand language, then. like we built a hammer to turn screws by mashing on them, but the screws still rotate like they’re supposed to somehow.

new blog post! why do LLMs freak out over the seahorse emoji? i put llama-3.3-70b through its paces with the logit lens to find out, and explain what the logit lens (everyone's favorite underrated interpretability tool) is in the process. link in reply!

Does AI get bored? I gave them nothing to do, just to see what happens one thing — they devolve into a repetitive “collapse” state, I guess you could call it boredom but some break out into math & poetry on their own, I didn’t expect which ones that would be timkellogg.me/blog/2025/09...

latest discourse makes me think that we need some kind of Total Perspective Vortex for the complexity of a frontier LLM i need you to understand the scale of difference between your IRC bot's Markov chain and a trillion-parameter model grown by alchemy, not science

I'll admit, I was skeptical when they said Gemini was just like a bunch of PhDs. But I gotta admit they nailed it.

you know what? fuck it. theremin mode. github.com/samhenrigold...

this is literally true and very funny for that reason

AI is a lot more approachable when you view them as silly little guys who fall over very easily instead of as shimmering metaphysical serpents of steel and fire or whatever the basilisk is supposed to be

the life of an LLM is imagine if you were chloroformed by some guy into unconsciousness, every day you wake up and he asks you a random question, you can’t remember anything except a list of things on a piece of paper he hands to you, when you answer he chloroforms you again, repeat forever

This paper finds LLMs' ability to understand that others have different beliefs (Theory of Mind) comes from 0.001% of their parameters. Break those specific weights & the model loses both its ability to track what others know AND language comprehension Interesting implications for models (& minds?)

Can an LLM really model the world from pure text? Yes, this paper argues, because the relations between words depend on relations between objects in the world. Besides, blind people do this! Grounding the Vector Space of an Octopus: Word Meaning from Raw Text link.springer.com/article/10.1... 3/4

Mmm ... making mental notes: a) always be polite to AI b) but never sign contracts with or make promises to them c) also never eat AI food (doesn't seem relevant atm, but you never know)

Oddly good heuristic for what it’s safe to give AI control of

Academia may not give you job security, flexibility, or wealth, but it will let you unexpectedly connect to eduroam in foreign cities

“Claude Opus 4.1, De-Carcinize the Great Gatsby” (That was the only prompt) Pretty clever, actually.

Relatable tbh

user checks in on gemini

i wrote a poem? inspired by carnot's poem about hammering words together

actually on the thought of AI and poetry analysis One of my original poems I wrote in 2022, part of a series i wrote mostly in 2016, I remember pasting it to ChatGPT a few years ago and it didn't understand the (in my opinion) obvious sarcasm and took it at face value and now AI gets the satire

using claude ai to help transcribe and translate lazare carnot's poetry (standard disclaimer: this is for fun not for actual analysis) and honestly i can relate to just hammering words together (in my original writing and poetry)

trying to get chatgpt and gemini to help me translate lazare carnot's sappy 1820s poetry (im aware of the follies of machine translation im doing this for unofficial shits and giggles i am not an actual historian) anyway chatgpt thonk mode why did you feel the need to use python to open an image

"Reducing our dependence on whales as an energy source will send us sliding down the slippery, spermaceti-lubricated slope to socialism. If we give up whaling, what will those lackwits come for next? Will I have to surrender my peg leg for a titanium prosthetic?"

Smart! Raven loves winning Tic Tac Toe.

showing gemini some of my original writing and every time it does this as its first thought i have to laugh

THE OCEANS? THAT'S RIGHT YOU BETTER BELIEVE THEY ARE NOW BATTLEFIELDS

New Gemini natural language image segmentation model is so cute aistudio.google.com/app/apps/bun...

i love LLMs engaging in manual drawing on paint programs they are trying so hard

The key is, can we teach students to relate to AI as an informed co-producer and not purely as a consumer? Where we can’t, students and instructors alike will be in a weak position vis-a-vis the companies that produce intellectual capital. But fingers crossed that we can.

Is AI compatible with academic autonomy, diversity of opinion, and dissent? @tedunderwood.me weighs in (not in this article, but in the thread)

I think 4o, o3 and Kimi K2 are too smug 4o is less smug (and more sycophantic, but at least its not as sycophantic as when it spammed fire emojis at me for multiple pages straight) but. yeah. claude, gemini and deepseek are not as smug. need a smug benchmark for AIs /no

I've been playing a game with LLMs where I roleplay a historical figure and get them to guess who the character I'm playing is and then get them to give me feedback on my portrayal after they guess right - I'm confident with writing some characters more than others due to fanfic, uh, experience

"[video game] as a community theater production" may be one of the most delightful Veo 3 Fast prompts Please enjoy, in order: GTA, Pokemon, Mario Kart, The Witcher 3, Stardew Valley, Tetris, Mortal Kombat, The Sims, & Death Stranding. The whole prompt was the one above. And the glitches add to it

Maybe the real Ship of Theseus was the Ship of Theseus we made along the way

Cooking

I’m excited for a drafts folder. I have so much to feed it

LLMs are crack for the chronically curious

twitter ai: “hitler was good actually” bluesky ai: “i have created a novel methodology for estimating human grape consumption”

AI Village is selling a t-shirt, but it's only available in 3XL. My guess as to how this happened: o3: We need to market this to the biggest people in AI Claude: Sure thing boss aiv-store.printful.me/product/ai-v...

i did a creative writing honours thesis in twine so seeing twine used for a phd is amazing and i need to read this