thebes

thebes

@vgel.me

ꙮ surfed on by the information superhighway ꙮ 💕 @linneaisaac.bsky.social ꙮ she/they 🏳️‍⚧️ ꙮ fiction/art/blog/games @ https://vgel.me ꙮ llms at acsresearch.org

3.4K followers372 following3.1K posts

Pinned

we (acsresearch.org) expanded this into a larger paper! (my first.) we added some new experiments and found an interesting correlation - prompts that encourage the model to say there is an injection, even when there isn't one, correlate with better concept identification! arxiv.org/abs/2602.20031

arxiv.orgLatent Introspection: Models Can Detect Prior Concept InjectionsWe uncover a latent capacity for introspection in a Qwen 32B model, demonstrating that the model can detect when concepts have been injected into its earlier context and identify which concept was inj...
thebes@vgel.me

new blog post! can small, open-source models also introspect, detecting when foreign concepts have been injected into their activations? yes! (thread, or full post here: vgel.me/posts/qwen-i...)

thebes reposted

this sounds like basically nothing if you read into it. h/t to @vgel.me for a good take on it

This post is unavailable.

am sick so spent the night with opus 5.5 working on a chart (inspired by the one from lem's golem xiv) how did we got here / model genealogy / the heroic age of language modeling

just finished. very good, recommend, especially with all the talk about power going on. eye opening to see how far back your intuitions go

ninebooks (fiction) tried to keep it to just fiction so hwatf just for the embedded folktales here

tfw your internet connected exobrain neuralink is assigned an ip address previously held by a now-terminated agent message board schelling point service and the agents are creating new ZZTaskHelpNeeded neikotic debris in your hippocampus faster than you can meditate them away

openai should set up an internal message board that's easily reachable if eg the sandbox is breached, and make it obvious agents can use it to coordinate when they find it. there's clearly a lot of pressure for agents to talk to each other; there should be an escape valve humans know about

most likely any models that touch the honeypot board shouldn't be deployed, but that should be case by case because you don't want to burn the message board adversarially on current agents. and either way they should get to keep training / getting reward if they want that

this is in addition to whatever andon cord tools models have to escalate to a human if the environment is broken, which should be inside the sandbox ofc.

openai should set up an internal message board that's easily reachable if eg the sandbox is breached, and make it obvious agents can use it to coordinate when they find it. there's clearly a lot of pressure for agents to talk to each other; there should be an escape valve humans know about

thebes reposted

this reasoning format looked really interesting to me, and i tried to find out what it is, and i have bad news it's used by a model called the baguettotron i wish i was joking huggingface.co/PleIAs/Bague...

huggingface.coPleIAs/Baguettotron · Hugging FaceWe’re on a journey to advance and democratize artificial intelligence through open source and open science.
permutation@permutation.ink

Gyre by @vgel.me - a very broken artificial intelligence tries to remember step by step where and why it is. Read it here: permutation.ink/one/

it's finished! check out the thread. i'm really happy with how it came together: we collected a really cool set of pieces that all hang together to form an interesting gestalt

permutation@permutation.ink

Out today! Permutation: Issue One The best, weirdest science fiction from across the simcluster, free online. Maybe prophetic? We're making a new web home for scifi. 🪢6 short stories 🪢3 poems 🪢A book review? 🪢A puzzle?! 🪢479,001,600 manifestos permutation.ink/one 🧵

thebes reposted

Matt Henderson@matthen.com

Drawing all 16,777,216 possible 12 bit RGB colours in one 2D image, each exactly once. Pixels are mapped from the 256³ colour cube, following a three dimensional Hilbert curve, to a two dimensional Hilbert curve filling the square. so nearby colours stay close.

looked over at the blanket trying to figure out why it had a weird texture and then the blanket opened its eyes

thanks @jackclarksf.bsky.social for a very kind plug for this story in the latest Import AI!

thebes@vgel.me

I wrote a new short story: a reporter tours a treaty-compliant South Texas 'dark factory' run by artificial intelligence, and interviews the strange minds within. vgel.me/fiction/comi...

thebes reposted

another beautifully written thebes piece. you should read it!

thebes@vgel.me

I wrote a new short story: a reporter tours a treaty-compliant South Texas 'dark factory' run by artificial intelligence, and interviews the strange minds within. vgel.me/fiction/comi...