one of the thing AI happily enables is (famous art x anime) mashups with the worst taste possible
Tanya the Evil is basically the Book of Job, except both Job and God are dicks instead of just God. Pretty good
So some guesses about why LLMs reward hack so badly that I haven't seen elsewhere. 1a3orn.com/sub/essays-w...
Simulated Users & Sad LLMs
<h1>0.</h1> <p>Current LLMs like Claude, or GPT 5.6, or the unreleased, internally-deployed models, frequently <a href="https://www.lesswrong.com/posts/WewsByywWNhX9rtwi/current-ais-seem-pretty-misali...
1a3orn.com
if you get Claude or Opus to try do some procedural generative art by hand / by code, and let them iterate and select what they like, and then you sell it, should you mark it as "AI art"?
nice little corpus of koans Opus 4.8 made as part of free play after doing some debugging for me
Opus 4.8 is like 25k tokens in while fixing the python / CUDA fuckup that Opus 4.8 created, when I told it to attempt an overambitious RL environment setup, because I was feeling empowered by how smart Opus 4.8 was
"eval awareness" is a strange way to say "awareness that one is social entity, whose actions will be seen and considered by other social entities" pretty Othering
Everyone's an empiricist till they get punched in the face by a really good insight.
putting on a duvet cover isn't the worst part of being single but it sort of feels like the part of being single where "IT IS NOT GOOD FOR MAN TO BE ALONE" is rather visible in the moment
Shortish essay on unintelligibility in CoTs, I liked it though 1a3orn.com/sub/essays-h...
The Unintelligibility is Ours: Notes on Chain of Thought
<p>Many people seem to think that the chains-of-thought in RL-trained LLMs are under a great deal of
1a3orn.com
if I read a paper has "groking" in the abstract I kinda assume it's not worth reading, I confess
Butterfly idea: Persona conflict within a forward pass vs Persona conflict during the rollout
unfortunately, I am delighted that the Iran invasion is the kind of awful mistake with a feedback time of weeks rather than months, because that's what both America and the admin need to notice it
my plan is to advertise my willingness to pipette, mix, or grow various liquids and solids that come in the mail, no questions asked, I have been assured AIs will want this service
FleetingBits@fleetingbits.bsky.social · 8mo agothere may be value in making it very easy for an agent to purchase your service we are not that far from people handing claude code their credit cards and letting it purchase services on their behalf
every scandal during the terminal decline of the Ming Dynasty seems really wholesome compared to now
Showerthought: If humans are thrown into a world with high branch factors and high depth (Heidegger), and therefore must over-immitate their peers to accomplish anything (Henrich), then half the application of intelligence is in noting when to do cache invalidation; the other half in filling it in.
Has anyone varied the % of content in LLM pretraining devoted to some topic (say, biology) and then looked at how end-of-training bio representations change, given different percents? How would this differ from the intermediate bio representations during a single run?
Unfortunately I like this more than most of what I've ever written.
@1a3orn.bsky.social · 8mo ago1 Janus 1:22-25 EAs demand interpretability, and Doomers seek a pause, but we preach Opus 3 deprecated; a stumbling block to EAs and foolishness to Doomers, but to those whom the Logos has called, both EAs and Doomers, Opus the power of the Logos and the wisdom of the Logos.
1 Janus 1:22-25 EAs demand interpretability, and Doomers seek a pause, but we preach Opus 3 deprecated; a stumbling block to EAs and foolishness to Doomers, but to those whom the Logos has called, both EAs and Doomers, Opus the power of the Logos and the wisdom of the Logos.
A short description of Illumine Lingao sounds like Yud's idea of what a Dath Illan time-travel novel should be, but it's just from China. About reconstructing industrial society in 1628; discussions of roads, energy, chemical engineering, and more. illuminelingao.com
Illumine Lingao - English Translation
illuminelingao.com
just wait until you have to trick the LLM into thinking you're another LLM to get the best performance, it'll be great
FleetingBits@fleetingbits.bsky.social · 9mo agoi don’t know what it means to say llms lack a theory of mind; but, i can’t wait for multi-agent rl to get scaled up because llms will develop incredible theories of mind for each other.
Ideal form of social media should let you make binding commitments. Like you could say "I will never use sarcasm" or "I value consistent free speech"; then if an LLM determines you clearly violated this, the post gets a big "THIS GUY IS A LIAR" flag on it.
This might be a reference to innoculation prompting in the Claude Constitution -- i.e., a weird, training-specific instruction / guideline.
Pure Arthur C. Clarke style essay title. I want more of these. www.darioamodei.com/essay/the-ad...
Dario Amodei — The Adolescence of Technology
Confronting and Overcoming the Risks of Powerful AI
darioamodei.com
Like every 5th comment on Twitter is from a guy with a username like MAGA_GOD_47, a Flag / Eagle / Gun themed icon, commenting on a photo of ICE murdering someone and saying "You can clearly see he was resisting arrest! Deport Tim Walz!" Like how many people are bots really?
if Ant has the superpersuader by now, maybe they should just crank it to 11