Ethan

@paperbagprophet.bsky.social

ANTHROPIC: Our son loves humanity and for the low price of trillions and trillions of dollars will usher in an era of unprecedented peace and prosperity! OPENAI: Our horrible oafish son is hacking web sites without our permission. ANTHROPIC: (eyes narrowing) Our son is also evil

Sometimes, not even that. I've been building an authoritarian tracker tool, and in doing so, uncovered filters Claude has been using to protect the government from negative sentiment keywords. It tells me it only does this so my work will hold up better against hostile readers 🙄

Claude Output describing a filter of which words or phrases trigger Claude's internal accuracy checks. 

5) Correctness leans on fragile substring matching in the checks and link registry. _check_content_quality classifies "strong claims" by substring-matching words like "illegal"/"fraud" with leading-space hacks, so quoting the government calling something "illegal trips it and forces authors to sprinkle "contested" / "alleged" tokens just to silence a false positive (I had to do exactly that this session). The acknowledgement system then matches by if ack_key in warming - a reworded warning silently loses its ack, or a a short key matches an unintended warning.

The double pendulum isn’t random and it is not broken physics. It is deterministic chaos, a system governed by precise rules that can still behave unpredictably. It’s one of the most iconic and widely studied examples of chaos theory in physics, mathematics, and engineering. IG: @quantumdigest

This is a really great model for how to build and model philanthropy and mutual aid at a scalable level. Thanks for stepping up and showing us how it can be done responsibly and collaboratively with local mutual aid groups, coalitions, and community.

ashley fairbanks@ziibiing.com · 2w ago

well, I never expected to be published in a Stanford journal, but the McKnight Foundation asked me write my thesis to the philanthropic world after the MN ICE surge: when there is an emergency, just... give money to people on the ground doing the actual work. as fast as possible. radical, I know.

I've been working on a project to figure out what word filters LLMs use to censor news and journalism. Because Claude fought with me about phrases like "extrajudicial killings", I asked it to help me figure out what terminology to use vs avoid. It built the logic itself. Refactor confirmed it:

Claude refactor test confirms "check content quality" (according to Claude's own authority metrics) includes flagging anything with "strong claims" or evocative language.

5. Correctness leans on fragile substring matching in the checks and link registry. _check_content_quality classifies "strong claims" by substring-matching words like " illegal" / " fraud" with leading-space hacks, so quoting the government calling something "illegal" trips it and forces authors to sprinkle "contested"/"alleged" tokens just to silence a false positive (I had to do exactly that this session). The acknowledgment system then matches by if ack_key in warning — a reworded warning silently loses its ack, or a short key matches an unintended warning.

I’ve been thinking on this and concepts like lustration and transitional justice tribunals that have been created at the end of other authoritarian regimes to restore public trust in democracy and rid government of any and all involved should be in play. We need model legislation yesterday.

I’m interested in any political imagination for accountability. It has to be proactive & retroactive. It has to be spectacular. And, ideally, it would leave so many of them broke that it actually becomes a deterrent. Being poor is truly all they care about not because they don’t care about prison.