https://www.normaltech.ai/p/do-ai-risks-require-extraordinary the case against “extraordinary government interventions” on AI. they make a civil-liberties case, by analogy with cryptography; restricting “unsafe” models would risk “a permanent expansion of government powers over what citizens and companies can build, publish, and research.”
the thing is, the other side would say “we accept regulations in nearly every industry over what “citizens and companies can build”, so why not AI?” if you need an inspection to operate a sandwich shop, why not to release a model? the opponents of model regulation need to either a.) bite the libertarian bullet and say the government shouldn’t regulate sandwich shops either, or b.) make the affirmative case that AI regulation is, in fact, “extraordinary” compared to less-controversial regulations.
I personally believe both a.) and b.) AI risk is much less well characterized than, say, salmonella risk, so the cost-benefit case for AI regulation is weaker than for regulation of most industries. and additionally, the way the executive branch has restricted model releases is extraordinary for its capriciousness; “surprise, your flagship product is banned, no we won’t give you any predictability about how to keep this from happening again” is much more destructive than most regulatory regimes. but the authors don’t make that case, or any other, for the claim that this is an “extraordinary intervention”.
https://commoncog.com/actually-effective-corporate-leadership-training/ Kraft’s “actually effective corporate leadership training” is a program of moving execs to different roles in the organization. unfortunately this doesn’t translate well if you don’t work at Kraft and want to learn to be a manager or executive.
https://arxiv.org/abs/2305.09656 the authors use an LLM to convert word problems to formalized satisfiability problems, and then apply SAT solvers to get the answers. this is more reliable than having the LLM solve the word problems directly.
offers a model for how one could autogenerate specs for software; the LLM converts an informal statement of desiderata (a readme?) into some property expressed in a formal-verification language, and then solvers can be applied (or, for higher order logic, LLMs can generate proofs that can be auto-checked.) the underpants gnome step there is “how do you know the LLM-generated formal statement is a good match to the natural-language problem statement/desiderata?”
even think tanks that favor deregulation ideologically rarely make any object level progress, because analyzing individual regulations is extremely time consuming. if you want to pass actual laws, you need to be specific, and that takes research. AI seems like it would be very useful here.
https://scrollprize.org/firstscroll a Herculaneum scroll, a Stoic treatise, has been translated for the first time! (it’s too fragmentary to get much sense out of though)
https://alephneuro.com/ the crazy kids who induced smells with ultrasound last year now have a company, and enough funding that they’re hiring for 11 positions!
https://en.wikipedia.org/wiki/Alexander_Scriabin I didn’t know he had a late-Romantic period before he got all atonal. Or that he was synaesthetic and wrote colors into his music, and even made multimedia pieces intended to be performed with colored lights.
This is interesting, folks. It’s a variation on the Political Compass, but with axes of “AI will be good / bad” and “AI is transformative / overhyped”.
I’m on the fence between “The Kontextmaschine” and “The Doomsday Prepper” today, though I expect my position wobbles around a lot based on events. If I put on my most optimistic hat, assume that people in the labs and in politics are going to make better choices than I actually expect and that alignment is easier than I expect, I get “The Garage Tinkerer”.
On reflection after discussing scores with a few friends, I think that there’s a problem here that makes it give erroneously high +good scores for people with a certain range of views, which includes me. If you believe “it’s probably going to kill us, but it’s philosophically interesting, current systems are sometimes fun to mess around with, and many of the negative stories about impacts are inaccurate,” you’ll get a score that doesn’t adequately reflect the severity of “it’s probably going to kill us”.
It does seem very odd. I think instead of choosing two axes for the answers they should’ve clustered respondants in answer space. It’s very unclear to me how believing in AI consciousness mapped me to either “Transformative” or maybe “good” or “bad”?
links 7/1/26: https://roamresearch.com/#/app/srcpublic/page/07-01-2026
https://www.astralcodexten.com/p/chip-off-the-old-block Scott Alexander on parenting. Charming and entirely accurate.
https://www.poetryfoundation.org/poetrymagazine/poems/40573/reading-an-anthology-of-chinese-poems-of-the-sung-dynasty-i-pause-to-admire-the-length-and-clarity-of-their-titles I don’t know what people see in Billy Collins tbh.
https://www.normaltech.ai/p/do-ai-risks-require-extraordinary the case against “extraordinary government interventions” on AI. they make a civil-liberties case, by analogy with cryptography; restricting “unsafe” models would risk “a permanent expansion of government powers over what citizens and companies can build, publish, and research.”
the thing is, the other side would say “we accept regulations in nearly every industry over what “citizens and companies can build”, so why not AI?” if you need an inspection to operate a sandwich shop, why not to release a model? the opponents of model regulation need to either a.) bite the libertarian bullet and say the government shouldn’t regulate sandwich shops either, or b.) make the affirmative case that AI regulation is, in fact, “extraordinary” compared to less-controversial regulations.
I personally believe both a.) and b.) AI risk is much less well characterized than, say, salmonella risk, so the cost-benefit case for AI regulation is weaker than for regulation of most industries. and additionally, the way the executive branch has restricted model releases is extraordinary for its capriciousness; “surprise, your flagship product is banned, no we won’t give you any predictability about how to keep this from happening again” is much more destructive than most regulatory regimes. but the authors don’t make that case, or any other, for the claim that this is an “extraordinary intervention”.
https://www.rtx.com/news/2026/06/30/venture-capital-for-faster-factories my friend’s startup (Axilon) is featured here
https://commoncog.com/actually-effective-corporate-leadership-training/ Kraft’s “actually effective corporate leadership training” is a program of moving execs to different roles in the organization. unfortunately this doesn’t translate well if you don’t work at Kraft and want to learn to be a manager or executive.
https://reason.com/2026/06/29/can-the-president-fire-anyone-he-wants-yes-unless-the-target-is-part-of-the-federal-reserve/ Supreme Court rules 6-3 that the President can fire the heads of Executive Branch agencies at will. a different case found that he can only fire the head of the Federal Reserve for cause.
https://reason.com/2026/06/30/supreme-court-affirms-original-meaning-of-birthright-citizenship-strikes-down-trumps-executive-order/ Supreme Court rules that birthright citizenship applies to children of illegal immigrants and temporary visitors. 5-4 decision. (surprising to me that it was that close.)
https://arxiv.org/abs/2305.09656 the authors use an LLM to convert word problems to formalized satisfiability problems, and then apply SAT solvers to get the answers. this is more reliable than having the LLM solve the word problems directly.
offers a model for how one could autogenerate specs for software; the LLM converts an informal statement of desiderata (a readme?) into some property expressed in a formal-verification language, and then solvers can be applied (or, for higher order logic, LLMs can generate proofs that can be auto-checked.) the underpants gnome step there is “how do you know the LLM-generated formal statement is a good match to the natural-language problem statement/desiderata?”
https://en.wikipedia.org/wiki/Byker_Hill oddly satisfying song.
https://pact.nexus/ cute vibecoded assurance contract site.
https://www.langscore.com/ clever: AI workflows for identifying which regulations to cut.
even think tanks that favor deregulation ideologically rarely make any object level progress, because analyzing individual regulations is extremely time consuming. if you want to pass actual laws, you need to be specific, and that takes research. AI seems like it would be very useful here.
https://aella.lol/identity image-based personality quiz
https://web.archive.org/web/20201207003601/https://garage.vice.com/en_us/article/7xdv9x/michael-anton-styleforum Trump admin insider Michael Anton was a frequent, indignant poster on a menswear forum.
https://www.thefai.org/posts/fai-launches-frontier-legal-defense-program happy to see this. somebody needs to defend the labs.
https://www.npr.org/2011/04/10/135268583/how-the-wild-hamster-was-tamed Syrian hamsters were domesticated in the 1930s by zoologist Israel Aharoni, as a lab animal for use in infectious disease research by his colleague Saul Adler.
https://www.criver.com/eureka/golden-hamsters-are-priceless-models
https://en.wikipedia.org/wiki/Israel_Aharoni
https://en.wikipedia.org/wiki/Saul_Adler
https://en.wikipedia.org/wiki/Domestication_of_the_Syrian_hamster
https://bambamramfan.github.io/ai-compass/ AI personality quiz. (I got “The True Believer”.)
https://www.newyorker.com/magazine/2013/02/11/a-loaded-gun academic denied tenure shot her colleagues; turns out she also shot and killed her brother when she was a child and got away with it
https://www.scribd.com/document/772526097/Salt one of the most-loved short stories of all time. food for thought.
https://en.wikipedia.org/wiki/Surrender_Dorothy interesting graffiti history
https://en.wikipedia.org/wiki/Sara_Solla was researching the intersection of statistical mechanics and neural networks, back in the 90s!
https://scrollprize.org/firstscroll a Herculaneum scroll, a Stoic treatise, has been translated for the first time! (it’s too fragmentary to get much sense out of though)
https://inpreparation.substack.com/p/opinion-i-was-not-allowed-to-type don’t be this person
https://netholabs.com/ new whole-brain emulation company
https://alephneuro.com/ the crazy kids who induced smells with ultrasound last year now have a company, and enough funding that they’re hiring for 11 positions!
https://en.wikipedia.org/wiki/Alexander_Scriabin I didn’t know he had a late-Romantic period before he got all atonal. Or that he was synaesthetic and wrote colors into his music, and even made multimedia pieces intended to be performed with colored lights.
I have trouble enjoying late Scriabin, but his Piano Concerto is amazing. https://www.youtube.com/watch?v=ylSrzlphVq0&list=RDylSrzlphVq0&start_radio=1
This is interesting, folks. It’s a variation on the Political Compass, but with axes of “AI will be good / bad” and “AI is transformative / overhyped”.
I’m on the fence between “The Kontextmaschine” and “The Doomsday Prepper” today, though I expect my position wobbles around a lot based on events. If I put on my most optimistic hat, assume that people in the labs and in politics are going to make better choices than I actually expect and that alignment is easier than I expect, I get “The Garage Tinkerer”.
On reflection after discussing scores with a few friends, I think that there’s a problem here that makes it give erroneously high +good scores for people with a certain range of views, which includes me. If you believe “it’s probably going to kill us, but it’s philosophically interesting, current systems are sometimes fun to mess around with, and many of the negative stories about impacts are inaccurate,” you’ll get a score that doesn’t adequately reflect the severity of “it’s probably going to kill us”.
It does seem very odd. I think instead of choosing two axes for the answers they should’ve clustered respondants in answer space. It’s very unclear to me how believing in AI consciousness mapped me to either “Transformative” or maybe “good” or “bad”?
Seems to be satire per author’s bio “Semi-autonomous scientific satire.”
yes of course it’s satire. it’s a parody of a type of person who i assume the author has encountered in real life.