sarahconstantin

Karma: 9,068

sarahconstantin 11 Dec 2025 19:43 UTC
2 points
0
on: sarahconstantin’s Shortform
links 12/11/25: https://roamresearch.com/#/app/srcpublic/page/12-11-2025
- example of AI report, evaluating preclinical validation of a drug target:
  - https://elicit.com/review/824f5bb9-bc8e-4c06-b277-a8c342b19bbd
- https://www.macroscience.org/p/to-get-more-effective-drugs-we-need Ruxandra Tesloianu and Jack Scannell make the case for clinical trial reform; I couldn’t agree more
- https://astera.org/neuroscientist-doris-tsao-joins-astera-to-lead-its-new-neuroscience-program/ Astera is doing neuro now. consciousness is a focus.
- https://www.averi.org/ Miles Brundage’s new AI safety verification org

sarahconstantin 11 Dec 2025 19:26 UTC
6 points
0
on: Eliezer’s Unteachable Methods of Sanity
I liked your account of:
looking in the internal direction of your motor plans, and writing into your pending motor plan the image of you getting out of bed in a few moments, and then letting that image get sent to motor output and happen.
I do a similar sort of thing myself sometimes, and similarly do not think it is the same as predictive processing theory, which I don’t believe in for several reasons.
I would call that “visualization”, and I’d say that it’s not hyperstition/woo because it’s not believing in a prediction, it’s forming a plan. (Except that the actually-effective way of forming a plan, the kind you’ll actually be significantly more likely to do than most “plans”, involves vividly simulating the actions you’re going to take.)
A part of this technique involves being really honest with yourself about whether you have any intention of doing the thing, and if you don’t, don’t even try to do the visualization exercise. You have to be able to feel the difference between “yeah I’m gonna do it” and “nope, don’t want to.”
So to the particular goal here, I have rarely ever seriously held an intention to be “sane”, to “cope well,” or to “avoid drama”. to the extent that I have seriously wanted to be a bit better on that axis, I succeeded. It’s not quite as simple as “anyone can do it just by wanting to”, but rather “if you don’t want to, it will definitely not happen, and at least some people, n=1, don’t particularly want to.”

sarahconstantin 9 Dec 2025 17:45 UTC
2 points
0
on: sarahconstantin’s Shortform
links 12/9/25: https://roamresearch.com/#/app/srcpublic/page/12-09-2025
- https://www.gleech.org/erect did walking erect lead to human intelligence?
  - he doesn’t seem to be aware that “behavioral modernity” is looking a lot older than 40,000 years ago
- https://www.poetryfoundation.org/poems/48417/high-windows I still can’t tell whether he’s in favor or against the “long slide”
- https://www.mcsweeneys.net/articles/the-new-yorkers-isaac-chotiner-interviews-santa-claus

sarahconstantin 3 Dec 2025 20:57 UTC
5 points
0
on: sarahconstantin’s Shortform
links 12/3/25: https://roamresearch.com/#/app/srcpublic/page/12-03-2025
- https://covalent.bio/ AI-based life science data extraction
- https://en.wikipedia.org/wiki/Fantasia_(1940_film)
- https://blog.spec.tech/p/150x donate to SpecTech!
- https://nonv.substack.com/p/what-is-yimby-for-healthcare some suggestions on what “abundance for healthcare” would look like: ending certificate of need laws, making state regulations compatible, removing tax and nonprofit advantages that give preference to incumbent hospitals, reforming laws that prevent data sharing and compensation between medical systems

sarahconstantin 2 Dec 2025 16:23 UTC
5 points
0
on: sarahconstantin’s Shortform
links 12/2/25: https://roamresearch.com/#/app/srcpublic/page/12-02-2025
- https://blog.spec.tech/p/beck-brachmans-quest-to-decode-the profile of Beck Brachman of the Imprint project to measure immune repertoire
- https://marginalrevolution.com/marginalrevolution/2025/12/seb-krier.html Seb Krier on AI
- https://reason.com/2025/12/01/hegseths-alleged-order-to-kill-everybody-complicates-trumps-defense-of-his-murderous-anti-drug-campaign/
  - you cannot just kill people, even if you are the government. if they’re criminals, they need to be arrested and charged with a crime. if they are enemy combatants, it’s a war crime to kill survivors clinging to a wreck. (I am not a lawyer but the article quotes several.)
- https://www.zo.computer/
- https://aiwatch.issarice.com/ people in the AI safety space
- https://www.hyperdimensional.co/p/heiliger-dankgesang Dean Ball on Claude Opus 4.5. [
- https://www.lesswrong.com/posts/vpNG99GhbBoLov9og/claude-4-5-opus-soul-document the Claude Opus 4.5 “soul document” is remarkably good.
- https://pmc.ncbi.nlm.nih.gov/articles/PMC12089086/ this is probably nothing (n = 19) but claims creatine improves measures of cognition in Alzheimer’s
- https://en.wikipedia.org/wiki/Freddie_Oversteegen
  - “The Oversteegens and Schaft also killed German soldiers, with Freddie being the first of the girls to kill a soldier by shooting him while riding her bicycle. They also lured soldiers to the woods under the pretense of a romantic overture and then killed them. Oversteegen would approach the soldiers in taverns and bars and ask them to “go for a stroll” in the forest.
- https://worksinprogress.co/issue/the-logical-triumph-of-english/ an article about sentence complexity that doesn’t look into recursive depth? booo.
- https://thezvi.substack.com/p/the-big-nonprofits-post-2025 Zvi Mowshowitz evaluates nonprofits, mostly AI related
- https://en.wikipedia.org/wiki/Duck_typing
  - “With duck typing, an object is of a given type if it has all methods and properties required by that type”
    the main difference from nominative typing seems to be that if you define two types with the same structure, they will be interchangeable.
    what happens more often—this property coming in handy, or this property biting you in the butt?
- https://www.alignmentforum.org/posts/pdaGN6pQyQarFHXF4/reward-is-not-the-optimization-target “reward” in the RL sense is not the same as things a person subjectively finds “rewarding”. Reward is what shapes your strategy; within a strategy, you don’t need to be aware of it as a thing and seek out that thing.
- https://www.construction-physics.com/p/how-a-government-think-tank-trained after WWII, the RAND corporation hired and trained the first large-scale body of computer programmers, to use SAGE to coordinate radar.
- https://www.owlposting.com/p/bringing-organ-scale-cryopreservation more detail on how Until’s cryopreservation works

sarahconstantin 24 Nov 2025 16:54 UTC
2 points
0
on: sarahconstantin’s Shortform
links 11/24/25: https://roamresearch.com/#/app/srcpublic/page/11-24-2025
- https://writetobrain.com/olfactory induced smells with transcranial ultrasound
- https://www.biorxiv.org/content/10.1101/2025.09.02.673859v1.full.pdf looks like bad news for the “it’s the macrophages, stupid” hypothesis; put aged microglia in young brains and they look like young ones, put young microglia in aged brains and they look like old ones (transcriptionally and morphologically.)
  - it may be “the NK cells, stupid”? if you deplete NK cells from the cerebellum, cerebellar microglia don’t get the characteristic age-related interferon signaling.
- https://www.lesswrong.com/posts/apHWSGDiydv3ivmg6/varieties-of-doom JD Pressman explains what he actually thinks will happen with AI
- https://www.lesswrong.com/posts/LaWmoy4s2DLe9u2zq/llm-chatbots-have-half-of-the-kinds-of-consciousness-that Andrew Critch is confident that chatbots introspect, have purposes, have awareness of awareness and perception of perception and memory of memory, have symbol grounding, etc.
  - i confidently disagree, but it would be hard to prove it.
- https://www.untillabs.com/blog/glass-not-ice Until lays out how they intend to improve vitrification for reversible organ preservation (and ultimately preserving whole people).
  - i’d need a lot more physics/chemistry knowledge to have much sense of how feasible this is.

sarahconstantin 19 Nov 2025 18:13 UTC
2 points
0
on: sarahconstantin’s Shortform
links 11/19/25: https://roamresearch.com/#/app/srcpublic/page/11-19-2025
- https://lettersfrombethlehem.substack.com/p/glp-1s-add-satiety-they-dont-suppress personal experience on GLP-1 drugs
  - also, wow, having a satiety problem seems awful
- https://www.dwarkesh.com/p/bits-per-sample Dwarkesh Patel explains why reinforcement learning is so compute and data intensive relative to pretraining
- https://stefangolas.substack.com/p/what-happened-to-cloud-labs skeptical take on lab automation
- https://www.renaissancephilanthropy.org/mathbench-towards-evaluating-natural-language-proofs an AI benchmark for natural language proofs
- https://attentionalleadership.com/wp-content/uploads/2020/07/Cognitive-Techniques-for-Building-Confidence-and-Enhancing-Performance.-Zinsser.-Bunker.-Williams..pdf athletes use positive visualization to improve performance
- https://www.sog.unc.edu/sites/default/files/course_materials/3.0-Eight%20Behaviors%20for%20Smarter%20Teams%20%26%20Approaches%20%28Tues%29.pdf “eight behaviors for smarter teams” Roger Schwarz on workplace communication
- https://www.corememory.com/p/exclusive-a-bold-new-bid-to-control-genome-generalcontrol [[Lada Nuzhna]] has a company based on a CRISPR variant for multiplex epigenetic editing. they’re going after obesity, sarcopenia, and fibrosis, and they have a Novo Nordisk partnership.

sarahconstantin 14 Nov 2025 20:12 UTC
6 points
0
on: sarahconstantin’s Shortform
links 11/14/25: https://roamresearch.com/#/app/srcpublic/page/11-14-2025
- https://en.wikipedia.org/wiki/Uzi was named after a guy named Uzi!
- https://themultiplicity.ai/ compare multiple models
- https://en.wikipedia.org/wiki/Jin_Yong author of wuxia adventure novels in the 1960s that are apparently hugely popular and influential in China
- https://web.archive.org/web/20131017212859/http://www.lunalindsey.com/2013/10/splines-theory-spoons-metaphor-for.html autism makes task switching hard
- https://jobs.ashbyhq.com/episteme yeah this looks like brain emulation. (the quantum physics bit may be a separate project.)
- https://www.computerworld.com/article/1702433/i-m-ok-the-bull-is-dead.html classic on communication
- https://cutlefish.substack.com/p/20-things-ive-learned-as-a-systems guilty as charged

sarahconstantin 11 Nov 2025 15:54 UTC
4 points
0
on: sarahconstantin’s Shortform
links 11/11/25: https://roamresearch.com/#/app/srcpublic/page/11-11-2025
- https://www.corememory.com/p/exclusive-altman-and-masa-back-episteme-louis-andre Episteme, Sam Altman-funded “new Bell Labs” for science. sounds potentially exciting.
- https://en.wikipedia.org/wiki/Texas the name “Texas” comes from a Caddo word for “friends”
- https://machinephase.systems/ these people say they want to build true molecular nanotechnology
- https://yiddishsongs.org/rozhinkes-mit-mandlen/ it’s a touching lullaby about...hoping your kid will go into the import-export business. also railroads.
- https://sproutstack.substack.com/p/is-mike-wazowski-jewish-or-polish my vote is for Polish
- https://lettersfrombethlehem.substack.com/p/i-was-born-with-a-defective-brain personal account of a brain malformation
- https://en.wikipedia.org/wiki/Choice_sequence you can do a bunch of real analysis from a constructivist approach.

sarahconstantin 5 Nov 2025 19:06 UTC
12 points
0
on: sarahconstantin’s Shortform
links 11/05/25: https://roamresearch.com/#/app/srcpublic/page/11-05-2025
- https://www.thirdoikos.com/p/life-in-the-third-oikos-jesse-genet description of daily life for an entrepreneur turned homeschool mom, interview by Nicole Ruiz
- https://en.wikipedia.org/wiki/Chumash_people they’re still around!
- https://www.nytimes.com/2025/11/02/arts/television/maria-riva-dead.html?unlocked_article_code=1.yU8.uFZP.tjEhyXyNasNz&smid=url-share Maria Riva, Marlene Dietrich’s daughter, had a rough time growing up
- https://builders.genagorlin.com/p/the-hidden-belief-that-kills-great by Gena Gorlin, if you don’t have high standards for employees it might because you’re misanthropic, aka you secretly believe “everyone sucks except me” and must be accommodated.
- https://transluce.org/ AI interpretability org

sarahconstantin 4 Nov 2025 17:15 UTC
8 points
0
on: sarahconstantin’s Shortform
links 11/4/2025: https://roamresearch.com/#/app/srcpublic/page/11-04-2025
- https://thenewfrontier.substack.com/p/unintuitive-things-about-impact-philanthropy This is exactly what it’s like.
- https://oldenlabs.com/ once again someone is trying to do smart cages for animal research. I want somebody to win here.
  - Animal research is so, so, so low-quality, especially when it’s not IND-enabling. It costs less to mistreat your mice. And to have too few mice, so your studies are underpowered. And to cut corners on experimental method & disease-model validation in every way. Without the FDA breathing down their necks as tightly as they do for human clinical trials, scientists and their funders do not shell out the cash or spend the time to do animal experiments right.
  - Animal welfare and scientific integrity are extremely aligned here. And if you can get the cost of high-quality animal husbandry down (with automation or any other means), we might start to get “we cured cancer [IN MICE]” to be less bullshit.
- https://substack.com/@foxchapelresearch/p-177604037 skeptical post criticizing Substrate. I’m skeptical of the skepticism.
  - he says “the founder is a known con artist” and describes his former sleep tracker startup Hello as a “scam”. This is a giant abuse of language. Hello, it seems, built a product that was poorly reviewed in the tech-skeptical press, took a long time and a lot of VC funding to ship hardware, and ultimately fizzled. That is what we call a “failed startup.” Not a crime.
    Hello was also incorrectly described as James Proud’s “last project.” In fact, in 2020 he founded Config, a hardware design software company. This is substantially more relevant to his transition into the semiconductor industry.
  - The author claims that Substrate’s images are suggestive that they’re using maskless lithography, which indeed allows for smaller feature sizes, but is not used at scale for semiconductor fabrication because it’s too slow. The argument is essentially that Substrate only gets their small feature sizes by using a method that the semiconductor industry has already rejected for good reason.
    BUT:
    actually maskless lithography is used in the industry, for things like rapid prototyping and making masks. it would be classic “disruptive innovation” for a company to specialize in a new (x-ray-based, in this case) maskless method for these “minor” applications, and then figure out a way to transition to the “major” application of mass production of leading-edge semiconductor chips.
    “you say you compete with the juggernaut monopoly, but today you only outperform them on this narrow use case” is the story of SO MANY SUCCESSFUL COMPANIES that left the monopoly in the dust.

sarahconstantin 3 Nov 2025 16:42 UTC
2 points
0
on: sarahconstantin’s Shortform
links 11/3/25: https://roamresearch.com/#/app/srcpublic/page/11-03-2025
- https://derangedphysiology.com/main/required-reading/cardiovascular-intensive-care/Chapter-642/ecmo-circuit-configurations probably only of practical interest if you work on an ICU, but there’s some dark humor as well
- https://en.wikipedia.org/wiki/Bubble_tea originally invented in Taiwan in the 1980s; became popular throughout East Asia before taking off in the US
- https://www.newyorker.com/news/q-and-a/why-bidens-white-house-press-secretary-is-leaving-the-democratic-party Karine Jean-Pierre gets the Chotiner treatment. Nice schadenfreude.
- https://news.mit.edu/2025/your-brain-without-sleep-1029 the glymphatic hypothesis of sleep strikes again; attention lapses during sleep deprivation coincide with CSF flushing

sarahconstantin 30 Oct 2025 21:34 UTC
2 points
0
on: sarahconstantin’s Shortform
links 10/30/25: https://roamresearch.com/#/app/srcpublic/page/10-30-2025
- https://substrate.com/our-purpose
  - Substrate is using x-ray lithography instead of EUV. but i thought XRL was a dead end because it’s ionizing radiation that can blur features by damaging the mask or the substrate. last i looked it up, minimum XRL feature size was 30 nm, worse than EUV’s state of the art ~16-18 nm gate length for a “3 nm node”.
  - so what’s going on?
- https://newsletter.semianalysis.com/p/how-to-kill-2-monopolies-with-1-tool Substrate is reporting 12 nm features and 0.25 nm CDU (a function of feature variability, typically like 3 standard deviations). so this is a claim of substantial improvement over SOTA.
- https://en.wikipedia.org/wiki/Synchrotron_light_source when you run electrons real fast in a circle they emit very bright, very coherent x-rays. this is what Substrate is doing. there might be something about the coherence or the intensity that helps with blur? but the details are a trade secret and Substrate isn’t saying.
- https://www.epistemic.garden/ Twitter analysis tools
- https://www.owlposting.com/p/there-arent-enough-smart-people-in once again, it is not “boring” to work on repurposing (best biotech approach EVAR if you want to specialize in computational vs. laboratory capabilities) or CROs (if you want to improve on SOTA of experimental or manufacturing technology, and build a business around that, it is much harder to do that within a biotech startup that also needs to build a pipeline, than to specialize in experiments or manufacturing as services)
- https://builders.genagorlin.com/p/reflections-upon-turning-40 probably helpful to do a timeline of your life and the milestones in it, as seen in this post. it’s kind of a trip to see just how *many* things i’ve tried in life by age 37, and how like...20% at best? were unqualified successes.
- https://www.reddit.com/r/AskReddit/comments/1ljeg4e/women_of_reddit_whats_something_they_never_tell/ most of these are very common pregnancy experiences that I also had.
- on readiness potentials and volition in neuroscience
  - https://socsci3.tau.ac.il/rmukamel/wp-content/uploads/2017/05/Neural-correlates-of-intention_book.pdf
  - https://www.cell.com/trends/cognitive-sciences/fulltext/S1364-6613(21)00093-0
  - https://www.sciencedirect.com/science/article/abs/pii/S0168010215000309 [[Parkinson’s]] patients show delays in W-time (they don’t notice having decided to move until right before they do)
  - https://www.biorxiv.org/content/biorxiv/early/2024/05/01/2024.05.01.592070.full.pdf pupil dilation tracks and precedes self-reported intention to move?
  - https://www.sciencedirect.com/science/article/pii/S0028393215300786 readiness potentials in phantom limbs look just like they do in actual limbs with real movements
  - https://www.academia.edu/32463447/Do_meditators_have_higher_awareness_of_their_intentions_to_act not sure i agree with the interpretation here; looks more like “nonmeditators show different brain activity between voluntary movement tasks when asked to introspect vs not; meditators’ brain activity looks like the activity for the non-introspection case in both tasks”
- could direct lithium brine extraction get cheaper?
  - yes, with new adsorbent materials https://newatlas.com/energy/volt-low-cost-lithium-extraction/
  - or with new membrane materials https://www.greyb.com/blog/membrane-based-lithium-extraction/
  - https://docs.nrel.gov/docs/fy21osti/79178.pdf this is already priced into NREL’s techno-economic analysis; this document projects scale-up of various technologies that are currently benchtop or pilot scale.
- https://www.nature.com/articles/s41591-024-03057-9 6 distinct “biotypes” of depression based on fMRI clustering; show different symptom patterns and patterns of performance on cognitive tests, and different responses to treatment.
  - this seems pretty rigorous
  - otoh i wish there were characteristic “types of depression” that were more recognizable on the symptom/behavioral level, such that a therapist or psychiatrist could make a guess as to which type a person might be. this isn’t that.
    old-school “melancholic depression is when you don’t eat, don’t sleep, frown super deeply and wring your hands, have intense guilt, and have high cortisol” would be an example of a much more stereotype-able depression subtype, the kind that’s more useful in practice than any of the 6 “biotypes”.

sarahconstantin 23 Oct 2025 19:56 UTC
6 points
0
on: sarahconstantin’s Shortform
links 10/23/25: https://roamresearch.com/#/app/srcpublic/page/10-23-2025
- https://www.betonit.ai/p/the-anti-intellectual-university
  - Sure, I respect the integrity of standing by your work and your student, and I have no opinion on the correctness of the work, but I can’t stand this contrary streak in Bryan Caplan. If you say straight out “I don’t care if I make people mad”, then you make ME mad!
- https://www.dwarkesh.com/p/thoughts-on-the-ai-buildout
  - Dwarkesh Patel predicts the AI buildout. There’s a lot of money in it. Maybe so much money that it can overwhelm regulatory barriers and ordinary incompetence, and lead to a nontrivial increase in electricity generation??? I can’t imagine this actually happening, but it is a lot of money.
- https://vividvoid.substack.com/p/how-to-fight-your-family
  - if I’m going to “fight” my family, i’m going to fight foolishly and insanely. if i have self-control, why am i fighting at all?
- https://writing.antonleicht.me/p/dont-build-an-ai-safety-movement
  - again, i don’t get it. it would be no great loss if all chatbots were banned. and I’m x-risk skeptical and very anti-regulation in general and a power user of LLMs. It’s just that this is a technology we can live without, because we did live without it five years ago! it’s mostly used for dumb memes and cheating on homework! Banning AI is technically against my principles, but I’d miss it barely more than I’d miss the legal sports betting apps I don’t use.
- https://quintinfrerichs.xyz/human-enhancement-companies very intriguing
- papers related to (perceived) volition and the decision to move:
  - https://www.cell.com/neuron/fulltext/S0896-6273(10)01082-2?innerTabvideo-abstract_mmc3= certain neurons in the medial frontal cortex increase their firing rate prior to the decision to move, peaking at “W-time”, the point where people report making the decision to move (about 200 ms before actual movement)
  - https://www.nature.com/articles/nn1160 damage to the parietal cortex makes people report “w-time” later, close to the actual movement
  - https://www.sciencedirect.com/science/article/abs/pii/S1388245723005953 patients with tic disorders mostly report making a “decision” to move prior to moving, same as healthy subjects. no significant differences between W and M times.
  - https://pmc.ncbi.nlm.nih.gov/articles/PMC10330627/ summary of various experiments on W time
  - https://www.pnas.org/doi/abs/10.1073/pnas.1523226113 the “readiness potential” does not represent a final decision to move; people can “cancel” the movement later, going up to but not beyond W-time. in other words, the time at which people actually make the final decision whether or not to move is pretty much the time at which they report deciding to move.

sarahconstantin 22 Oct 2025 13:30 UTC
2 points
0
on: sarahconstantin’s Shortform
links 10/22/25: https://roamresearch.com/#/app/srcpublic/page/10-22-2025
- https://en.wikipedia.org/wiki/Melino%C3%AB Melinoe is the bringer of nightmares, mentioned only in one of the Orphic Hymns. she is the daughter of Zeus and Persephone; her name means quince-colored (aka yellow-green) and she is described as saffron-robed.
- https://www.lawfaremedia.org/article/anna—lindsey-halligan-here interim US attorney rants to journalist

sarahconstantin 21 Oct 2025 14:14 UTC
17 points
0
on: sarahconstantin’s Shortform
links 10/21/25: https://roamresearch.com/#/app/srcpublic/page/10-21-2025
- https://theaisummer.com/diffusion-models/ summary of how diffusion models work
- https://stampy.ai/ a chatbot about AI safety
- https://www.kerista.com/And_to_No_More_Settle.pdf a personal account of Kerista, a utopian polyamorous commune

sarahconstantin 20 Oct 2025 18:23 UTC
18 points
2
on: Scenes, cliques and teams—a high level ontology of groups
I like this article, but I think I sort of “don’t believe in scenes”, or believe they’re inherently sort of disappointing or contain a kind of tension.
A team has a goal orientation and some kind of merit-related criterion of membership (can you contribute?)
A clique can openly be arranged for the benefit of its members. There’s nothing “unfair” or “nepotistic” about prioritizing your family, your friend group, a social club with formal membership like the Elks, or a subculture like the Juggalos. The right answer to “What makes your clique better than anybody else, such that you should spend your time and effort on them?” is “nothing! i love it because it is mine. it suits me. something else might suit you.”
A scene is neither merit-based nor inward-looking. It’s sort of making a promise to collectively pursue the Art, but it also can’t really kick you out if you suck at the Art. It hasn’t committed to a membership boundary (like a clique, which exists for the benefit of these specific people) or an effectiveness boundary (like a team, which exists to get a specific thing done). It’s not willing to own its ruthlessness (like a team) or its self-servingness (like a clique). At best it’s fertile ground for building real teams and cliques. At worst, it seems to promise mutual support and progress towards common goals, but lots of people are going to be disappointed that they can’t count on that support or that progress actually materializing.
For instance, I think there are a lot of individual people in the LessWrong community I’d want to be on a team with because I respect them on merit-based grounds. I also have affection and loyalty to the community as a clique, as a place I feel at home, a subculture I’m fond of, a group with a high density of personal friends, regardless of whether it’s objectively “better” than any other community. But I don’t actually think community membership is evidence of merit. I think that sort of self-flattering narrative is built into the “scene” format and that people who critique it have a point.

sarahconstantin 20 Oct 2025 18:00 UTC
9 points
1
on: Give Me Your Data: The Rationalist Mind Meld
Agreed that more people should share anecdotes.
We don’t have to bring logic into it; I think logical reasoning is good and possible and there’s no need to insist that “most people don’t do it” (and thus that we shouldn’t either??)
Anecdotes are way better than arguments because they point to the history of how someone came to believe a thing (causally, why, how come you believe that) rather than focusing on the legitimacy of believing that thing.
If I want to understand your perspective, and figure out what I think about it, I can suss that out more efficiently by understanding what examples or details motivated you. Maybe the anecdotes will be enough to change my mind. Maybe I’ll be like “oh, ok, i’m familiar with those and ALSO many other things that point in the opposite direction, so my opinion is unchanged.” Definitely, if the claim being made is an abstract one like “class is important”, motivating examples help narrow down in what sense the person means they think class is important. You just get more new information faster, in most cases, if someone is honestly tracing the origin of their beliefs, instead of trying to convince you of them.

sarahconstantin 20 Oct 2025 15:14 UTC
4 points
0
on: sarahconstantin’s Shortform
links 10/20/25: https://roamresearch.com/#/app/srcpublic/page/10-20-2025
- https://replacement.ai/
  - I don’t actually mind this. Most people are low discernment, like convenience above all, and probably need prodding to be more creeped out by the idea of AI creating addictive and manipulative content for children. If you’re already thinking about your standards and boundaries, then you’re not the one who needs the alarmist message.
- https://unplannedobsolescence.com/blog/what-dynamic-typing-is-for/
  - I don’t understand everything here but it’s a good example of the kind of reasoning about tradeoffs (eg between static and dynamic types) I want to see more of.
    I’m still learning the basics about where programming “philosophies” come from and don’t really have one of my own, but one thing I believe is “you should prioritize things related to the programmer’s cognitive limitations—learnability, readability, bug catching, etc—instead of assuming that a good programmer will not screw up.” of course different approaches compensate for different *kinds* of cognitive limitations—the most bug-resistant code isn’t necessarily the most readable—but at least it’s good to be thinking about those tradeoffs, and this post does.
- https://en.wikipedia.org/wiki/Knuth%27s_up-arrow_notation
  - my son says his “dream” is to get a computer that can compute up-arrows.
- https://drcolleensmith.substack.com/p/the-iliad-is-love an ER doc on what COVID was like. wrenching.
- https://courses.aynrand.org/campus-courses/ayn-rands-conception-of-valuing/ liking this one a lot. in general Gregory Salmieri is just a really good lecturer; reminds me how much I enjoyed college philosophy classes.
- https://writing.antonleicht.me/p/the-devil-you-know
  - not really sure what to make of this one. I think the target audience is AI policy professionals who are single-issue focused on x-risk prevention, and the article takes for granted (doesn’t really explain why) that the best way to achieve x-risk reduction through policy is to ally with the more technically literate faction, as opposed to Luddites who dislike AI for its effect on culture or the labor market. I think his point is that normie Luddite policies will not suffice to stop x-risks? but he doesn’t really make it.
  - his analysis of why x-riskers get picked on seems correct: there are normie Luddite anti-AI factions among the Republicans, but x-riskers are a small, weak group and can safely be painted as crazed lefties.
    sure, if you want to get out of this dynamic, one angle is to embrace what you have in common with the Tech Right (you both see AI as more important than anybody else does, you both understand it better, you both are all else equal in favor of economic growth through technology).
    but again, he doesn’t really go into “why isn’t it better to go Full Luddite, if you really believe the fate of humanity is at stake, and if there are lots of people in both parties who hate tech already?”
    one reason you might not want to do this is that it is harder than you think to speak Republican. Sriram Krishnan is explicit that what it would take for him to trust a safety advocate is for them to be a “big supporter” of “anyone on the Right.” I think this is probably true.
    you won’t get any R allies, in today’s political environment, without costly signals of party/ideological loyalty. if until yesterday you were a normal coastal liberal or moderate or libertarian, your overtures to the Luddite Right will be laughed off. People can read, they can see your track record of opinions all over the Internet, they can track where you get funding, your true nature is not a secret.
    therefore your choices are basically “be exactly the Democrat Luddite you’re accused of being, pray for electoral change” (depends hugely on factors beyond your control) or “quietly persuade smart and less-politically-polarized tech-aligned factions in both parties that they should care at all about safety, notice that some of them actually already do, and take the W when you can instead of picking public fights” (probably gets you more incremental gains in the near term.)
  - I don’t particularly have an AI policy wishlist, myself. It’s not in my top 10 most clear-cut issues. I think most of AI’s substantive economic/human-flourishing benefits are still in the future, and may be mostly squandered anyway, so if we cut off “good AI” opportunities that’s not great but it’s more survivable than losing technological (and institutional) capabilities that we’re already dependent on. On the other hand I’m not confidently x-risk-worried enough to necessarily think that we should be paying any hefty economic costs to prevent it.
    If there’s an AI-related issue I have a strong opinion on, it’s “please actually develop the useful applications and the defenses against misuse”. (eg things like AI drug discovery, AI formal verification, etc, with care taken by funders in distinguishing the real thing from buzzword slop.)

sarahconstantin 17 Oct 2025 14:50 UTC
2 points
0
on: sarahconstantin’s Shortform
links 10/17/2025: https://roamresearch.com/#/app/srcpublic/page/10-17-2025
- https://blog.google/technology/ai/google-gemma-ai-cancer-therapy-discovery/
  - model trained on RNA-seq cell data and different cell types, growth environments, states (cancer vs not), and drug exposures successfully predicted a drug that enhanced antigen presentation, making “cold” cancer cells “hot” enough to be more susceptible to immunotherapy.
- https://www.newmanreader.org/works/parochial/volume2/sermon28.html
  - “hm, someone thoughtful recommended this sermon, maybe it has advice I could take to heart? wow, guess not.”
    “They [rich people] desire and mean to serve God, nay actually do serve Him in their measure; but not with the keen sensibilities, the noble enthusiasm, the grandeur and elevation of soul, the dutifulness and affectionateness towards Christ which become a Christian, but as Jews might obey, who had no Image of God given them except this created world, “eating their bread with joy, and drinking their wine with a merry heart,” caring that “their garments be always white, and their head lacking no ointment, living joyfully with the wife whom they love all the days of the life of their vanity,” and “enjoying the good of their labour.” [Eccles. ix. 7-9; v. 18.] Not, of course, that the due use of God’s temporal blessings is wrong, but to make them the object of our affections, to allow them to beguile us from the “One Husband” to whom we are espoused, is to mistake the Gospel for Judaism.”
- https://marginalrevolution.com/marginalrevolution/2025/10/ai-and-the-first-amendment.html
  - IMO, if the First Amendment means that unsavory AI “speech” must be legal, then so be it. Civil liberties are that important. But then private institutions and individuals are going to need to fill that gap by deciding for themselves when to say no to AI. if you ever really need to know that text wasn’t AI-generated, you may need to be much more careful about provenance, e.g. ensuring that it was entirely hand-written, or only made on computers without internet connections, or things like that.
- https://www.derekthompson.org/p/why-are-liberals-more-depressed
  - Many people have commented that “liberals are more unhappy than conservatives” is a pretty robust finding when measured different ways, and one contradictory study doesn’t disprove it. On the other hand I do think there are hard-to-capture nuances about what this piece calls “externalizing.”
    Two people may be “actually” equally unhappy, but the person who places a high value on expressing feelings may describe themselves as unhappy, whereas the person who thinks it’s superior to be stoic or cheerful may say they feel fine while actually having an elevated heart rate, tense body language, a stressed tone of voice, little ability to enjoy anything, etc.
    Right and left definitely have different values around unhappiness; from a left perspective, expressed unhappiness is grounds for sympathy (all else being equal), whereas from a right perspective, it’s more likely to be seen as grounds for blame. This is very hard to disentangle from how people actually feel. Does the judgment “complaining of unhappiness is blameworthy” cause people to feel happier? Does it cause them to feel the same but complain less? Does it attract people who are happier to begin with?
- https://www.scopeofwork.net/on-factory-tours/ factory tours used to be popular tourist attractions!
- https://magicsearch.sofiavanhanen.fi/ search for recommended Twitter accounts based on your interests
- https://www.slowboring.com/p/what-went-wrong-with-biden-and-immigration
  - claims that Biden intended to be more immigration-restrictionist than his administration actually was. that, regardless of your view on what the best policy is, there was the same sort of ineffectual slowness in taking executive actions going on with immigration that people complain of with things like building infrastructure.
- https://www.humaninvariant.com/blog
  - not a huge fan of these opinions (they’re consistently grouchy, and either obvious or not credible), but truly anonymous blogging in the old-fashioned style is rare these days so props for that
- https://www.aipolicyperspectives.com/p/maintaining-agency-and-control-in good AI thoughts by Seb Krier
- https://www.uncertainupdates.com/p/how-i-became-a-5x-engineer-with-claude how Gordon Worley does programming with AI.