sarahconstantin

Karma: 8,535

sarahconstantin May 27, 2025, 2:58 PM
4 points
0
on: sarahconstantin’s Shortform
links 5/27/25: https://roamresearch.com/#/app/srcpublic/page/05-27-2025
- https://www.plex.tv/ appears to be a place you can make a curated hub for streaming videos?
- https://www.descript.com/ record and edit video; generate transcripts from uploaded video or YouTube links. very convenient and fun to use.
- George Priest, writing in the 1980s, about the need for tort reform and the excessive strictness of product liability
  - https://openyls.law.yale.edu/handle/20.500.13051/5001
  - https://openyls.law.yale.edu/handle/20.500.13051/5008
- https://arxiv.org/abs/2504.11501 Dean Ball’s proposal for AI governance
- https://nexus-tool.com/docs/getting-started collective opinion sharing tool
- https://www.nejm.org/doi/full/10.1056/NEJMoa2405847?query=WB orexin agonist helps with the type of narcolepsy that comes from orexin deficiency
  - https://en.wikipedia.org/wiki/Orexin
    peptide produced from hypocretin
    receptors in lateral hypothalamus
    promotes wakefulness, food intake, and increased energy expenditure—the “SEEK system”!
- Anthropic’s AI for science program—apply for free credits
  - https://www.anthropic.com/news/ai-for-science-program

sarahconstantin May 23, 2025, 7:18 PM
2 points
0
on: sarahconstantin’s Shortform
links 5/23/25: https://roamresearch.com/#/app/srcpublic/page/05-23-2025
- Convergent Research is hiring https://jobs.lever.co/convergentresearch
- https://marginalrevolution.com/marginalrevolution/2025/05/no-evidence-of-effects-of-testosterone-on-economic-preferences.html testosterone doesn’t appear to affect economic preferences like risk-taking, in a large (n=1000) randomized trial
- https://www.ribbonfarm.com/2016/07/13/welcome-to-nixonland/ Carlos Bueno on Nixonland
- https://www.gutenberg.org/files/1322/1322-h/1322-h.htm Leaves of Grass
- https://www.poetryfoundation.org/poems/45477/song-of-myself-1892-version Song of Myself
- https://www.ribbonfarm.com/2016/02/25/the-epic-struggle-between-good-and-neutral/
  - yet another “feeling + premodern culture vs. thinking and markets” dichotomy that seems to be more about his own personality than a universal axis of variation
- https://www.ribbonfarm.com/2011/08/26/the-scientific-sensibility/
  - same deal—science is unsentimentality? sort of, but is that the essence of the thing?
- https://www.ribbonfarm.com/2016/04/28/immortality-begins-at-forty/ content warning: aging, cynicism. but actually a piece that has stayed with me

sarahconstantin May 21, 2025, 4:50 PM
2 points
0
on: sarahconstantin’s Shortform
links 5/21/25: https://roamresearch.com/#/app/srcpublic/page/05-21-2025
- https://www.hathitrust.org/ books online
- https://en.m.wikipedia.org/wiki/Database_consumption when people are more interested in the sorts of “worldbuilding” you could put in a database than in stories themselves
- how many people get personalized medicine?
  - https://www.theguardian.com/science/article/2024/aug/18/rise-in-dna-tests-used-to-claim-citizenship-of-other-countries-brexit-eu more than 40M people worldwide have gotten a genetic tests
  - https://www.bbc.com/future/article/20231004-these-people-took-dna-tests-the-results-changed-their-lives 26 millions of Americans have gotten a genetic test
  - https://www.npr.org/sections/shots-health-news/2025/03/28/nx-s1-5342632/concierge-membership-primary-care-doctor-shortage-rural-health-access there are 12,000 concierge medicine practices in the US
  - https://www.businesswire.com/news/home/20220607005447/en/The-Institute-for-Functional-Medicine-Releases-2020-2021-Annual-Report-Highlighting-Impact-of-Critically-Important-Mission-Driven-Work there are 1800 certified functional medicine practitioners in the US
  - https://www.grandviewresearch.com/industry-analysis/genetic-testing-market-report the global genetic testing market is $11.71B
  - the global direct-to-consumer testing market is $1.7B https://www.grandviewresearch.com/industry-analysis/direct-to-consumer-genetic-testing-market-report
  - https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(24)01924-X/fulltext the global market for direct-to-consumer medical testing is projected to be $9B by 2033
- Pharma proteomics Project with the UK Biobank shows lots of proteomic disease associations
  - https://www.nature.com/articles/s41591-024-03142-z.pdf
  - https://www.nature.com/articles/s41586-023-06592-6#MOESM1
- https://en.wikipedia.org/wiki/Occludin OCLN is associated with depression; something to do with gut permeability maybe.
- https://en.wikipedia.org/wiki/GDF15 GDF15 rises sharply w age, probably a compensatory response to inflammation
- https://en.wikipedia.org/wiki/Elastin ELN goes up in the blood w age, as it leaves blood vessels that are becoming more rigid
- https://pmc.ncbi.nlm.nih.gov/articles/PMC3159481/ PEA is a proteomics technique that’s a lot cheaper to multiplex than ELISA or mass spec; uses tiny quantities of antibodies to barcode proteins and then sequences the barcodes.
  - https://olink.com/technology/pea-technology-video

sarahconstantin May 20, 2025, 3:24 PM
2 points
0
on: sarahconstantin’s Shortform
links 5/20/2025: https://roamresearch.com/#/app/srcpublic/page/05-20-2025
- psalm 51:
  - https://en.wikipedia.org/wiki/Psalm_51 English, Hebrew, Greek
  - https://kpshaw.blogspot.com/2015/03/miserere-mei-deus.html Latin
  - https://www.youtube.com/watch?v=36Y_ztEW1NE Allegri Miserere
- https://www.understandingai.org/p/i-got-fooled-by-ai-for-science-hypeheres plasma physicist gets disappointing results trying to implement “physics-inspired neural networks” for solving PDEs. the example in the paper works, but a different equation doesn’t converge at all to the known closed-form solution & he can’t find any settings that get it to work.

sarahconstantin May 19, 2025, 7:07 PM
6 points
0
on: sarahconstantin’s Shortform
links 5/19/25: https://roamresearch.com/#/app/srcpublic/page/05-19-2025
- https://arxiv.org/html/2502.15631 diminishing performance returns to reasoning token number.
  - more difficult questions get allocated more reasoning tokens.
  - higher reasoning “level” doesn’t
- https://arxiv.org/html/2503.04697 there is an an optimal amount of reasoning to do at inference-time and doing more degrades performance. (“overthinking”). you can design models to estimate this for themselves.
  - see also:
    https://arxiv.org/html/2504.05185
    https://arxiv.org/html/2504.21370
    https://arxiv.org/abs/2412.18547
- more on “overthinking” degrading performance:
  - https://arxiv.org/html/2410.21333
  - https://arxiv.org/abs/2502.07266
- https://www.paulgraham.com/safe.html this is an analytical argument by Paul Graham; you can be “nice” (not capture the most value, in zero-sum contexts) and it won’t affect your revenue/profits much at all compared to your growth rate.
- https://en.m.wikipedia.org/wiki/Monte_Cristo_sandwich -- it’s dipped in powdered sugar (!)
- things I looked up while reading about the French Revolution
  - https://en.m.wikipedia.org/wiki/Anne_Robert_Jacques_Turgot physiocrat of my heart!
  - https://en.m.wikipedia.org/wiki/Charles_Alexandre_de_Calonne spendiest minister ever
  - https://en.m.wikipedia.org/wiki/Guillaume-Chr%C3%A9tien_de_Lamoignon_de_Malesherbes critic of royal absolutism
  - https://en.m.wikipedia.org/wiki/Montesquieu
    https://en.m.wikipedia.org/wiki/The_Spirit_of_Law
  - https://en.m.wikipedia.org/wiki/Germaine_de_Sta%C3%ABl political theorist & salonniere

sarahconstantin May 16, 2025, 2:58 PM
7 points
0
on: sarahconstantin’s Shortform
links 5/16/25: https://roamresearch.com/#/app/srcpublic/page/05-16-2025
- https://www.nytimes.com/2025/05/15/health/gene-editing-personalized-rare-disorders.html
  - baby treated for a rare genetic disorder, CPS1 deficiency, with a CRISPR base editing therapy in LNPs. Children’s Hospital of Philadelphia and UPenn researchers. they came up with a brand new gene editing therapy and got FDA approval in 6 months. two IV infusions seem to have fully done the trick!
  - https://en.wikipedia.org/wiki/Carbamoyl_phosphate_synthetase_I is a mitochondrial gene involved in the production of urea. without it you get ammonia buildup whenever you consume protein.
- https://www.nejm.org/doi/full/10.1056/NEJMoa1910549 pembrolizumab is real good with chemo for stage 2-3 triple-negative breast cancer. either way we’re looking at >50% complete responses.
- https://scientific-paper-planner-test.vercel.app/ beta AI app for planning scientific research projects

sarahconstantin May 15, 2025, 2:28 PM
4 points
0
on: sarahconstantin’s Shortform
links 5/15/25: https://roamresearch.com/#/app/srcpublic/page/05-15-2025
- https://www.worksinprogress.news/p/nuclear-reactors-for-dummies classifies nuclear reactors
- https://www.imprint.org/press-release-may-2025
  - “Imprint is building tools to reconstruct immune histories from memory B cells and T cells at scale. This ‘forensic immunology’ approach aims to unlock the secrets held within memory immune cells, which contain an archive of all immune exposures throughout a person’s life – essentially a web browser history of diseases and their hidden triggers.”
- https://www.construction-physics.com/p/fixing-naval-ship-acquisition the Navy can streamline ship procurement by designing its own ships (!), using simpler, special purpose ships, and completing design before construction.
  - up until the 1960s, in-house ship design was the norm.
  - now, military contractors do not have full-time design teams employed (!) and design is routinely concurrent with building, which increases costs.
- https://arcadiamedicine.com/ intended to develop short-acting analogues of MDMA for therapeutic use
- https://usefulfictions.substack.com/p/i-used-to-be-a-high-performing-robot no, when Cate Hall talks about “agency” she doesn’t just mean being high-achieving. being a “high-performing robot” drove her to alcoholism. agency, to her, means the opposite of all that.
- https://www.nature.com/articles/s41586-023-06592-6 UK Biobank proteomic study
- https://backofmind.substack.com/p/boredom-is-a-form-of-pain Dan Davies says you should not ignore boredom while reading something; that’s a sign that you’re forcing yourself to process info you don’t understand.

sarahconstantin May 14, 2025, 5:26 PM
2 points
0
on: sarahconstantin’s Shortform
links 5/14/2025: https://roamresearch.com/#/app/srcpublic/page/05-14-2025
- https://classics.mit.edu/Aristotle/prior.1.i.html
- https://epoch.ai/data/ai-benchmarking-dashboard?view=table#explore-the-data
- https://en.m.wikipedia.org/wiki/Ana_Montes Cuban spy
- https://qwenlm.github.io/blog/qwen2.5/
- https://ouraring.com/blog/oura-healthiest-cities/ man, people seem healthy in Scottsdale
- https://en.wikipedia.org/wiki/Organon catchall term for Aristotle’s works on logic
- https://rsc.byu.edu/civil-war-saints/latter-day-saints-civil-war Did Mormons fight in the US Civil War? Barely, but yes.
  - “For many Latter-day Saints, the outbreak of the Civil War was seen as a fulfillment of a prophecy made by the Prophet Joseph Smith nearly thirty years before (D&C 87).(https://rsc.byu.edu/civil-war-saints/latter-day-saints-civil-war#_note-4) Of course, most of the fighting occurred over a thousand miles away… “The modest number of soldier Saints is partially due to the decided lack of enthusiasm demonstrated by Mormon leadership regarding the conflict. Strong and able young men were not encouraged to fight for a nation that only a few years earlier had essentially banished the Mormon people from her borders. Church leaders spoke out often about the war, and the opinions expressed were generally negative toward both the North and the South.”
- https://www.interconnects.ai/p/openais-o3-over-optimization-is-back
  - “spiky”—good at some tasks, not others, hallucinates actions
- https://www.vals.ai/benchmarks/mmlu_pro-05-09-2025
  - benchmark
- https://www.vellum.ai/blog/llama-3-3-70b-vs-gpt-4o
  - Llama, GPT4o, and a few others on benchmarks
- https://moodle2.units.it/pluginfile.php/732281/mod_resource/content/1/ForceConceptInventory.pdf measuring intuitive misconceptions in physics students; most classes do not effectively dispel these.

sarahconstantin May 13, 2025, 3:25 PM
4 points
0
on: sarahconstantin’s Shortform
links 5/13/2025: https://roamresearch.com/#/app/srcpublic/page/05-13-2025
- https://www.cnn.com/2025/05/12/asia/india-pakistan-kashmir-ceasefire-intl-hnk India/Pakistan ceasefire
- drug price controls executive order
  - https://www.cato.org/blog/trump-attempts-price-controls-prescription-drugs
  - https://www.whitehouse.gov/presidential-actions/2025/05/delivering-most-favored-nation-prescription-drug-pricing-to-american-patients/
- https://blog.sentinel-team.org/p/global-risks-weekly-roundup-202025
  - US/China trade deal reduces tariffs
  - cuts in US military officers, federal research funding
  - Stephen Miller, homeland security advisor and White House deputy chief of staff, says they’re “actively looking at” suspending habeas corpus for immigrants
- https://languagehat.com/why-adjectives/ not all languages have adjectives! read the comment thread too, it’s fascinating.
- https://classics.mit.edu/Aristotle/interpretation.2.2.html “On Interpretation”
- https://dynomight.net/titles/ how to title blog posts
- Claude Sonnet 3.7 benchmark & compute references
- Grok 3 benchmark & compute references
  - https://x.ai/news/grok-3
  - https://www.helicone.ai/blog/grok-3-benchmark-comparison
- GPT4 benchmark & compute references
  - https://www.linkedin.com/pulse/supposed-leak-gpt4-architecture-alvaro-duran-tovar/
- Llama 2 benchmark & compute references
  - https://ai.meta.com/research/publications/llama-2-open-foundation-and-fine-tuned-chat-models/
  - https://www.cursor.com/en/blog/llama-inference
- GPT-3.5-turbo benchmark & compute references
  - https://www.vellum.ai/blog/gpt-4o-mini-v-s-claude-3-haiku-v-s-gpt-3-5-turbo-a-comparison
- calculating inference FLOPs and other compute measurements from model architecture, in transformers
  - https://kipp.ly/transformer-inference-arithmetic/ a now-classic reference post, explains the kv cache
  - https://www.adamcasson.com/transformer-flops.pdf
  - https://jax-ml.github.io/scaling-book/transformers/
  - https://www.artfintel.com/p/where-do-llms-spend-their-flops
    inference time (and hence FLOPs on a given machine)scales linearly with output length if you use a KV cache
  - https://www.baseten.co/blog/llm-transformer-inference-guide/
  - https://www.lesswrong.com/posts/g7H2sSGHAeYxCHzrz/how-much-ai-inference-can-we-do
- https://marginalrevolution.com/marginalrevolution/2025/05/who-wants-impartial-news.html who prefers impartial news over news that shares their point of view? apparently old, rich, male, non-ideological people. (I did not read the paper & thus do not have an opinion on its credibility.)

sarahconstantin May 13, 2025, 3:21 PM
5 points
2
on: It’s Okay to Feel Bad for a Bit
Agreed.
Here’s how I think of it: very happy people exist. Their default state is cheerful, they rarely worry, they only feel bad if something seriously bad happens to them, and they recover faster than average from misfortunes. I think it makes sense to aspire to be a very happy person.
I don’t think the same is true of people who literally never feel even slightly bad, no matter how bad the situation is. What I’ve heard is that even advanced meditators will e.g. feel grief if their children die. They just won’t add to the grief by acting it out unskillfully and being dicks to everyone around. Literal equanimity, in literally all circumstances, is an abstraction. I haven’t seen it in real life, I’m not sure it exists, and I doubt it would be good if it did exist.

sarahconstantin Apr 25, 2025, 3:33 PM
2 points
0
in reply to: Viliam’s comment on: sarahconstantin’s Shortform
thanks, fixed

sarahconstantin Apr 24, 2025, 3:17 PM
5 points
0
on: sarahconstantin’s Shortform
links 4/24/25: https://roamresearch.com/#/app/srcpublic/page/04-24-2025
- https://en.m.wikipedia.org/wiki/Spasmodic_torticollis cervical dystonia: a striatal problem!
- https://pubmed.ncbi.nlm.nih.gov/22234840/ the “geste antagoniste” or “sensory trick”—many motor disorders benefit from touching or prompting the body (either manipulating oneself or having someone else do it) to “get unstuck” from the undesired posture, immobility, or involuntary movement. at the turn of the 20th century, it was thought that this meant the disorder was psychogenic, but no; motor problems with real, visible brain lesions respond to such “sensory tricks.”
- https://austinvernon.site/blog/rockweathering.html Austin Vernon does a deep dive of rock weathering for removing CO2 from the atmosphere.
- Raman spectroscopy and its variants can be really powerful. Non-contact, label-free detection of biological molecules.
  - surface-enhanced Raman can detect viral RNA & even distinguish (some) different viruses https://pmc.ncbi.nlm.nih.gov/articles/PMC10088700/
  - live & dead cells have different Raman spectra: https://www.mdpi.com/1424-8220/7/8/1343
  - an EBV viral membrane protein in infected cells is detectable with Raman spectroscopy https://www.nature.com/articles/s41598-024-56864-y
  - https://pmc.ncbi.nlm.nih.gov/articles/PMC10316396/ CARS (a Raman variant involving lasers) can tell apart DNA, RNA, and some different kinds of proteins and lipids (bands can be determined by the type of bond that’s prevalent in the molecule)
  - https://chatgpt.com/share/680a46e6-fac8-800f-bd52-fc3a2c34e998 OpenAI o3 suggests that if you want to remotely & nondestructively detect the presence of a virus on a surface, your best bet is probably CARS. (I don’t understand CARS well enough to sanity-check the analysis, but individual linked papers say pretty much what o3 claims they say)
- https://asteriskmag.com/issues/09/abundance-at-home Jasmine Sun, Clara Collier, and Kelsey Piper talk about what the “Abundance” philosophy means for civic engagement.
  - shouldn’t you be able to get your neighborhood together to build a playground, instead of only using “community input” as a veto? “Abundance” can easily fall into idolizing Robert Moses-esque government megaprojects and being nostalgic for the days when the community couldn’t stop them...but an important part of small-a abundance is the freedom to build things for ourselves.
- case studies of big pharma companies have shown quantitative efficiency gains from AI:
  - https://www.bioprocessonline.com/doc/ai-breakthroughs-revolutionizing-pharma-tech-ops-at-roche-0001 Roche used predictive apps for yield and CQAs (critical quality attributes, pharma manufacturing speak for KPIs) resulting in 5-10% yield improvements and up to 50% reduction in CQA writeoffs (=bad batches)
  - https://www.sanofi.com/en/media-room/press-releases/2023/2023-06-13-12-00-00-2687072 Sanofi’s AI programs allowed it to accelerate some research processes “from weeks to hours”, and increased the speed of identifying the best lipid nanoparticle for an application from “months to days”. they also used AI to identify 80% of low inventory positions in their supply chain.
  - https://www.pfizer.com/sites/default/files/investors/financial_reports/annual_reports/2022/story/data-and-ai-are-helping-to-get-medicines-to-patients-faster/ Pfizer used AI to analyze patient data in Paxlovid trials 50% faster, and reduced manufacturing cycle time of one step in the process by 67%.
  - https://www.theinformation.com/articles/ozempic-maker-says-ai-is-finally-reliable-enough-to-produce-sensitive-documents LLMs allowed Novo Nordisk to cut its team of clinical documentatation writers from 30 to 3, and finish tasks in minutes that once took weeks. nobody got fired, but they froze hiring for writers and gave them more to do.
- I vaguely remember an example of a medicinal chemist using AI in his workflow and finding it sped up manual tasks, but I can’t find it now and most of the examples I can find are of medicinal chemists being disappointed in LLMs:
  - https://iwatobipen.wordpress.com/2023/04/11/can-llm-understand-chemistry-memo-journal-jcim/
  - https://practicalcheminformatics.blogspot.com/2024/10/silly-things-large-language-models-do.html
- https://www.research.va.gov/research_in_action/Diabetes-drug-from-Gila-monster-venom.cfm GLP-1 degrades fast in the human body; for GLP-1 inhibitor drugs to be a thing, they had to find a long-half-life version, which was discovered serendipitously in Gila monster venom in 1992.
  - https://en.wikipedia.org/wiki/Exenatide#History John Eng, (student of Rosalyn Yalow! ) was intrigued by a lecture he saw where gila monster venom caused lots of inflammation in the pancreas & thought he’d identify the chemical agent behind the effect. Turned out to be the GLP-like hormone exendin-4. he licensed it to pharma in 1996 and it became the diabetes drug exenatide, approved 2005.
    clearly, more had to be done before we could get modern GLP-1 inhibitors—even longer half-life and better potency through molecular modifications.
- https://www.science.org/content/blog-post/you-can-probably-smell-planet-here Derek Lowe, ever the skeptic, isn’t saying it’s not life. The planet K12-18 b is full of dimethyl sulfide and dimethyl disulfide, which produce a strong cabbage-y odor, and on Earth are pretty much only naturally produced by phytoplankton and algae.
  - is an abiotic origin physically impossible? of course not. but this is definitely an Intriguing Thing.
  - because these molecules would break down under light exposure, they must be being continually replenished by some process. that’s also consistent with life.
- what we know about which drugs work in the clinic:
  - https://www.sciencedirect.com/science/article/abs/pii/S1471489203001188 gene knockout phenotype (in mice) correlates well with the success of drugs targeting that gene
  - https://www.nature.com/articles/s41467-019-09407-3 drug side effects in an organ system occur more when there’s genetic evidence linking the drug target to a disease in that organ system.
  - https://www.nature.com/articles/s41586-024-07316-0 drugs with genetic evidence linking the target to the disease are 2.6x more likely to succeed in the clinic
  - https://ascpt.onlinelibrary.wiley.com/doi/pdfdirect/10.1111/cts.12980 drugs with unknown targets or mechanisms of action fail more
  - https://jamanetwork.com/journals/jamainternalmedicine/fullarticle/2565686 drugs are more likely to fail because of poor efficacy than because of safety problems. cancer drugs and drugs from small companies have a worse success rate.
  - https://www.nature.com/articles/nrd.2017.194 it’s common for drug programs to continue even after evidence has come in that the mechanism/action pair doesn’t work (sunk cost fallacy?) and these drug programs are especially likely to fail.
- Biopharma looks at very, very few drug targets. There are only 854 targets for all FDA-approved drugs—compared to 10,000-20,000 distinct proteins in the human body. About 1857 proteins both have experimental evidence of a link to a disease and belong to a known druggable class.
  - https://www.proteinatlas.org/humanproteome/tissue/druggable
  - implication being: it is SUPER HIGH VALUE to open up more classes of proteins to being “druggable”.
    for example by getting drugs inside cells (now possible with nanoparticles!) so that cytosolic and nuclear proteins are reachable.
- inventors of new methods in biology often have physical or computational science backgrounds:
  - proximity labeling (localizing molecules within the cell): Kyle Roux
    https://news.sanfordhealth.org/research/kyle-roux-lab-work/
  - GCamp6 calcium imaging, which allows you to see single neurons firing in real time: Tsai-Wen Chen, Karel Svoboda, Loren Looger
    https://pubmed.ncbi.nlm.nih.gov/23868258/
    https://www.janelia.org/people/karel-svoboda
    https://loogerlab.ucsd.edu/
  - genome sequencing: George Church (biochemistry)
    https://en.wikipedia.org/wiki/George_Church_(geneticist)
  - CRISPR:
    Emmanuelle Charpentier: https://en.wikipedia.org/wiki/Emmanuelle_Charpentier
    Jennifer Doudna (biochemistry): https://en.wikipedia.org/wiki/Jennifer_Doudna
  - single cell RNA sequencing: Aviv Regev (CS, computational biology):
    https://microbiome.mit.edu/team/aviv-regev/
    https://en.wikipedia.org/wiki/Aviv_Regev
  - https://en.wikipedia.org/wiki/Edward_Boyden Ed Boyden, inventor of both optogenetics & expansion microscopy, started in EE/CS before getting his PhD in neuroscience
  - PALM (super-resolution microscopy with fluorescence): Eric Betzig, physics
    https://en.wikipedia.org/wiki/Eric_Betzig
- cryo-EM is old (80s) but atomic resolution cryo-EM of single particles didn’t come along till 2020 https://www.nature.com/articles/s41586-020-2829-0
- https://www.nature.com/articles/s41573-022-00552-x that great Jack Scannell article on predictive validity and why you want it

sarahconstantin Apr 22, 2025, 2:34 PM
2 points
0
on: sarahconstantin’s Shortform
links 4/22/25: https://roamresearch.com/#/app/srcpublic/page/04-22-2025
- https://reason.com/2025/04/21/alito-and-thomas-may-trust-trump-to-follow-a-court-order-but-gorsuch-kavanaugh-and-barrett-seemingly-do-not/ midnight Supreme Court order, 7-2, against Trump sending Garcia to prison in El Salvador without due process
- https://visakanv.substack.com/p/why-frame-studies Visakan Veerasamy’s outline for the new media studies project
- https://www.leash.bio/ building a very big new database of molecules and their protein interactions, designed for training ML models on. This is the Way.
- https://newsletter.rootsofprogress.org/p/sci-fi-without-dystopia Jason Crawford says yes, science fiction in high-tech settings can be dramatic without taking an anti-tech stance.
  - most of his examples are movies & TV but I think Vernor Vinge does an especially good job at telling stories that are centrally about tech, and involve tech far beyond our own, and are even in some ways “utopian”, but aren’t dull. Mostly they center on the problems at the edges of where the tech breaks down, or interpersonal conflicts where the tech defines the means at each party’s disposal.

sarahconstantin Apr 22, 2025, 12:51 PM
8 points
1
in reply to: Dagon’s comment on: The Uses of Complacency
My experience is not, of course, that it happens constantly, but that many people expect you to be able to do it immediately on demand. As someone who’s easily startled it isn’t always easy for me!

sarahconstantin Apr 21, 2025, 2:53 PM
2 points
0
on: sarahconstantin’s Shortform
links 4/21/25: https://roamresearch.com/#/app/srcpublic/page/04-21-2025
- Pope Francis dead:
  - https://apnews.com/live/pope-francis-dies-catholic-church-updates
  - it’s pneumonia
- Gavin Leech on how he uses & doesn’t use LLMs
- https://www.gleech.org/llms
  - he’s not actually that abstemious compared to me. some differences include that he actually knows regexes fluently so doesn’t use em for formatting; and that he doesn’t like flattery.
  - Mostly I also feel creepy using it as a replacement for directly reading a source or for writing anything I actually care about. I don’t want a replacement for my own mind. My most common LLM use case is “replacement search engine”, followed by “accountability buddy/cheerleader”
- https://statmodeling.stat.columbia.edu/2025/04/17/drowning-in-junk-science-is-there-any-hope-at-all/ Andrew Gelman now thinks in at least some fields the problem isn’t (only) bad research methodology, it’s the mindset that research is a formality you have to get through in order to do the advocacy you actually want to do.
- https://malcolmocean.com/2024/01/guru-dynamics-i-can-show-you-how-to-trust-yourself/ Malcolm Ocean on self-trust is just correct, and has a wholesome goodwill that I’m very fond of. A lot of his ideas are obvious to me (having thought through a lot of the same topics independently) but I wish I lived in a world with a thousand Malcolms. He is actually interested in how minds work and how you can manage em well without self-harm or cruelty or authoritarianism; in an era when a lot of people are leaning into “maybe self-harm and cruelty and authoritarianism are good actually” he has stuck to his values, even through marriage and fatherhood.
- https://www.lesswrong.com/posts/rH492M8T8pKK5763D/agree-retort-or-ignore-a-post-from-the-future old Wei Dai post making the point that obviously one ought to be able to call in arbitration and get someone to respond to a dispute. people ought not to be allowed to simply tap out of an argument and stop responding.
  - this is not what I believe.
  - first of all, it goes against ordinary norms about consensual social interactions, where any party can withdraw association whenever they want unless they have made a binding commitment not to.
    “What about courts? people can be compelled to testify.”
    Ok, cool, but that’s the law of the land. Your idea for norms in a rational society is an idea, and it has neither persuaded me to accept them nor taken political power in the territory I live in.
  - second of all, it is costly. Are you saying that my time and energy can be called up, indefinitely, to participate in an argument, and I have to either concede that you’re “right” or keep arguing until I persuade you? what if I really don’t want to be on the record saying things that I may have been pressured into saying, or confused about, especially on a sensitive topic? what if you are incredibly stubborn and refuse to be persuaded no matter how strong my arguments are?
    “no, you get to stop, but you should have to explicitly say you are no longer participating in the discussion before you can stop.”
    well, that’s better, but I still think it’s too much of an imposition to require. it still takes some time and willpower to explicitly assert one is tapping out. I think it’s not unreasonable to expect people to take a hint sometimes and accept silence or a change of subject as a sign of lack of interest in continuing the discussion.
  - here is my position:
    I do not commit to be accountable to just anybody for being fully logically rational.
    I do, as a point of honor, try to avoid leaving up false or misleading information in my writing, and will try to correct it or remove it if I’ve found I’ve made a factual error. I do not want anybody to mistakenly believe a falsehood on my account.
    But I do not commit to admitting I believe anything that I can’t win an argument against. Even if (someone claims) that given other things I’ve said I logically must believe it. Frankly, if for whatever reason, including aesthetic and social ones, I am not willing to say something in print or out loud, then by gum you do not get to make me say it.
    You can judge me for my choices however you like. But you don’t get to compel them unless you have the literal legal or physical power to do it.
    I also do not commit to continue engaging in conversations I don’t want to, answering questions I don’t want to, etc, outside of an explicit committed obligation (e.g. if it’s part of my job.) I am unusually generous with my time and candor, as people go, but that is my choice, not my obligation.
- https://en.m.wikipedia.org/wiki/Lebo_M this is the guy responsible for the African choruses in The Lion King as well as many other movies. Close collaborator with Hans Zimmer
- https://en.wikipedia.org/wiki/Muscari grape hyacinths are not hyacinths at all. Genus Muscari rather than Hyacinthus. (Same family though.)
  - family Asparagaceae, containing asparagus. https://en.wikipedia.org/wiki/Asparagaceae
    wait. I thought asparagus was a fern? no it isn’t! despite their feathery fern-like leaves, asparagus ferns, tree ferns, plumosus ferns, and others are angiosperms, or flowering plants. they have seeds, not spores.
- https://classicalchristian.org/the-lost-tools-of-learning-dorothy-sayers/ Dorothy Sayers’ famous essay proposing a medieval-inspired “classical education” curriculum for the present day

sarahconstantin Apr 4, 2025, 2:20 PM
2 points
0
on: sarahconstantin’s Shortform
links 4/4/25: https://roamresearch.com/#/app/srcpublic/page/04-04-2025
- https://www.science.org/doi/10.1126/science.adr3675
  - thalamic nuclei are involved in consciousness, active in a task that distinguishes conscious from unconscious visual perception
- https://www.precigenetics.com/ nondestructive image-based epigenetics of single cells. how do they do it??? founded by Parmita Mishra
- https://transformer-circuits.pub/2025/attribution-graphs/methods.html
  - individual neurons are polysemantic; not great for interpretability. we need larger interpretable “circuits” or chunks that have a definite purpose.
  - “cross-layer transcoders” read from one layer and write to all following ones; they are meant to replicate the output of the multi-layer perceptron of the input layer & residual stream. the CLT’s feature activations are given by the inner product of the CLT encoder matrix and the residual stream activations at layer l. but we also have the decoder matrix so we can run this in reverse and get an estimate of the original model’s MLP at any layer.
  - then you just train the model with the CLTs, all the while trying to minimize a loss given by the squared errors between estimated & actual MLPs, and a sparsity penalty for the CLT matrix’s feature activations
  - we can replace the original network with a network made of features (rather than the original neurons) which are sparsely encoded (because we asked them to be.) the sparsity means we’re getting rid of polysemanticity and the new neurons “mean” more uniquely-defined things.
  - new network:
    Its input is the concatenated set of one-hot vectors for each token in the prompt.
    Its neurons are the union of the CLT features active at every token position.
    Its weights are the summed interactions over all the linear paths from one feature to another, including via the residual stream and through attention, but not passing through MLP or CLT layers. Because attention patterns and normalization denominators are frozen, the impact of a source feature’s activation on a target feature’s pre-activation via each path is linear in the activation of the source feature. We sometimes refer to these as “virtual weights” because they are not instantiated in the underlying model.
    Additionally, it has bias-like nodes corresponding to error terms, with a connection from each bias to each downstream neuron in the model.
  - from this they build an “attribution graph”—the input string is encoded into tokens, each token can “flow” towards one or more neurons, the weights tell you which paths are important, and you can observe that neurons do things like “continue this acronym” or “say d”
    features can be associated with words, text placements like “numbers ending in 6” or “first two letters of an acronym”, or concepts (like sports)
    there’s also “say [string]” features that activate immediately preceding [string]
  - you can test interpretations by replacing a layer with the feature decoder equivalent, and suppressing given features; this changes what the LLM outputs in predictable ways.
- https://www.science.org/doi/10.1126/science.adr3675 [[neuroscience]]
  - thalamic nuclei are involved in [[consciousness]], active in a task that distinguishes conscious from unconscious visual perception
- https://www.precigenetics.com/ nondestructive image-based epigenetics of single cells. how do they do it??? founded by [[Parmita Mishra]]
- https://transformer-circuits.pub/2025/attribution-graphs/methods.html [[mechanistic interpretability]] [[AI]]
  - individual neurons are polysemantic; not great for interpretability. we need larger interpretable “circuits” or chunks that have a definite purpose.
  - “cross-layer transcoders” read from one layer and write to all following ones; they are meant to replicate the output of the multi-layer perceptron of the input layer & residual stream. the CLT’s feature activations are given by the inner product of the CLT encoder matrix and the residual stream activations at layer l. but we also have the decoder matrix so we can run this in reverse and get an estimate of the original model’s MLP at any layer.
  - then you just train the model with the CLTs, all the while trying to minimize a loss given by the squared errors between estimated & actual MLPs, and a sparsity penalty for the CLT matrix’s feature activations
  - we can replace the original network with a network made of features (rather than the original neurons) which are sparsely encoded (because we asked them to be.) the sparsity means we’re getting rid of polysemanticity and the new neurons “mean” more uniquely-defined things.
  - new network:
    Its input is the concatenated set of one-hot vectors for each token in the prompt.
    Its neurons are the union of the CLT features active at every token position.
    Its weights are the summed interactions over all the linear paths from one feature to another, including via the residual stream and through attention, but not passing through MLP or CLT layers. Because attention patterns and normalization denominators are frozen, the impact of a source feature’s activation on a target feature’s pre-activation via each path is linear in the activation of the source feature. We sometimes refer to these as “virtual weights” because they are not instantiated in the underlying model.
    Additionally, it has bias-like nodes corresponding to error terms, with a connection from each bias to each downstream neuron in the model.
  - from this they build an “attribution graph”—the input string is encoded into tokens, each token can “flow” towards one or more neurons, the weights tell you which paths are important, and you can observe that neurons do things like “continue this acronym” or “say d”
    features can be associated with words, text placements like “numbers ending in 6” or “first two letters of an acronym”, or concepts (like sports)
    there’s also “say [string]” features that activate immediately preceding [string]
  - you can test interpretations by replacing a layer with the feature decoder equivalent, and suppressing given features; this changes what the LLM outputs in predictable ways.

sarahconstantin Apr 1, 2025, 3:22 PM
2 points
0
on: sarahconstantin’s Shortform
links 4/1/24: https://roamresearch.com/#/app/srcpublic/page/04-01-2025

https://www.worksinprogress.news/p/the-prehistoric-psychopath using a nice big dataset of modern hunter-gatherers and prehistoric archaeological sites, what do we find about ancient violence?
- [[Stephen Pinker]] is right that hunter-gatherers and subsistence farmers (contemporary and prehistoric) had much higher rates of violent death than moderns, even including the world wars of the 20th century.
- however, hunter-gatherers seem much less violent than subsistence/preindustrial agricultural societies, which make more and bigger wars.
  - low populations and few resources are an incentive to avoid wars with neighboring tribes, generally by physically staying out of their way.
- most hunter-gatherer deadly violence is murder, not war.
  - murder declined in Europe with the rise of states in the Middle Ages/Early Modern period. stateless societies had more revenge killings and less law enforcement.
  - murder in hunter-gatherer societies is not considered acceptable; most murder is done by “sociopaths” violating the community norms; hunter-gatherers have methods, like capital punishment, to keep “sociopaths” from terrorizing the group.
    possibly “sociopaths” are more likely to become chiefs or leaders in larger, more hierarchical agricultural societies, and lead them in wars?
- humans are far less prone to violently assaulting other humans than primates are to members of their own species, but human violence is much deadlier than primate-on-primate violence. we have more self-control, better coordination and weapons, and more vulnerable bodies.
- https://arxiv.org/pdf/2006.15191
  - “Is SGD a Bayesian sampler? Well, almost.” empirically, the functions learned by common neural net architectures with stochastic gradient descent are close to the functions learned by Bayesian inference over the data with Gaussian processes.
- https://www.cs.toronto.edu/~duvenaud/distill_bayes_net/public/ how Bayesian neural networks can improve generalization
- https://www.reddit.com/r/DrWillPowers/comments/1iy708b/at_this_point_for_me_an_estradiol_lab_level_in/ Will Powers on hormone therapy
- https://www.sully.ai/ AI agents for health care
- https://astera.org/residency/ Astera residency applications open
- https://press.asimov.com/articles/nobel-duel the story of the race to characterize CRF
- https://jasonbenn.com/tools-for-thought-social-media-posts/50-days meditation personal experience
- https://jasoncrawford.org/the-dilution-of-precise-concepts Jason Crawford on how concepts get diluted

sarahconstantin Mar 31, 2025, 4:04 PM
5 points
0
on: sarahconstantin’s Shortform
links 3/31/25: https://roamresearch.com/#/app/srcpublic/page/03-31-2025
- https://www.noahpinion.blog/p/how-brazil-built-a-world-beating
  - The 7th biggest airplane manufacturer is Embraer, a Brazilian company. How did that happen?
    it came out of a state aeronautical institute founded by aviation pioneer Casimiro Montenegro Filho in the 1940s; military-run, well-funded, and famously competitive as an engineering school.
    Embraer was founded in 1969 as a state-run spinout company using a prototype developed by a student. They focused on a specific niche: small commuter planes for the local market.
    Brazil is very big and has bad roads; there’s high demand for domestic air travel.
    Embraer was privatized in 1994 as part of an overall economic liberalization strategy in response to the hyperinflation and falling GDP of the early 1990s. At this point, Embraer got more investment from the private markets and branched out to regional jets and exports.
    “classic industrial policy” playbook: strong government support at first, which was later taken away; focus on exports and keeping up with the technological frontier.
- https://www.cell.com/cell-metabolism/pdf/S1550-4131(23)00303-0.pdf
  - “Death-seq”—identify dead or dying cells as those which have detached from the surrounding tissue, and do RNA-seq on dead vs. live cells to identify the expression properties characteristic of cell death.
  - Then do a genome-wide CRISPR knockout screen on cultured fibroblasts, induce senescence with doxorubicin, and see which gene knockouts make senescent cells more or less susceptible to being killed by the senolytic compound ABT-263. then Death-seq the lot.
  - they find hits, particularly SMAC/DIABLO, a mitochondrial protein whose knockout inhibits senolysis with multiple BCL-inhibiting senolytics. Going the other way, SMAC mimetic drugs promote senolysis.
    Senescent but not proliferating cells get their cell membranes depolarized during senolytic treatment.
    systemic SMAC] mimetic + senolytic combination therapy in a mouse model of idiopathic pulmonary fibrosis reduces senescent cell markers & reduced the histological severity of pulmonary fibrosis.
    the combo also helps with liver function in a mouse model of NASH
  - what about doing Death-Seq with no drug? what gene knockouts predispose to all sorts of cell death (apoptotic or not)?
    they get 13 markers of senescent cell death. most are novel. one is expected (BCL2L1)
  - bad news for senescence markers: the Tabula Muris Senis (single-cell sequencing atlas of mice at different ages) simply does not contain cells with multiple genetic markers of senescence in the aged animals!
  - “Despite an increase in both the percent of cells expressing Cdkn2a and expression levels of Cdkn2a increasing when pooling cells across 23 organs, we were unable to find cells with multiple markers of senescence in physiologically aged mice in either atlas, in accordance with another recent independent cell atlas dataset.”
- https://www.chicagotribune.com/1994/02/08/sour-note-metropolitan-opera-dumps-soprano-kathleen-battle Operatic soprano Kathleen Battle got fired from the Met for being unprofessional and “nasty”, which is really saying something.
- https://957benfm.com/2025/02/18/timothee-chalamet-method-a-complete-unknown/ Timothee Chalamet sang and played his own songs as [[Bob Dylan]] in “A Complete Unknown”, got a dialect coach and movement coach, didn’t use a cell phone, and gained 20 pounds for the role.
- https://www.thecrimson.com/article/2025/3/28/hms-researcher-detained/ Kseniia Petrova, a bioinformatician at Harvard Medical School, is currently detained by ICE in Louisiana after her visa was revoked because she didn’t declare some biological samples at customs when reentering the country.
  - “When she told a CBP officer who questioned her that she feared political persecution if she returned to Russia, the agency decided to detain her instead, according to one of the petitions filed by her lawyer, Gregory Romanovsky.”
- https://www.lesswrong.com/posts/JFWiM7GAKfPaaLkwT/the-vision-of-bill-thurston Legendary topologist Bill Thurston, famous for his extraordinary abilities to visualize in 4 and 5 dimensions, had no visual depth perception, due to strabismus. He said that he did high-dimensional visualization exactly the same way he had to do 3D visualization—by reconstructing it from 2D projections.
- https://en.m.wikipedia.org/wiki/Jharkhand
  - heavily forested state in India
- https://hbd.gg/play/
  - very tricky “guess where these people originated from their faces” quiz

sarahconstantin Mar 28, 2025, 6:37 PM
2 points
0
on: sarahconstantin’s Shortform
links 3/28/25: https://roamresearch.com/#/app/srcpublic/page/03-28-2025
- https://docs.google.com/forms/d/e/1FAIpQLSc5RvEk4ufotM52v5mKHF7iZOh0Ls_cMDLd5TXn7YOVa62VZQ/viewform Anna Gat offers free mentoring calls to women & immigrants
- https://chatgpt.com/share/67e6ae4a-2138-800f-8e33-f1c50857859f Deep Research report says that 19th-20th c metals companies (Rio Tinto, US Steel, Alcoa, Timet) could routinely make double-digit profits.
- https://www.science.org/content/blog-post/alkanes-mars there are alkanes—big organic molecules—on Mars. these can be produced by abiotic processes, but usually that makes shorter chains than these. so....life? We Shall See.
- https://archive.is/20250326022231/https://www.nytimes.com/2025/03/25/opinion/ezra-klein-podcast-santi-ruiz.html Ezra Klein and Santi Ruiz on DOGE.
  - these are reasonable people. it’s clear to both of them that you do not solve the federal debt with workforce cuts—most of the money is in entitlements.
  - Ruiz is a bit more hopeful than Klein that the conspicuous cuts to (small and sometimes even cost-effective) liberal-coded programs like PEPFAR are the necessary concession to Trump voters (who want to own the libs but may want to keep their Medicare) before serious budget cuts can begin.
  - Klein is suspicious that the whole program is just punishing Trump’s political enemies, which, as a Democrat, even a moderate who accepts the need for fiscal discipline, he has no reason to support.

sarahconstantin Mar 27, 2025, 2:54 PM
2 points
0
on: sarahconstantin’s Shortform
links 3/27/25: https://roamresearch.com/#/app/srcpublic/page/03-27-2025
- https://www.secondperson.dating/p/navigation-by-moonlight
  - essentialist? sure. but this seems like an approximately right version of what “feels natural” to me and gets missed by most men I know.
    most self-help revolves around “if you have a problem, here’s how to take action to fix it”, and there’s a frame from which every version of this is missing the point.
    sometimes the thing that is needed is not solving the problem but encouragement & support in the existing situation, reassurance that it’s a normal situation to be in and doesn’t reflect poorly on you, hope that things will get better...
    sometimes the real “need” isn’t a solution to the alleged problem but sympathetic social support itself. the “problem” being complained about is just a pretext; if it were “solved” you’d pick another problem to complain about; the point is that you need friends who care about you. this is a real need! people suffer without it! yet another book of advice on how to solve today’s problem yourself is not going to help you acquire friends who would take care of you if you were completely helpless.
    obviously if you like a guy and you are following vaguely feminine social scripts your first option is going to be “make yourself conspicuously attractive and show interest in him” rather than “ask him out.” it’s not that asking is totally taboo, but it’s not the default first move! most people start out by trying to execute the briefs assigned to them by society, and only reject those expectations if they find them intolerable.
- https://www.ggd.world/p/how-do-radical-ideas-go-mainstream
  - mostly, reforms to social mores (and IMO feminism was a good one) happen through a process of “mainstreaming” that sands off the radical rough edges and makes them palatable to the general public. IMO this is also a good thing; it’s what makes social change a process of persuasion and winning-over rather than an arbitrary dictate.
    the weirdness of the past 15 years or so is that a bunch of changes skipped straight from Tumblr to the boardroom, leaving everyone else puzzled and dismayed. For a lot of little social-mores things—like “should we have people in videos describe their appearance for the benefit of the blind?”—it’s not clear what the objectively right answer is, but it is clear that persuading most people to agree (on virtually any halfway-sane policy) is better than springing a baffling new policy on them and then having predictably bitter conflict over it.
- https://www.startrek.com/news/shirts-and-skins-in-tos
  - the command uniforms in Star Trek were green, not gold. their shifting color is an artifact of lighting.
- https://thechinaproject.com/2023/06/15/chinas-war-on-drugs-from-incarceration-to-rehabilitation/ a laudatory piece on China’s approach to drug addiction. it’s described as “compassionate” and “humane” but the bottom line is that drug use is not treated as a crime but rehab is mandatory for all. outpatient first, and then in state-run facilities if the person isn’t going to the outpatient rehab.