there are lots of smaller, newer companies with varied business models, mostly “too early to tell” for sure if they have the potential to get huge, but I expect in principle many of them should be viable.
sarahconstantin
Karma: 8,130
will fix
What’s Behind the SynBio Bust?
links 1/30/25: https://roamresearch.com/#/app/srcpublic/page/01-30-2025
https://www.astralcodexten.com/p/why-recurring-dream-themes Scott Alexander on recurring dreams
https://darioamodei.com/on-deepseek-and-export-controls Dario Amodei on DeepSeek and export controls
he’s right that efficiency improvements in AI training shouldn’t reduce demand for chips; In This House We Believe in Jevon’s Paradox
I believe him if he says he’s priced in efficiency improvements into his estimates for time and cost to human-level AI performance. it would be stupid not to, and he’s not stupid.
the case he’s making is “no, DeepSeek doesn’t show that export controls backfire by motivating Chinese AI researchers to pursue efficiency.”
I think I agree narrowly.
China’s top AI researchers aren’t stupid either; they’re pursuing efficiency anyhow. So is everyone.
I think it’s never wise to bet on the best people in a competitive field being complacent, even when they can “afford to be.”
Poverty is a cost, not a benefit. It’s not clear to me that China’s chip-poverty (or overall poverty) made DeepSeek more innovative than its American counterparts.
I know there are DeepSeek stans out there who think everything about American culture is toxic, but the world is big, you have not read every AI paper under the sun, and you simply do not have enough information to conclude that they are uniquely innovative.
also, whether we have export controls or not right now, it is Known that the US-China relationship is chilly and that it’s in both countries’ national interest to pursue independent tech stacks.
otoh, while I don’t think DeepSeek’s strong performance was caused by export controls, I also think we can’t make strong claims in the other direction like “if we have these export controls, Chinese AI will be such-and-such amount far behind American AI”. Progress towards more cost-efficient and compute-efficient models is something everyone expects, but it’s very hard to predict how fast it will happen, or whether it’ll be steady or bursty, or how it will be implemented.
if what we really care about, as Dario claims, is ensuring victory in a potential US-China war, rather than economically immiserating the Chinese, then I think it is also very questionable whether the AI that wins wars is the most “advanced” AI.
People like Dario whose bread-and-butter is model performance invariably over-index on model performance, especially on benchmarks. But practical value comes from things besides the model; what tasks you use it for and how effective you are at deploying it.
If you want to win a war, does it help to have an AI that can win math Olympiads? I dunno. All things equal, I’d bet on the country that’s three years behind on the benchmarks and has a functional military procurement system.
we should also note that this is a push from the CEO of a private company who wants to outlaw selling advanced chips to his competitors. there’s a self-interested motive here.
stuff I looked up while reading about Venice
https://en.m.wikipedia.org/wiki/Euboea
home of https://en.m.wikipedia.org/wiki/Georgios_Papanikolaou, father of the Pap smear. apparently, as with so many scientists, his wife was his unpaid lab tech.
https://en.m.wikipedia.org/wiki/History_of_Thessaloniki
https://en.m.wikipedia.org/wiki/History_of_Thessaloniki
former Salonika; second largest city in the Byzantine Empire; crucial port
https://en.m.wikipedia.org/wiki/Pietro_Loredan_(admiral)
https://explorer.globe.engineer/ AI tool for learning about topics; generates heavily nested, illustrated “intros”. I don’t love the workflow; it’s overwhelming rather than sequential. If I’m trying to learn something in earnest, I need to go through it systematically rather than bopping around.
https://www.emilybynight.com/p/why-makeup-works
frustratingly, not a detailed how-to.
https://www.sciencedirect.com/science/article/pii/S2213671118304776 I don’t know how to compare this to anything, but indeed surface microenvironment makes a big difference for mammalian cell culture & cell differentiation.
3D culture is hard; 2D culture can’t get big; solution: tuuuuuubes.
https://www.nature.com/articles/s41467-024-46314-8 possibly the solution to the high cost of cultured meat is by using edible fungi instead.
https://nintil.com/why-so-few-women-in-cs-the-google-memo-is-right Jose Luis Ricon back in 2017 on the Damore memo
links 01/29/25: https://roamresearch.com/#/app/srcpublic/page/01-29-2025
https://marginalrevolution.com/marginalrevolution/2025/01/its-time-to-build-the-peptidome.html link of a piece I’ve already read by Maxwell Tabarrok about the value of building a dataset of peptides and their sequences, structures, and properties, for training models that can discover peptides that can be used pharmacologically (for instance, as antimicrobials).
https://www.biorxiv.org/content/10.1101/2025.01.24.634830v1 using modified T-regulatory cells can be turned against the macrophage foam cells that produce atherosclerosis plaque, reducing heart disease (in mice).
how do they do it?
oxidized LDL is a marker of inflammation and atherosclerosis. It’s also a protein (the acronym is for Low Density Lipoprotein) so you can target it with an antibody!
regulatory T cells (Treg) genetically modified to contain one of these anti-OxLDL antibodies prevents the OxLDL from getting into macrophages and turning them into foam cells.
inject them into live mice on a high-fat diet and they also get less atherosclerosis than control mice
interestingly they make Tregs from CD4 (helper) T-cells by getting them to express FOXP3. this will be useful for human clinical application, since natural Tregs are rare and it can be difficult to extract them in sufficient quantity.
https://talentmarket.org/cato-psych-policy-analyst/ ugh. they basically have already written the bottom line on what they want a new hire to think about the “psychology of progress”, and it’s mostly talking points that have been made ad infinitum already.
surely if you were serious about overcoming psychological barriers in the general public that make them ill-disposed to objectively beneficial economic/technological changes, you’d want to look for new ideas (since clearly past ones haven’t worked). you’d also want to focus on starting with empathy and common ground, if the hope is to actually change minds. this is a job description that, ironically, guarantees stasis, not progress.
https://stephenmalina.com/post/2023-11-04-biologizing-the-stack/ I’m glad he admits this is a contrarian position because the straightforward lesson of the past several decades is that you want as few things to be biomanufactured as possible, living things make incredibly cost-inefficient factories and should be a last resort.
https://unstableontology.com/2022/05/02/on-the-paradox-of-tolerance-in-relation-to-fascism-and-online-content-moderation/ I like this post, but on the general topic, how could you possibly implement Popper’s standard of “we must not tolerate speech that incites people to not listen to arguments”? That would, itself, require an extensive censorship apparatus.
1st Amendment Law seems much more practical in protecting all political opinions and having exceptions only for “true threats” or narrow “incitement to violence”. Saying “don’t read this book, it’s by a fascist” is intolerance by Popper’s definition, but it’s straightforwardly protected speech under US law and I think it should be; we don’t trust a government agency to adjudicate questions of epistemic vice.
1st Amendment law has almost a refreshingly nihilistic attitude to discussion—all “opinion”, including iirc all discussion on social media, is assumed to be neither true or false, so it can’t be considered defamatory. basically, under the law, “this is all just yapping, people get to yap, call me when they make a false factual claim that costs you money, or make an actual plan to physically hurt someone”. sometimes, in addition to being a safer standard for one’s government to hold, this is a healthy attitude to adopt oneself!
Stephen Wolfram on machine learning: https://writings.stephenwolfram.com/2024/08/whats-really-going-on-in-machine-learning-some-minimal-models/
first, we visualize how values at the nodes of a neural network (during inference) change based on the inputs. complicated!
then we look at how weights on edges change as a neural network is trained. also complicated! though you can see the changes getting smaller as the network approaches convergence.
a mesh neural net—each node is only connected to its neighbors in a grid—is a cellular automaton. Cellular automata are a special case of neural networks, in other words.
you can do an analogue of “training a model to fit a function” with something called a “rule array”—you have a grid, and each square can have a local update rule relative to its neighbors, cellular-automata style, but the rule might be different for different squares. If your “input” is a black square at the top row, then “running” the automata rule repeatedly may propagate down a black-and-white pattern on the grid. you can then “adapt” the rule array iteratively to get a desired pattern, like “generate one that lasts for exactly 30 steps”. or a function: you want to get something that gives certain “outputs” on the bottom row, given an “input” at the top row. then “mutate” cells at random, keeping only the results where the distance from the “training examples” doesn’t change.
you can observe which cells mutate along the training runs; they seem to be the ones along the “ideal path” between x (at the top) and f(x) (at the bottom).
doing mutations at random is inefficient; you can do an equivalent of “steepest descent” in cellular automata too, but this tends to get stuck
there’s an equivalent of “backpropagation” in automata-land too
in general, why translate to automata? he says they’re easier to “inspect” because simpler, but I kinda don’t get it.
he even builds a discrete analog of a transformer!
“It could have been that machine learning would somehow “crack systems”, and find simple representations for what they do. But that doesn’t seem to be what’s going on at all. Instead what seems to be happening is that machine learning is in a sense just “hitching a ride” on the general richness of the computational universe. It’s not “specifically building up behavior one needs”; rather what it’s doing is to harness behavior that’s “already out there” in the computational universe.”
yep this is a convergent idea. the secret sauce in machine learning is just having a rich enough space of possible functions, and a means of variation and selection. you can get similar results using things like genetic algorithms or cellular automata that aren’t “neural networks” at all.
Stephen Wolfram on formal verification: https://writings.stephenwolfram.com/2025/01/who-can-understand-the-proof-a-window-on-formalized-mathematics/
Wolfram Language allows proof verification!
you can also visualize the dependency graph of lemmas.
sometimes the “high-degree nodes” are legible things like commutativity
but often they, and “commonly used lemmas” in randomly generated proofs, are weird elaborate non-human-interpretable things.
What about proof-to-proof equivalences?
this is where the “homotopy” metaphors come from. you could find a “path” from one proof to another...but what if there are “holes” in proof-space? “Then a “continuous deformation” of one proof into another will get stuck, and even if there is a much shorter proof, we’re liable to get “topologically stuck” before we find it.”
https://kmill.github.io/informalization/ucsc_cse_talk.pdf
nice explanation of formalization (in Lean), auto-formalization and auto-informalization
if you just throw ChatGPT at this, it sucks.
so let’s do a little ontology hard-coding.
https://moreisdifferent.blog/p/german-scientific-paternalism how Germany in the late 19th-early 20th century trained scientists
https://kordinglab.com/about/ an approach to figuring out what neurons do by mapping between simulations, electrophysiology data, and psychophysics. “what algorithm is being implemented here?”
https://www.hypothesisfund.org/ a Reid Hoffman project: funds breakthrough research, mostly life sciences.
this doesn’t “smell” aggressive enough to me—the projects look fine but i’m surprised they’d be unfundable elsewhere—but maybe it’s just an insufficiently pointed communication style.
https://en.wikipedia.org/wiki/Robert_Duggan_(venture_capitalist)trange fellow. apparently he’s a Scientologist and knows nothing about biology but has a knack for picking winners. picked up ibrutinib!!!
links 01/27/25: https://roamresearch.com/#/app/srcpublic/page/01-27-2025
https://asteriskmag.com/issues/09/where-the-wild-things-arent Agnes Callard is strange and unsettling here. I wonder what this is “really” about.
“Being a person is too hard a job to leave to a single person. We can’t do it on our own, not even as adults. Figuring out how to be a person is a group project, and we have to help each other. But the catch is that we don’t really know what we are doing, so sometimes we end up hurting each other instead. When you are weird, you experience this hurt. Social categories have been poorly constructed and fail to conduce to human happiness. The weird person is a record of the mistakes we have made.”
it sounds like her own “weirdness” is experienced as a source of pain, not as a stable and beloved personal identity. and that she’s uncomfortable with public celebration of “weirdness”.
admittedly it is a bit paradoxical that contemporary culture celebrates “weirdos” in a not-very-individualistic way. the nonconformists who are celebrated do, actually, conform to their own little club rules.
this never particularly bothered me, though!
if i find a club i want to be a member of, that’s fantastic. i’m not attached to being literally unique.
if i find that existing “labels” or “groups” don’t entirely suit me as an individual, that can be a bummer, but i don’t think i would be better off if no fine-grained identity categories existed and i was expected to conform with everyone in my geographical location.
there’s a very natural explanation about why children’s books star alienated weirdos: writers are not typical people!
beloved children’s book authors were writing to children like me, the children who read a lot of books and might grow up to be writers ourselves.
this isn’t some paradoxical thing.
what is “normal” (both common and normative) in the world of books is what is “normal” for text-native obligate readers and writers, which does in fact mean being different from the majority! Bookishness is a minority trait!
Bookish people, as a rule, are glad we are this way, and eager to acculturate potential kindred spirits into bookishness. This seems generally healthy to me.
sure, be a little thoughtful about not making depictions of alienation into self-fulfilling prophecies, but I think a little bit of care and taste suffices. no need to angst about “what if we are BAD ROLE MODELS”. it’s okay to like your own quirks.
https://www.ams.org/journals/notices/202502/noti3114/noti3114.html anatomy of a Lean proof
https://en.m.wikipedia.org/wiki/Marquis_de_Sade
he sounds genuinely awful, though i haven’t read his writing
links 1/24/25: https://roamresearch.com/#/app/srcpublic/page/01-24-2025
https://traditionsofconflict.substack.com/p/traditional-methods-of-black-magic maybe “black magic” beliefs persist because the “cursed” person gets scared, contacts a shaman/magician/etc for help, and “recovers”
https://tori.gg/ AI productivity browser extension?
https://www.biospace.com/isomorphic-labs-announces-strategic-multi-target-research-collaboration-with-novartis only a year ago, Isomorphic Labs announced this partnership, for developing small molecules with AI (via in silico screens or ab initio)
https://www.bloomberg.com/news/articles/2025-01-21/deepmind-expects-clinical-trials-for-ai-designed-drugs-this-year now they’re announcing they’re going to hit the clinic this year?! that is fast. and it’s a bit surprising that it’s small molecules rather than antibodies, when the literature really seems to point to protein structure/affinity prediction being stronger than small molecule binding prediction
https://www.isomorphiclabs.com/ they don’t want to tell us anything. no disease indication. bupkis.
https://www.complexsystemspodcast.com/episodes/banking-crisis-two-years-later/ Patrick McKenzie says the things he couldn’t say in 2023 about the banking crisis
if he, making inferences from public information, said that he believed specific banks were insolvent when that was not yet public knowledge, he’d be accused by the press of deliberately causing a financial crisis or destroying banks.
the government lied about banks being fine, which were not fine, in the interests of preventing the crisis from getting worse.
i dunno whether this is outrageous because the culture that is finance is so unfamiliar to me; afaik saying true negative things about banks will make them fail faster and maybe spread the “contagion” more broadly in the economy? maybe it is in the public’s best interest for the government to lie in such cases?
but at any rate, this implies you should not necessarily believe announcements that a bank is fine.
“Don’t spend time alone with one minor away from the group or conduct private interactions with minors in enclosed spaces or behind closed doors”, “Don’t relate to minors as if they were peers, conduct private correspondence, or take on the role of “confidant” (outside of a professional counseling relationship)” and “Don’t privately email, text, or engage with minors through social media. Group messages and posts are acceptable and must be viewable by all participants” are the objectionable rules here.
obviously the motive was to avoid child sexual abuse, which is important, and I’m even on board with erring on the side of avoiding the appearance of impropriety. indeed, don’t sext the minors or offer them alcohol/drugs.
but a normal, appropriate mentorship relationship will involve one-on-one meetings and emails. teachers talk to students. how can you ban this???
it’s especially outrageous because this is MIT. child prodigies need opportunities to do real things in the world; this will generally involve working with adults. rules like this mean “no more Terence Taos.”
https://www.jstor.org/stable/44162589
only relevant Google result I could find for a James Joyce quote I remembered part of:
Joyce made a specific comparative observation in his notes: “Europe is weary even of the Scandinavian women (Hedda Gabler, Rebecca Rosmer, Asta Allmers) whom the poetic genius of Ibsen created when the Slav heroines of Dostoievsky and Turgenev were growing stale. On what woman will the light of the poet’s mind now shine? Perhaps at last on the Celt. Vain question. Curl the hair how you will and undo it again as you will” (E 125).
this is from the notes/stage directions of Exiles. The play has “full text” versions online but they don’t include the notes.
links 01/23/2025: https://roamresearch.com/#/app/srcpublic/page/01-23-2025
https://www.hyperdimensional.co/p/invitations like so many visions, whether i’d actually like this or not depends on execution. AI-powered smart devices? no thank you. Interpretive “power tools” with lots of configurable options, that allow me to do a lot of media analysis and view embedding coordinates? i’d like that. AI tools to interface with bureaucracies, or to assemble a media diet based on custom preferences, or to resolve disputes, or to provide traditional governmental functions where those are absent or defective? I begin to be intrigued. but you need taste. and taste is a little orthogonal to being at the leading edge of the zeitgeist; the zeitgeist-surfer always wants to “embrace” the direction we’re going, whereas the tasteful person is opinionated and likes some new things but not others, and doesn’t always favor the new things that are destined to be the biggest.
https://psychcrisis.org/mania-guide/ this is the best guide to how to handle a friend or loved one with mania I’ve ever seen. It offers lots of options for getting professional help that are less risky & restrictive than “call 911” which is most people’s first and only thought about what to do when someone has suddenly “gone crazy”.
https://meltingasphalt.com/a-codebase-is-an-organism/ a clear intro to “what do people mean about code “rotting” anyway? (spoiler: big codebases, unlike school programming assignments, get used and changed by many people. this introduces many new problems because you can’t singlehandedly control everything everyone does.)
http://www.laputan.org/mud/ what do people mean when they say code is “sloppy”, “spaghetti code”, or whatever?
poor performance at large scale
Data structures may be haphazardly constructed, or even next to non-existent. Everything talks to everything else. Every shred of important state data may be global.
why is this bad?
Claude explains:
it’s bad if “everything talks to everything else” or “all state data is global” because if you change one part of the code everything else may break
it’s bad if there are no data structures or if they’re chosen randomly, because the right data structure is much more time & space efficient to search/sort/etc than the wrong one
“complicated, convoluted” code that’s hard to read
https://marginalrevolution.com/marginalrevolution/2024/07/an-overly-simple-model-of-positive-and-negative-contagion.html Tyler Cowen: negative emotions are contagious & so focusing on the negative (even to critique bad ideas) has harmful externalities
does he apply this insight to himself? does he try to avoid over- focusing on the negative? i’m not sure...
possibly someone who’s constitutionally non-neurotic will never be able to really understand negative contagion risks. they’ll say things that they assume are fine, and have no real intuition for what stressed-out people are hearing.
links 1/21/2025: https://roamresearch.com/#/app/srcpublic/page/01-21-2025
https://sammatey.substack.com/p/the-weekly-anthropocene-interviews-84c news is about entertainment rather than information...yeah i guess, but i’m not sure what to do about that, i probably don’t actually need to know most of what’s going on!
https://contraptions.venkateshrao.com/p/as-above-so-garage I kind of liked this Venkatesh Rao essay. the “garage project” sort of hackerish engineering is a “hobby” project that has potential to connect to the “big story” of advanced technology (unlike, say, woodworking, which may be personally or artistically fulfilling but is clearly opting out of the race towards the cutting edge) but also is a sort of “pure” personal project that has no obligation to deliver a result or make a profit or please anyone but oneself. it’s that intersection of “probably nothing” but “could be something”, limitless potential at the far horizon but tiny in the here-and-now/
“tinkering in your garage” is the tech equivalent of “starting a band” in the days where rock bands had a shot at stardom
there’s a Fun Theory thing here. it’s not Fun to do things that you know are trivial/meaningless and will definitely never “matter” in a big or deep sense; it’s also not Fun to be burdened with expectations and obligations. Fun thrives in the zone of “playing around that could someday grow into Something”
https://en.m.wikipedia.org/wiki/List_of_largest_cities
NYC is only ranked 35, behind Moscow??
things I looked up while reading about Venice
https://en.m.wikipedia.org/wiki/Carlo_Zeno
Venetian admiral
https://en.m.wikipedia.org/wiki/Andrea_Dandolo
doge
https://it.m.wikipedia.org/wiki/Pagano_Doria
Genoese admiral
https://en.m.wikipedia.org/wiki/Niccol%C3%B2_Pisani
Venetian admiral
https://en.m.wikipedia.org/wiki/Venetian_Gothic_architecture
https://www.ncbi.nlm.nih.gov/sites/books/NBK559170/
https://www.thepopverse.com/movies-nosferatu-count-orlok-bill-skarsgard-voice-robert-eggers yep, Count Orlok in Robert Eggers’ Nosferatu is speaking Dacian
https://www.politico.com/news/2024/08/28/trump-zuckerberg-election-book-00176639 is this a serious threat? who can tell any more. i suppose if people respond to it, it’s an effective threat.
https://substack.com/home/post/p-154556883 Dan Davies on the problem of governments that outsource everything to contractors and NGOs, which then fail to do the job and can’t be held accountable or monitored because the government no longer has the staff or the expertise to oversee the project in detail.
outsourcing was, itself, a response to the high cost (and overspending and overstaffing) of governments running programs directly through civil service departments. the hope was that outsourcing would impose market discipline through competition.
from this point of view i’m not sure what one can do.
if the government refuses to provide public services at all, people might overthrow it?
or at least vote it out in elections
if the government provides public services itself, it will predictably overcharge, underdeliver, and engage in direct abuses of power
if the government hires third parties to provide public services, those parties will predictably overcharge and underdeliver, except more opaquely and in ways that are less amenable to being changed when voters get disgruntled.
http://okayfail.com/2025/i-met-pg-once.html heartfelt and says something I’ve been concerned about myself. when you say you’re “anti-woke”, how can we tell whether that means you’re against specific, recent types of administrative overreach or whether e.g. you actually want to drive gender-nonconforming people out of public life? I’m sure there’s some of both going on, but there’s also a lot of uncomfortable (or strategic) ambiguity, which is a much more pressing concern for those with personal reason to worry whether they’re going to experience a huge rise in discrimination.
links 1/17/25: https://roamresearch.com/#/app/srcpublic/page/01-17-2025
https://davekasten.substack.com/p/welcome-to-the-essay-meta Dave Kasten, known Washington Insider and Dangerous Professional, says that the way to increase the chance your ideas make it into policy is to write essays online.
interestingly I’m reading this just as I’m hearing other people discuss how fruitless it is to write publicly, because nobody will understand you and take action based on your words. I think it probably depends how you write and what concepts you’re trying to get across, and what your bar is for “action.” People who think essays are useful are probably trying to reach across shorter inferential distances and make smaller nudges to people’s behavior.
https://www.rationalistjudaism.com/p/the-kezayis-post what is the minimum amount of matzah one is required to eat on Passover? traditionally it’s “the size of an olive.” but some people claim it’s about the size of a large pizza...because Biblical olives were huge??? (spoiler: no, they were not. we literally have thousand-year-old olive trees producing normal-sized olives.)
https://www.read.ai/ this is a fascinating app. Link it to your calendar and Zoom and it will accompany on you on all your meetings, summarize them, generate action items, and grade you on things like how fast you speak and how much other people seem to like you.
“bias” is their word for “how much positive emotion do you show”, which is a terrible word for that! did they let the statisticians do too much product design?
they will also keep track of how many “non-inclusive” things you say, apparently. i’m still at zero, not being much for slurs myself.
transcription accuracy is pretty good, but the app is often wrong about who’s speaking.
https://marginalrevolution.com/marginalrevolution/2013/04/the-culture-of-guns-the-culture-of-alcohol.html old Tyler Cowen post saying that if we blame guns for homicide/suicide we should, with equal or greater emphasis, also be blaming alcohol. Believable!
links 1/15/25: https://roamresearch.com/#/app/srcpublic/page/01-15-2025
https://www.proteinatlas.org/ seems like a good resource. Swedish.
https://en.m.wikipedia.org/wiki/Human_cloning human cloning was first discussed by JBS Haldane in a 1969 speech!
https://en.wikipedia.org/wiki/Protalix_BioTherapeutics they seem pretty successful. enzyme replacement for Gaucher disease. Israeli.
https://en.wikipedia.org/wiki/Phillip_Frost interesting guy. “served as a lieutenant commander, U.S. Public Health Service at the National Cancer Institute, from 1963 to 1965.” Major pharma investor.
What happened to Amyris?
they used to be a biofuel company but couldn’t get production up and costs down:
they pivoted to low-volume, high-price beauty & personal care ingredients, which actually generated a bunch of revenue, but not enough to cover costs. and then also bought a ton of celebrity beauty brands, which didn’t. 2022 stock plunge, 2023 bankruptcy.
they’re not terrible at industrial fermentation (compared to other synbio unicorns) and have some lessons learned
they got in trouble with the SEC for recognizing more revenue than they actually made (according to standard accounting)
https://www.science.org/content/article/synthetic-biology-once-hailed-moneymaker-meets-tough-times bad times for biomanufacturing/synbio overall
there are kind of...zero large profitable firms founded after 2000 that specialize in industrial fermentation/biomanufacturing, EXCEPT a couple of biotechs that make enzyme drugs.
there’s plenty of biomanufactured products but pretty much all from very large old boring firms at sorta commodity prices?
possibly-intrusive question: are you Russian?
links 1/13/2025: https://roamresearch.com/#/app/srcpublic/page/01-13-2025
https://www.construction-physics.com/p/why-skyscrapers-became-glass-boxes
plain glass box skyscrapers were, in fact, more cost-effective for developers. it’s not all about architectural tastes. architects in real life are very far from all-powerful.
in fact, I really think people should stop writing books/movies/etc about auteur architects; it only encourages more young people to go into architecture and become unemployed. I’m looking at you, Francis Ford Coppola
https://www.betonit.ai/p/the-typical-man-disgusts-the-typical
pointing in the right direction, but overstated/inflammatory. women don’t go around being “disgusted” by every man they interact with socially.
rather, most women find the idea of having sex with a randomly selected unfamiliar man disgusting, even if there’s nothing particularly the matter with him. typical straight women are cautious/selective about sex and fairly slow to warm up sexually to new people. not much “lust at first sight.”
but yeah, getting rejected when you ask women out does not in fact mean you are inadequate or unattractive! getting rejections in dating is normal, just like every author gets rejected manuscripts and every job applicant gets rejected from jobs. the average man gets a lot of “no”s and at least one “yes”, and eventually marries a “yes.”
also i share Bryan Caplan’s view that women shouldn’t be offended by being asked out by someone they aren’t interested in. sure, persistent harassment can be a problem, but a simple question isn’t.
https://nabeelqu.substack.com/p/principles Nabeel Qureshi
I agree with most of this, but “you don’t need much sleep” is very individual. some of us very much need plenty of sleep and our lives improve dramatically when we face that fact.
things I googled while reading about Venice
links 1/10/2025: https://roamresearch.com/#/app/srcpublic/page/01-10-2025
https://en.m.wikipedia.org/wiki/History_of_Romania
https://en.m.wikipedia.org/wiki/Cucuteni%E2%80%93Trypillia_culture
https://en.m.wikipedia.org/wiki/Gheorghe_Gheorghiu-Dej
https://en.m.wikipedia.org/wiki/Nicolae_Ceau%C8%99escu
https://en.m.wikipedia.org/wiki/Romanian_rural_systematization_program
claims that the Securitate continued to exist de facto, even after the revolution that ended the Ceausescu dictatorship.
“This influence distorts democracy, says Alina Mungiu-Pippidi, a Romanian civic activist and a leading scholar on corruption. “There’s long been ample evidence of a ‘deep state’ in Romania with roots in the former security services,” she maintains.”
“The Securitate was declared defunct and, with no admission of previous crimes – including those committed in the revolution – or internal vetting, was chopped into nine separate services that corresponded to the Securitate’s organizational substructure. The new services were staffed and directed by virtually the same people as the old Securitate.”
“During the 1990s, the Securitate’s successors, the domestic SRI, the Foreign Intelligence Service, SIE, and others, exploited their wide-ranging resources and security monopoly to become oligarchs and form cartels in the post-Communist economy. So vast was their intelligence that they could blackmail compromised politicians, media professionals and judges. In post-communist Romania, just about everyone had something to hide.”
what went wrong at Zymergen?
https://www.sec.gov/files/litigation/admin/2024/33-11303.pdf
they failed to manufacture their first product to customers’ specifications, lost the orders, overstated the market for it (which the SEC sued them for), and admitted they had no revenue & would not be profitable for the foreseeable future shortly after their IPO.
ultimately the issue may be that they specialized in only one piece of the manufacturing process, strain optimization. they outsourced downstream processing and manufacturing scale-up to third party contractors, and those are often places where hiccups arise for new biomanufactured products.
“not enough controlled burns” and “climate change” are the causes of forest fires.
what’s going on in LA right now is a chaparral fire. controlled burns aren’t appropriate for chaparral. these are “wind-driven” fires spread by the Santa Ana winds, which are unusually strong this year, but the wind severity isn’t driven by climate change.
the human-preventable cause of these fires isn’t climate change or Smokey the Bear, but too many human-caused fires, in particular due to power lines, since PG&E has not adequately maintained the lines in California as population has risen and the power grid has grown.
https://www.npr.org/2006/12/15/6630791/name-calling-in-michael-crichtons-next amusing story about Michael Crichton putting an unflattering portrait of a journalist who criticized him in his (fiction) book.
https://topos.institute/work/ Topos Institute builds collaborative modeling tools based on category theory, useful for formalizing mathematics, system dynamics, epidemiology, etc.
try their tool CatColab here: https://catcolab.org/analysis/f79f1894-601f-49cf-9da1-00a2ebcc0792
textbook about these concepts: http://davidjaz.com/Papers/DynamicalBook.pdf
https://github.com/PatrickMassot/leanblueprint the Blueprint library allows you to visualize the dependency graph of a proof in Lean
in retrospect, 6 years later:
wow, I was way too bearish about the “mundane” economic/practical impact of AI.
“AI boosters”, whatever their incentives, were straightforwardly directionally correct in 2019 that AI was drastically “underrated” and had tons of room to grow. Maybe “AGI” was the wrong way of describing it. Certainly, some people seem to be in an awful hurry to round down human capacities for thought to things machines can already do, and they make bad arguments along the way. But at the crudest level, yeah, “AI is more important than you think, let me use whatever hyperbolic words will get that into your thick noggin” was correct in 2019.
also the public figures I named can no longer be characterized as only “saying true things.” Polarization is a hell of a drug.
links 1/8/2025: https://roamresearch.com/#/app/srcpublic/page/01-08-2025
https://neuromatch.io/ company courses & networking opportunities in fields like neuroscience, climate science, AI, etc; matching collaborators
https://birdflurisk.com/ H5N1 risk dashboard
https://www.thetransmitter.org/neuroai/solving-intelligence-requires-new-research-and-funding-models/ the case for big funding of brain mapping & modeling, by David A. Markowitz
“The recent mapping of an entire adult fruit fly brain—a watershed achievement that made headlines worldwide—offers a glimpse of what’s possible. But this breakthrough almost didn’t happen. It required the serendipitous alignment of support from three non-traditional funders: Scientists at the Howard Hughes Medical Institute’s Janelia Research Campus imaged the complete fly brain; the Intelligence Advanced Research Projects Activity drove the development of tools for scalable neural-circuit mapping through its MICrONS program; and the National Institutes of Health BRAIN Initiative provided sustained support for data analysis.”
it might cost $1B to fully map & model the brain. ARPA-style and FRO-style research orgs are essential.
https://www.biorxiv.org/content/10.1101/2024.02.13.580186v4.full if you just use neural nets to model the output of C. elegans’ 302 neurons, bigger networks are better.
continuous-time RNNs scale the best—even better than transformers.
https://www.biorxiv.org/content/10.1101/2024.09.22.614271v2.full
optogenetically perturb each neuron in C. elegans and see what happens to neural output in all 302 neurons
fit this as a simple multivariate dynamical system—each neuron’s output at time t is a linear function of all the neurons’ output at time t-1, plus a linear function of the neurons’ history of optogenetic stimulation, plus error.
compare to a simpler, connectome-constrained model, where each neuron’s output is only a function of its presynaptic input neurons (and direct optogenetic stimulation). this is actually a good approximation!
in fact, it’s better than a fully-connected model, OR a “shuffled-connectome” model based on a made-up C. elegans connectome with similar topological properties to the real one. the true connectome matters.
if you train a connectome model without one neuron, it predicts something about what activity “should” be there. correlation at 0.30 with the real one, much higher than a “fake” connectome model’s correlation with reality.
model weights don’t reflect synapse counts, though. “multi-hop” trajectories have significant influence on correlations (i.e. neuron A and neuron B’s activity may still be highly correlated even if A and B are more than one step away on the “connectome graph”).
it would be shocking if connectomes didn’t matter, so in a sense this is not a surprising set of results; but this is a first example of collecting data with the optogenetic perturbation method, which is a major step towards true neural simulations.
a simulation should be able to predict what every neuron would do under various circumstances.
gathering data on the worms in varying behavioral/environmental contexts could approximate this, but manipulating each neuron one at a time gives a much more thorough picture of the input-output relationships of the nervous system.
https://arxiv.org/abs/2308.06578
roadmap document for full reverse-engineering simulation of the C. elegans nervous system, by Adam Marblestone and many others
includes perturbation!
https://en.m.wikipedia.org/wiki/Asha Zoroastrian concept of “truth” or “right”
https://www.complexsystemspodcast.com/episodes/outside-view-yatharth/ Patrick McKenzie interviewed by Yatharth
https://mad.science.blog/2020/07/07/desummation/ theory that NMDA receptor antagonists’ hallucinogenic and “psychotomimetic” effects come from inhibition of memory.
NMDA receptors (for glutamate) are important in long-term potentiation, in which “a neuron becomes highly sensitive to excitatory transmission for days or weeks” following recurrent stimulation.
long-term potentiation is important in associative memory. it “strengthens” a neural connection that has been sufficiently strongly/repeatedly made.
NMDAr inhibitor drugs reduce this effect.
so does schizophrenia
Neurons have “summation” effects.
“spatial summation”—if two neurons stimulate a third, the effect is stronger than if only one did.
“temporal summation”—repeated stimulation has a stronger effect than a one-off.
it is a form of coincidence detection; multiple “simultaneous” events are treated as a bigger deal, less likely to be flukes or errors.
the NMDAr inhibitor dexmethorphan inhibits this effect.
so does schizophrenia.
normal subjects have “prepulse inhibition”, aka the reaction to a loud startling noise is less intense if preceded by a small pulse sound.
this is a form of temporal summation; signals generated by the prepulse accumulate and prepare the brain for further sound, preventing the startle response.
schizophrenics don’t have this; they get startled both ways, indicating (?) that the prepulse sound doesn’t “accumulate” properly.
“desummation” or “cognitive atomization” is like a failure to anticipate; new stimuli are fresh, not expected.
this coincides with the subjective effects of NMDAr inhibitors: at low doses there is “increasing perceptual acuity for things usually unnoticed” and at high doses there is “inability to notice a lot of previously learned meaning”, “inability to recognize familiar stimuli”, added “noise”, and loss of “definition and meaning.”
visual agnosias are a common reported ketamine effect
NMDAr inhibition causes amnesia effects in patients
patients on low-dose ketamine do not retain things they learned while on the drug
“PsychonautWiki lists ‘memory suppression’ as a distinct effect from amnesia as a side effect of NMDAr antagonists. In the description of this effect it notes that short-term memory is suppressed much earlier than long-term memory. At very high doses, the Wiki suggests that one may even forget who they are, where they live, or even a failure to remember what humans are”
impaired short-term memory relates to sensations of unfamiliarity and dissociation (if you can’t remember stuff, you don’t know what it is)
https://en.wikipedia.org/wiki/Autostereogram
excellent explanation of how “magic eye” pictures work; i can now see them for the first time in my life!
https://smoothbrains.net/posts/2023-08-01-ketamine.html personal experience with ketamine
effects:
visual agnosias, “2D vision”, loss of “egocentric coordinates” in spatial perception, “expanded” awareness
music perception is different—complex music is confusing, repetitive drones are hypnotic
tactile feelings are pleasurable, muscle tension is strongly reduced
aversions are dampened, in particular by making them slower—from the usual “100 ms” to “500 ms”.
celebrity cases of ketamine deaths:
https://www.nbcnews.com/news/us-news/arrests-made-connection-accidental-death-actor-matthew-perry-rcna166676 Matthew Perry, found face down in a heated pool
https://en.wikipedia.org/wiki/Gary_Frisch Gary Frisch, found dead beneath an 8th floor window with ketamine in his blood and liver
notice that these are plausibly fatal accidents or suicides while under the influence. this is a potential danger of recreational use that won’t show up in stats about medical use or experimental administration to animals.
https://www.quantamagazine.org/elliptic-curve-murmurations-found-with-ai-take-flight-20240305/ wavy “murmurations” found in elliptic curves with machine-learning algorithms
https://www.quantamagazine.org/behold-modular-forms-the-fifth-fundamental-operation-of-math-20230921/ modular forms!
https://trevorklee.substack.com/p/regulators-almost-killed-biotech
“from the very first announcement of successes in cloning DNA (the foundational technology of recombinant insulin) in 1973, the first reaction of the National Academy of Sciences and the National Institute of Medicine was to try to stop research into it and, simultaneously, to prevent any patenting of it. The National Institute of Health followed shortly thereafter with their own restrictions, as did a number of cities, including Cambridge and Berkeley, the towns where the universities most likely to do the research were located”
Genentech, a tiny garage startup, was able to escape those restrictions.
https://polypharmacy.substack.com/p/whats-alprazolams-deal alprazolam (Xanax) is reputed to have a fast onset and fast diminution of its anti-anxiety effect, compared to lorazepam (Ativan), making it more abusable and less useful for anxiety disorders. but why? all the theories seem wrong!
also, diazepam (Valium) has exactly the same fast onset/fast diminution, but doctors don’t seem to worry about it the same way!
it looks like the difference isn’t about half-life, elimination, or the blood-brain barrier, but something about ligand-receptor binding.
links 1/7/25: https://roamresearch.com/#/app/srcpublic/page/01-07-2025
https://asteriskmag.com/issues/08/the-making-of-community-notes the team behind Community Notes describes how they do it
it’s people who were working on “Birdwatch” before Musk bought Twitter, and they use algorithms derived from PageRank.
these guys are not, even a little bit, “based.” Yet the Twitter userbase loves Community Notes!
if you have a capable team that firmly believes in “fairness”, in auditable, open, participatory processes that don’t put a top-down thumb on the scale on controversial issues, and they get to actually use the neutral algorithm instead of being pressured to make exceptions, you get solid results and community trust!
https://pmc.ncbi.nlm.nih.gov/articles/PMC7132523/ 90% of bronchitis is viral. doctors are cautioned not to give antibiotics for it, even for long-lasting coughs.
https://www.aafp.org/pubs/afp/issues/2010/1201/p1345.html no really. bronchitis symptoms typically last three weeks. it’s still not bacterial. antibiotics do not work.
https://harmonic.fun/ AI for formal mathematical reasoning
https://drive.google.com/file/d/1-FFa6nMVg18m1zPtoAQrFalwpx2YaGK4/view Tim Gowers’ research manifesto on AI for mathematics
it’s not machine-learning based; it’s a version of GOFAI that’s formalizing the types of “tactics” or “moves” that a human goes through in a proof, trying to get the formalism right such that a computer proceduralization only has a modest, human-like amount of trial-and-error & backtracking rather than vast amounts of brute-force search.
https://www.businesswire.com/news/home/20240514445088/en/Multiply-Labs-and-Retro-Biosciences-Announce-an-85-Million-Partnership-to-Advance-Cell-Therapy-Manufacturing-for-Age-Related-Diseases cell therapy manufacturing automation robots from Multiply Labs, used at Retro Bio.
https://gopiandcode.uk/pdfs/sisyphus-pldi23.pdf
“proof repair” is the problem of updating formally verified software; if you have a library of provably correct code, and you make any kind of software updates to the library, now you also have to update the proofs such that it’s still verifiably correct (assuming the update didn’t break anything.) Progress towards automating this.
https://openreview.net/pdf?id=RtTNbJthjV the Karp Dataset, 90 reduction-based NP-hardness proofs (in natural language) for training LLMs to write proofs.
https://web.stanford.edu/~jlmcc/ Jay McClelland, co-inventor (with Rumelhart of backprop fame) of the Parallel Distributed Processing theory of cognition, has been doing a lot with LLMs lately
https://web.stanford.edu/~jlmcc/papers/ChanEtAl22DataDistPropsDriveInContextLearning.pdf in-context learning works better when the training data is bursty and has lots of sparsely presented elements
https://web.stanford.edu/~jlmcc/papers/DasguptaLampinenEtAl22LMsShowHumanLikeContentEffectsInReasoning.pdf LLMs, like humans, show belief-bias effects on Wason and syllogism tests
https://web.stanford.edu/~jlmcc/papers/McClelland22CapturingAdvCogAbilitiesWithDeepNets humans learn human-made formal systems (like mathematics, computer science, logic) in order to solve certain kinds of difficult problems. perhaps AIs should also “go to school”, being trained on math problems, in particular with diagrams as well as text. also, goal-directed motivation may require fundamentally different architecture from the usual LLM transformer setup.
https://web.stanford.edu/~jlmcc/papers/NamEtAl22LearningToReasonWRelationalAbstractions.pdf an empirical example of “taking the AI to school”—fine-tuning on a McClelland-designed dataset intended to teach “relational abstractions” makes models perform better at math problems
https://web.stanford.edu/~jlmcc/papers/NamMcC21RapidLearningGeneralizationInHumansArxiv.pdf human mTurkers are better at abstract problem-solving tasks if they’ve taken high school algebra and geometry (no effect for other educational variables). they split pretty bimodally into people who learn a strategy and people who guess at random. this points towards “basic math education teaches systematic thought.” also, small RNNs generalize much worse than humans, but who cares.
https://en.wikipedia.org/wiki/Answer_set_programming ASP is used for difficult search & combinatorics or optimization problems. I’m struggling to understand whether it is in wide industrial use or if it’s more of a research specialization.
links 1/6/25: https://roamresearch.com/#/app/srcpublic/page/01-06-2025
https://en.m.wikipedia.org/wiki/Ralph_Fiennes descended from Norman nobility
https://meltingasphalt.com/wealth-the-toxic-byproduct/ nice essay on how (money) earnings generally represent your benefit to others while consumption spending is a cost to others
https://mises.org/mises-daily/defending-miser not only is the “miser” who invests productively benefiting others, but so is the “hoarder” who takes money out of circulation to put under a mattress—hoarders lower prices.
ehhh, shouldn’t lowering prices and raising prices be seen as equally neutral in real terms?
https://timmermanreport.com/2025/01/ai-needs-natural-language-to-give-structure-to-biology/ I’m sorry, you just named three discoveries in the latest table of contents of Science or Nature, and complained that biological “foundation models” can’t come up with them, and that instead you need LLMs? what are you even thinking???
you are in a hurry to replace human thought at the highest levels, when there’s a tremendous amount left to be done in developing AIs to replace tedious grunt work.
why??? why do you want to put the PI out of a job? There aren’t that many PIs. why are you starting here? it’s so backwards.
what they’re actually doing at FutureHouse is pretty cool: “write an accurate and cited Wikipedia-style article for nearly every protein-coding gene in the human genome.” I think accelerating lit review is a useful function of LLMs. there’s a lot of information out there! using automation to synthesize it is a win!
but my god, man, why are you framing this as “I really want the machine to do all the thinking for me” rather than “I want everyone to have a research assistant in their pocket so they can more efficiently explore hypotheses with the context of the whole scientific literature.” do you not like thinking and learning? do you wish you could quit or something???
https://blog.samaltman.com/reflections
“We believe that, in 2025, we may see the first AI agents “join the workforce” and materially change the output of companies.” Ok, this is a real prediction (that we can grade him on later) and it certainly seems technologically possible today—it remains to be seen whether it’ll be organizationally possible, useful in the first contexts where it’s used, or whether it’ll cause catastrophic errors.
if i had to guess, i think commercial AI agents will be launched by OpenAI, and their economic impact will be ambiguous, by end of 2025.
https://parahumans.wordpress.com/ trying to read this, again, for the fourth? time.
not sure why I keep bouncing off this. it’s recommended by friends, I’m excited by the promise of interesting conceptual things being done with the superhero-genre system, but my god there are many chapters of “boring”-to-me stuff (action scenes, description of city politics & class dynamics that doesn’t feel true to life, etc)
ketamine overdose risks:
https://pmc.ncbi.nlm.nih.gov/articles/PMC3168228/
typical recreational doses are 200-300 mg (orally) or 50-100mg IV, roughly 50x the effective IV dose for general anaesthesia.
https://www.psychiatrist.com/jcp/oral-ketamine-for-depression-practical-considerations/
similarly, oral ketamine for depression is given in doses ranging from 25mg to 300 mg
https://www.ncbi.nlm.nih.gov/books/NBK541087/ this StatPearls claims that the median lethal dose for a 70kg human is 678 mg, but references a study not cited
this is the Gable et al study mentioned in the StatPearls article
it mentions that the rodent LD50 for ketamine is 600 mg/kg, and divides by 10 (which is apparently a standard rule of thumb for converting between species) to get 4.2 g for a 70 kg human
they say a typical recreational dose is 175 mg, and so get a “safety ratio” (lethal dose/effective dose) of 25, safer than alcohol.
StatPearls may have misquoted this paper...I’m not sure where they’re getting their 678 mg number.
if the lethal dose really is only about twice the effective dose, as in the StatPearls article, that’s a very tight window.
for context:
the median lethal dose (LD50) of ethanol is the equivalent of about 25-40 standard drinks
the LD50 of aspirin is about 43 aspirin tablets
the LD50 of amphetamine is between 20-200 times a typical prescription ADHD dose
https://www.youtube.com/watch?v=ouJrOQfkeok how a Kite cell therapy (autologous) is manufactured (they ship your blood sample to a facility, and process/expand your cells in 5 days).
suicide note of a DeepMind researcher who took ketamine for depression, developed psychosis and then very severe depression unlike anything he’d experienced before; “the emergency alarm continues to strike blind panic and fear into my mind every second” for two years with no change.
https://marginalrevolution.com/marginalrevolution/2025/01/simple-points-on-immigration.html Tyler Cowen remains pro-high-skilled immigration.
https://impact-ops.notion.site/11c061ba6c7880b38073e8ddfb4f1db0?v=11c061ba6c788137aa62000c8aa8918b an instruction manual for setting up a nonprofit.
having started a nonprofit myself and screwed some of this up, i want to endorse taking a lot of care with legal compliance and administrative processes.
an “ordinary” middle-class person, in their personal life, usually doesn’t have to worry too much about accidentally breaking laws. a founder (of a nonprofit or small business) can ABSOLUTELY do illegal stuff by accident and get burned for it.
it’s also very easy to accidentally be the kind of terrible boss you hear horror stories about...just screw up payroll!
https://www.nature.com/articles/s41467-024-53922-x
neuroblastoma and embryonic kidney cell lines exhibit the “massed-spaced” effect discovered in neurons by Hermann Ebbinghaus where a repeated spaced stimulus has a stronger response than a “massed” (clumped) stimulus, almost as though “learning” the difference between an event that recurs and a (potentially erroneous) one-off.
in non-neural cells, instead of looking at neuron firing we’re looking at a particular cell signaling pathway engineered to carry a luciferase gene that glows when the CRE gene is transcribed.
The chemical TPA directly activates protein kinase C (PKC)
the chemical forskolin activated protein kinase A (PKA) via raising cAMP levels, which activate PKA
PKA and PKC both phosphorylate the transcription factor protein CREB1
CREB1 causes lots of genes, including our reported gene CRE, to be transcribed more
this pathway was selected to be similar to the way serotonin activates a signaling cascade during memory via PKA and PKC
people are sharing around popular articles claiming that this proves kidney cells have memory, but it really doesn’t.
it proves that the mechanisms of neuronal memory have shared properties with cell signaling mechanisms present throughout the body, which is what you’d expect; neurons necessarily evolved as a variation on some other kind of cell.
now, understanding exactly how this works is interesting! but no, it doesn’t mean you think with your kidneys.
links 1/3/2025: https://roamresearch.com/#/app/srcpublic/page/01-03-2025
https://thisgenomiclife.substack.com/p/this-weeks-finds-in-genomics-and-4bd “More than half of [a new dataset of 2.35 million candidate regulatory DNA] elements are not close to the transcription start sites of genes, meaning they act at a distance to control gene expression”
https://www.ams.org/notices/202501/rnoti-p6.pdf Terence Tao on machine-assisted proof
https://www.cato.org/blog/national-security-hoax Joe Biden blocked a Japanese company’s acquisition of U.S. Steel on “national security” grounds...even though Japan is a US ally and this would be an investment into America-based steelmaking facilities
https://marginalrevolution.com/marginalrevolution/2024/12/the-future-of-the-scientist-in-a-world-with-advanced-ai.html Tyler Cowen thinks that all the hypothesis generation in science will be done by AI and humans will only gather data (via lab work or confidentiality/data access negotiations.). This is...pretty backwards from what I expect the best uses of AI vs. creative humans are. I’m much more interested in AIs for lab automation and faster idea-generation/lit-review loops.
https://kyunghyuncho.me/i-sensed-anxiety-and-frustration-at-neurips24/ interesting perspective. as we’re entering the “productization” phase of AI, says the article, the days of PhDs getting fat industry salaries to do research are ending.
this...does not match my experience, though maybe I was seeing a different side of the elephant. In the mid-2010s, ML people knew that “deep learning” was the best at the benchmarks, but it was fiendishly hard to get business results from it in most contexts, so startups would typically aspire to use it and then...quietly not. in my corner of startupland, machine learning PhDs were definitely not being paid to do freeform research; these were the days of “feature engineering is king” and “data munging” and “let’s just use a logistic regression, it works better”
article’s probably right that it’s not a great time to be in non-LLM fields of ML, though there are exceptions
https://www.nature.com/articles/s41562-021-01286-3.epdf?sharing_token=FmzVoOJu5aSIdLKxYEmbs9RgN0jAjWel9jnR3ZoTv0Mqd2apSCbVNCoLnGZzPpZgrhSB6F3n6W7UifVpls202s_RwJ4kTaNjSgIQzVnS6hGeSq41cU3dWAYR23ygDu7de3fh4_6F6XJA2pD3xNO3c8hyjtnc_kr6GV5YDuwIdFQ%3D it’s very easy for grant funding schemes to be net negative in societal value.
“if the cumulative work that goes into an average grant application adds up to considerably more than a couple of days, these grant schemes draw more resources from the scientific community than they add.” in reference to two (actually existing) grant programs awarding €50,000 and €30,000 with success rates of 5% and 2.5%, respectively.
in other words: if grants are small, selective, and time-consuming, they’re using up more scientist-hours on grant applications than they are funding scientist-hours of research.
https://poetryarchive.org/poem/fiddler-dooney/ one of my favorites
https://www.feraleyes.xyz/p/god-written-by-a-girl personal essay, I don’t totally understand but it’s heartfelt
https://viaseparations.com/wp-content/uploads/2024/12/Via-Separations-White-Paper-Dec.-2024.pdf Via Separations produces membrane separations for industry that replaces (combustion-powered) evaporators, using 75% less energy without yield loss and with favorable economics without subsidies.
https://mathstodon.xyz/@tao/113721192051328193 Terence Tao on getting his papers declined; it happens to him about once a year. rejections are not unusual in math journals and even good mathematicians get them.
links 1/31/25: https://roamresearch.com/#/app/srcpublic/page/01-31-2025
https://hareesh.org/blog/2016/2/5/the-real-story-on-the-chakras the pop-culture chakra model comes from modern Theosophical books, not ancient Indian scripture.
https://solarfoods.com/ solar-powered industrial-fermentation food? they have a microbial protein called Solein. it’s made from bacteria, probably Cupriavidus necator.
https://pmc.ncbi.nlm.nih.gov/articles/PMC8719805/
is this competitive with whey, legume, or fungal protein, from a cost or sustainability standpoint? haven’t checked yet
https://vetmed.illinois.edu/i-tick/2019/08/09/iceman-lyme-mummy-tattle-the-tick-blog/Otzi the iceman had Lyme disease??
https://en.wikipedia.org/wiki/Xanthan_gum is made from bacteria
the inventor https://en.wikipedia.org/wiki/Allene_Jeanes, also pioneered mass production of dextran (from industrial fermentation) which saved many lives as an ingredient in an emergency blood substitute in the Korean and Vietnam wars
https://emalliaraki.com/