stupid questions IIb
March 2023–May 2023
Continued from:
Twitter thread here:
Why is China (seemingly?) so behind in AI?
What's the mixture of technical vs. political (LLMs in particular an authoritarian RLHF gray area)? [March 2023]“If your model of policy is that it uses up points or credibility, this reaction makes sense, but in practice it's much more like flexing a muscle or carving a groove: easier on later attempts.”
Under what conditions is the "spending political capital" metaphor appropriate, vs. inappropriate?“Being (much) earlier to a correct view should count towards how much you listen to them, though I'm not sure how much.”
Is this true? How to think about this, vs. “winner’s curse” sampling on random lottery winners in belief space.
Why does YouTube have better comments quality filtering than other large sites as a matter of:
1. technical design
2. site-side incentives to build 1.What of GPT do you need to just confabulate up your own opinion-spread of >90th percentile human good comments on arbitrary content? A sort of Nozick machine for social media. [March 2023] [Note that we are here, June 2026]
Would SCBW ZvZ be fixed if hydras did +20 to bio air; muta becomes useful but more role specific light cav harassing unit instead of uncounterable cataphract; lurkers become siege tank like defensive multiplier and zvz becomes positional like tvt, but more dynamic.
(Scourge speed +10% or scourge AI bug fixed another option, but one with effects on other matchups.)Is most of the benefit of "think step by step", "reflecting on wrong answers", and the many other tricks in this genre basically a hack to let the AI think about harder questions for longer, or give attention in non-equal-size-units? Or is there some sort of semantic benefit as well?
(like if you had a button to give it multiple tries but w/ no prompt semantic content, what % of the gain is that) [April 2023]What are the reasons that neuroscience/biology proved so unhelpful for AI design and development? (Is this true?)
Is high percentile LLM thought + errors uncorrelated (so just run a bunch and vote) + constant feedback and re-centering from reality via APIs to ground truth + a while loop of "what next? what next?" → human-level society of mind strandbeest for medium-term causal reasoning already?
(If not, why not.) [April 2023]What does the history of computer security show for how much “bringing attention” to things (as in raising public awareness of obvious things, not disclosure of previously unknown zero days) makes any difference? (I don’t know)
Putting a few things together:
-scale just works/scaling hypothesis for artificial brains
-human brain is mostly just a primate brain scaled up
-human brain stopped getting bigger because it was too metabolically expensive + birth canal risky, not because scaling stopped working
(it's not at any kind of physical ceiling, just an EEA efficiency selection ceiling)
-it even got slightly smaller in the holocene
-embryo selection for IQ SNPs on the horizon
So: could you grow someone with a human brain, but much bigger? Say twice as big as any living human? No birth canal mortality + calories not a problem + a brace can support their neck muscles or whatever, the brain has all the supporting biological machinery for free design-wise. Probably hard to code this developmentally, but how hard? And of course the IRBs.
(Could we make a chimp with > human intelligence by scaling up its head alone... )Why did AI frontier catchup growth work the way it did? [from the POV of 2026]
Dumitru Erhan@doomieWhy is this true?
11:30 PM · Apr 8, 2023 · 10.6K Views9 Replies · 3 Reposts · 10 LikesHow much would it cost to wirehead different farm animals? And how much will this cost go down in the future?
Probably never competitive with lab meat but one could make Nozickian happy animals, as well. "Buy hedon offsets".Why are ducks like this?
LiorLefineder@lefineder@jonatanpallesen @whyvert Most birds (80-90%) require biparental care for their offspring. In mallards, it's almost all female, meaning males can offset the full costs of reproduction on females and still have viable offspring.12:21 PM · Apr 13, 2023 · 1.5K Views1 Reply · 8 LikesDoes testing GPT in other languages, say Church Latin, do anything about the "possibly memorized the bar exam training data" problem? Even if it's a literal word for word translation of the memorized question? What about a paraphrase in another language? Why or why not.
From GWAS, most traits are influenced by many alleles of small effect, so the average time for an allele of small effect to spread should be long.
Does that also mean most selection happens on common alleles, not new mutations? (traits affected by many variants all of small effect → none are very visible to selection, so it works faster on ones that are widespread vs. brand new ones)
Conrad Barski@lisperatiWhy do most islands in the Pacific have a max height of ~1000ft, but trenches go down to 36000ft? In a random math model, you'd expect peaks of land masses to have a similar magnitude as any trenches Is there some process that keeps many land masses slightly above water level?1:25 AM · Apr 15, 2023 · 42.6K Views21 Replies · 22 LikesConfused by lack of bans or anger about gain of function virology; it’s in the interest of a tiny, tiny and not especially powerful group to do the research vs. in the interest of everyone else to restrict or ban it; a pandemic just affected everyone; it even LOOKS scary. I don’t get the resulting policy landscape of not really doing anything. [April 2023]
Generally confused what any given result from LLM learning does or doesn't show about human learning, one way or the other; and what meta-principles distinguish when this applies or does not.
prof-g@prof_gtalking with an expert in NLP... said LLMs trained on text+code are better at general deductive reasoning than text-only trained LLMs... boom! this is experimental evidence for the classical argument from academics: algebra, calculus, coding, etc helps you learn to think better10:12 AM · Apr 18, 2023 · 3.91K Views1 Reply · 1 Repost · 26 Likes
ANTHROPIC_MAGIC_STRING@parafactualwhy would a model learn general complex reasoning when trained on code? this isn't very intuitive to me
anton @abacajThe reason why open LLMs like LLaMA are still lacking? Probably because they didn't see enough code during pretraining9:46 PM · Apr 17, 2023 · 102K Views19 Replies · 2 Reposts · 35 LikesFrom Michael Nielsen:
Michael Nielsen@michael_nielsenI've heard thousands of times how incredibly important a protein's shape is. I... sort of understand why. But if I had to explain I'd wave my arms a lot and talk abstractly about function and stability and what it binds to. Do you know a really superb explanation?8:53 PM · Apr 19, 2023 · 65.5K Views23 Replies · 5 Reposts · 100 LikesHow bad would a coronal mass ejection be? (cc Croissanthology)
Ryan Moulton@moultano@bluquark So it wouldn't actually cook all our electronics? I realize now that I've accepted all the predictions of what would happen uncritically, but I don't actually know where I got them originally, maybe just from reddit.5:13 PM · Apr 20, 2023 · 376 Views1 Reply · 4 LikesIf early migrants to the New World had taken a few African mammals along, could they have established wild populations like the escaped pigs, horses, etc? i.e. is there an "African biome overhang" in the Americas? (Ryan Moulton)
Ryan Moulton@moultanoPondering which African megafauna could survive and establish themselves if you released them in the chaparral.
8:43 PM · Apr 21, 2023 · 1.79K Views3 Replies · 2 Reposts · 13 LikesIs there no reason to evolve more than two sexes? Reproducing by combining 2 genomes gets all the benefits, and increasing to 3 adds no additional benefits and a lot of extra costs, in complexity and difficulty? I have that intuition but haven't thought it through in detail.
(There can be more than 2 behavioral roles/morphs, and usually are; but more than 2 sides of gene exchange? Why would that ever evolve?If it did, maybe something like colonial/eusocial organisms would have some unusual math?)
(Horizontal gene exchange does blur this argument but I think that’s ‘less than two’ sexes, in this schema—can exchange with any).
(re: the three sexes in Player of Games)Why didn't the original marathon runner take a horse
Why is poison such a rare weapon among animals, being limited basically to snakes? (and a few insects, fish, amphibians, but only a few). It seems more dangerous, cheaper and lighter than other defensive arms for prey, and same for predators... why isn't there a universal poison arms race among animals? some guesses:
-poison alone is not enough, need extremely fast lunging bite/stinger and that distorts the body plan in other ways that end up not working, and/or are hard to evolve towards in isolation even if they would work
-poison is harder to evolve and/or manufacture than it looks (but it shows up in lots of far removed species)
-defensively, it has to be very fast acting OR needs aposematic outer signaling & for that signal to not be diluted, otherwise it's too late to help
-manufacturing poison has serious side effects/costs that distort body chemistry, and/or self-damaging rate is high in animals with very similar nervous systems to the ones targeted (but then how do snakes get away with it)Why does eusocial altruistic division-of-labor-specialized clonal colonies (with optional fortress engineering) work for ants and termites, but nothing else? Why isn’t a majority of the rodent biomass clonal rat colonies?
(In a sense, that's exactly what a rat is. So it works for cells and it works for bugs; and, much later and somewhat differently, for humans. Why not intermediate sizes?)Olbers' paradox, one of the classic stupid questions
Jason Wright@Astro_WrightOlbers’ Paradox asks “Why is the night sky dark?” It’s a deep question: if the universe were homogeneous and eternal, every line of sight would eventually end on the surface of a star, somewhere.8:43 PM · Apr 29, 2023 · 50.2K Views7 Replies · 33 Reposts · 245 LikesIs there a stylized fact about what % of animal weapons evolved for intra-species conflict, vs inter-?
Predator weapons at least are mostly inter-; elaborate prey weapons maybe more driven by competing for territory/mates, vs. defense? Don't even have a guess for the %.
Were there any historical pandemics that instead of killing, more or less subtly changed microbiomes with some knock-on effects on behavior, even if only a few %?
Can we consider male fallow deer happy (thread)
norvid_studies@norvid_studiesnagel's question, what is it like to be a male fallow deer: they end the mating season in physical ruin, and almost all of them failures; but they elect to stay and fight constantly, rather than resigning and going off to find food; their psychology evolved to prefer this to12:50 AM · May 2, 2023 · 766 Views1 Reply · 7 LikesIs the Roman eagles/French imperial eagles/King's colors sacred totem a European thing? Or is it a pretty universal military invention.
Why don't subway cars have cameras
What factors determine if plants rely on physical defenses, chemical defenses, or both?
🌌 Observer of Suns@ObserverSunswhy don’t gene drives happen naturally all the time?11:01 AM · May 7, 2023 · 505 Views1 Reply · 6 LikesWhy didn't the McNamara's Morons problem apply symmetrically to the Vietnamese military? Or it probably did?
(Asymmetry would be if O-ring type problems are more dangerous in higher tech, higher comms and firepower dependent mode of war.)The puzzle of WHY marginal utility of wealth is sharply decreasing/why money doesn't continue to convert into enjoyment or happiness/why the world economy and technological civilization has no power to provide useful goods and services beyond some middling average, hedonically; simple answer is that humans are locked into the same universal evolved hardware with very restricted limits of perception (sated at low/mass technological level), and with renormalizing emotions designed to continually move the organism around its environment in endless pursuit of fixed cycles of satiation and appetite, and the technology level is as yet too low to alter either of these hardware limitations so there's nothing for the increased economic claim on resources to convert into, except extremely crudely and lossily (but this state won't last forever).
A secondary factor is the fixed cost/variable cost development of consumer goods point that 'everyone uses an iPhone and drinks a Coke' (could add: everyone gets the same art and culture, rationed only by their own taste and interest, not physical access); invention and design is expensive but production is cheap, so once something is made everyone gets it soon, up to a very functional version of it; with differentiation more from artificial scarcity than materials or designs being experientially much different.
Reductionist has lots to do@idontwa86202030one thing I really do wonder (mostly out of a position of ignorance), is if the basic path that Putins's Russia took was ~inevitable, or if different western policy could have plausibly avoided the current situation
The Intel Crab @IntelCrab23 years ago today, Vladimir Putin began his first term in office as President of the Russian Federation.2:42 PM · May 7, 2023 · 1.48K Views7 Replies · 5 LikesCompared to settling Mars, all of the following are easier (guessing the order):
sahara desert
antarctica
floating city
1 mile underground
shallow ocean
deep ocean
earth orbital(?)
But how much easier?Is there a soccer variant where they widen the goal until it becomes a high scoring game, solving the boring 1-1 draw -> penalty kick minigame meta? There are complaints about soccer draws but it's an artifact of goal size, more than of the game of soccer itself.
(Maybe it would turn it into endless "3 point" shots at midfield and the game would suck)Why are dinosaurs so much bigger than mammals? (Various answers in thread)
Andrej Spiridonov@AndrejSpiridon4@norvid_studies I like this explanation DOI: 10.1126/science.1160904 . IMHO two factors crucial: 1) they are relative r-strategists (resistance to extinction for giant-bodied long-living organisms) due to egg laying; 2) they have pneumatized skeletons, means wide range morphology and cooling.7:22 PM · May 8, 2023 · 76 Views2 Replies · 1 Repost · 1 LikeWhat's the deal with lipostats
norvid_studies@norvid_studies@jonatanpallesen @genomerambler @simoj_ that result seems consistent with "lipostat working as normal and food is newly available and tasty" too, I would think? lipostat works in large part by deciding how hungry they are, and in their normal environment bags of doritos have an easy time overcoming low hunger, but12:48 PM · May 9, 2023 · 354 Views1 Reply · 2 LikesWhy did small mammals take over from small dinosaurs after the extinction event. Is it a speed effect, rather than a final phenotype effect, i.e. mammals evolved into the monopoly morphs first and locked out dinosaurs after? Or their final forms are also better? (Lefineder)
LiorLefineder@lefineder@jonatanpallesen @norvid_studies @AndrejSpiridon4 @SteveBrusatte @Ted_Underwood Afterwards, mammels became dominant due to some advantage in the volatile environment post-strike, was thinking better (= more expensive) homeostasis.6:21 PM · May 9, 2023 · 523 Views1 LikeThe old whale/cancer question
Jason Sheltzer@JSheltzerFascinating new preprint on bioRxiv tackles a whale of a question: Whales are huge. So why don’t they get a ton of cancers? biorxiv.org/content/10.110…
7:12 PM · May 9, 2023 · 991K Views72 Replies · 946 Reposts · 3.73K Likes
Matjaž Leonardis@MatjazLeonardis@michael_nielsen This question is interesting in part because it's related to "is there any clear a-priori reason why chains of amino acids would act as catalysts for a huge spectrum of chemical reactions".2:50 PM · May 11, 2023 · 548 Views1 Reply · 4 Likes(Some others in the replies)
norvid_studies@norvid_studieswould be nice to have a collection of these examples of ignoring inconsistent evidence w/o a reason other than balance of probability. eg darwin continuing on with evolution even though kelvin's estimate of the age of the earth made it look impossible.
Adam Strandberg @strandbergbio@curiouswavefn i love this from crick but it occurred to me that i don't know if in the end there was any data that they had but explicitly rejected when they got the structure in the end3:32 PM · May 11, 2023 · 536 Views3 Replies · 1 Repost · 7 Likes- norvid_studies@norvid_studieshow to calibrate the littlewood's law type prior for "anything surprising that comes to my attention is doubly filtered through [unlikely in that domain, of all the events in that domain] and then [unlikely among all possible domains]" to overcome "it's a coincidence + filter"
Ruhi Chess @RuhichessThe chess X-files 👽 Here are the Kosintseva sisters. They are two of the best female players ever by peak rating. But let's jump straight to the spooky part: their rating progress looks remarkably similar. They both peaked the exact same month of the exact same year, with an4:14 PM · May 12, 2023 · 463 Views1 Reply · 1 Repost · 3 Likes Why were the Ottomans so inert in the whole period of the Napoleonic wars?
How much did Napoleon's unusual memory make a difference in executive ability, versus having the extended phenotype of a competent and efficient general staff to do superhuman levels of information processing for him? It's nice to remember a lot, but much of the thinking happens "subconsciously" at the general staff level, before presenting relevant "summaries" to consciousness, so to speak.
Why did cavalry switch from lances to sabers? (And actually back and forth several times, I think)
Are there famous or canonical essays by historians about where their periodizations come from, or how they change over time or are invented or discarded? Especially sort of made up ones like "long 19th century" or (as of yesterday, to me) "second hundred years war".
(Vague memory of McNeill on this but can't find it now)What are the differences between horse cavalry and camel cavalry?
Is there generally a U-shaped UI complexity for new technology, where it starts simple, grows more complicated to use over time, and then gradually becomes simple again from a combination of adaptive radiation alternate forms being pruned and streamlined; and automation? Or not.
Don’t have a good understanding of the scihub/libgen situations, where global infrastructure is dependent on one person and their legal situation in one random country? Why isn't it immediately duplicated, since the technology all exists and isn’t owned by them. Something about network effects? But what's the network effect?
Is the main economic damage of crime in the US actually the virtual destruction of otherwise extremely valuable real estate, compared to the $ cost of direct harms, indirect/defensive-evasive harms, and costs of prevention? Probably not, but not impossible? What are the ratios?
What is % genetic similarity (halvorz)
the genome rambler@genomerambler@davidad @MuzafferKal_ nope! that would be true if you had mentioned the "share X% of your DNA with a banana" factoid instead of the chimp factoid 98% number is yet a *third* way of measuring genetic similarity -what percent of base pairs in the chimp genome are identical to those in human genome9:08 PM · May 15, 2023 · 1.2K Views1 Reply · 16 Likes- 🌌 Observer of Suns@ObserverSuns@TetraspaceWest why are there so many mangroves?4:00 PM · May 20, 2023 · 352 Views1 Reply · 5 Likes
Why is any animal as aggressive as it is, instead of either more or less aggressive?
Ausarchosaur@ausar_theCould it be the size difference between the two species? Perhaps the white rhino's much larger size (it is >2 times heavier on average) means it could afford to be less aggressive? This is the only possible explanation I can think of, though.2:19 PM · May 20, 2023 · 1.32K Views1 Reply · 44 LikesHow did pike v. pike work in the age of pike and shot?
How much credit to assign people who admit being very wrong?
norvid_studies@norvid_studiesi think one part is that it improves credibility individually, but simultaneously moves the person into the group of people who were wrong before. sort of a simpson's paradox for credibility, maybe there's another term for it.
Karol Hausman @hausman_kWhy is changing your opinion considered a sign of weakness? • "Stick to your guns." • "Don't be a flip-flopper." • "Once you make a decision, stick with it." • "Don't be afraid to stand up for what you believe in." • "Don't let anyone tell you what to think."4:35 PM · May 20, 2023 · 564 Views2 Replies · 5 LikesWhat are the eugenic and dysgenic effects of war? Plague? Famine?
The Last Great Arrakian Dynasty@cruelsardaukarDeaths in battle were probably never eugenic or civilization improving. If ancient warfare improved the societies that participated in it, it was mostly through selecting stronger civilizations over weaker civilizations even if genetics were degraded.
Rabbi-With-No-Knife 𓉱 🇺🇸🎗🇮🇱📟 @FreeManDespair@AviWoolf The ironic thing is, that the Fascist's foundational claim, that civilization is born and improved in war is invalidated by the very fact that the Europe that climbed out if the trenches was so degraded by the loss of its finest that it could produce Fascism2:16 AM · May 22, 2023 · 788 Views2 Replies · 8 LikesWhat are the similarities and differences of the Russian and French revolutions
Matt Clancy@mattsclancySuppose there are some economic activities that are zero sum, where R&D just steals market share. Maybe advertising, signaling education credentials, political campaigns. As an economy grows richer, do these sectors occupy a larger share of the economy? Is there a name for this?1:57 PM · May 23, 2023 · 14.4K Views7 Replies · 6 Reposts · 31 LikesIs there a name for how, if population risk of anything is some %, for the typical person looking at the stat it’s lower than that % because that % has to account for all the people much worse than average at risk management? Sort of pareto principle adjacent. For instance: if car crash deaths are some small proportional number of people per year, that number has to include all the people doing extremely stupid things with their cars (let's say 10x risk) AND "people like me" (1x risk) and so the metric is inflated compared to the real ratio, for a realistic you. (cc Croissanthology)
In what era of warfare were the *highest returns* to individual military genius, imagined as a comparable ability to chess genius, poker genius, etc.; some scale with GMs at one end and a few superGMs above, and then a few Carlsens and Kasparovs at the top.
Sort of imagining an inverted U-shape: too early, and there isn't the innervation of armies to be able to do very much with them except aim them & make a few decisions (where to set battle, initial arrangement of forces, when to commit reserves, fake retreats, committal charges); later, as armies get larger, operate over larger areas, there's a professional officer corps and staff system, and discipline in battlefield maneuver improves, real battlefield management and control is possible at a new level, with attendant greater returns to skill; but then, as the trend continues, "operation of war" becomes too vast for a single human brain to do much information processing, and the ability begins to reside more and more in corporate brains who have lower marginal effect and are more interchangeable, as seems the case in WW1/WW2 level and scale. But pretty uncertain.
So maybe Nader Shah or Napoleon were about the high point of this “unique contribution of the individual?”Somewhat confused about this because it seems like even in WW2, enormous discretion for decisions, including like when to invade or not, and where, rested with individual autocrats (still somewhat responsive to their general staffs)... so maybe the returns to skill were still going up. a lot going on, feel pretty unclear.
norvid_studies@norvid_studiesin what era of warfare were the *highest returns* to individual military genius, imagined as a comparable ability to chess genius, poker genius, etc. some scale with GMs at one end and a few superGMs above, and then a few carlsens and kasparovs at the top
The Last Great Arrakian Dynasty @cruelsardaukar@EgeErdil2 @norvid_studies @gleech I think that the ideal army still needs true greatness in the command chair but then a huge amount of bureaucratic competence at basically all other positions, because modern commanders can micro-manage to an unprecedented degree9:08 PM · May 23, 2023 · 321 Views2 Replies · 5 Likes- norvid_studies@norvid_studies@EgeErdil2 "how efficient is the tournament structure of a given era at surfacing genetic talent, from a larger pool available in that era, distinct from the training benefits of that later era": a question I don't have an answer to9:58 PM · May 23, 2023 · 211 Views1 Reply · 2 Likes
