I feel like taking in GenAI content, even if it makes me laugh, probably does something bad to my brain. It looks like real life, but the physics is just wrong in ways that range from obvious to very subtle. I don’t want to feed my brain videos of things that look photorealistic but do not depict reality, that just seems foolish somehow.
Like, imagine if you watched a bunch of GenAI videos of cars sliding on ice from the driver’s perspective. The physics is wrong, and surely it’s going to make you a worse driver because you are feeding your internal prediction engine incorrect training data. It’s less likely that you’ll make the right prediction in real life when it counts.
Back in the late 90s, when I first entered the video game industry to work (when it was quite scruffy, countercultural, and populated by some pretty odd people), one of the first things I encountered was a new co-worker who, next to his giant tower of used Mountain Dew cans, had a black and white TV in his cubicle. This struck me as very odd at that moment in time - as I understood things, obviously the point of work was supposed to be that it was a place where you worked, not a place where you watched TV. (Now, granted, everyone else was playing the recently released Diablo on their work PCs during lunch in network mode, and we were a game studio after all, so my reaction wasn't totally coherent). Still, no one else had a TV, and that guy was young and single with no work-life balance, he was a recent transplant, and it still seemed unusual at the time.
Fast forward 28 years later, and now everyone has an amazing TV in their pocket at all times when they commute, sit in their work space, go out for coffee or lunch, or go sit down in the bathroom, all with a near infinite collection of video via youtube, netflix, and even massive amounts of porn. How little did I know. And that's to say nothing of texting and twitter and reddit and instant messaging and discord and ...
Several years ago, I was working on a college campus, and there were giant corporate-flavored murals beside some of the city blocks students walked, full of happy multicultural clip art people and exciting innovative technological innovation, and adorned with the message, "Imagine a borderless world!" Clearly that message was meant to be rhetorical, not a call to reflection, critique, or reevaluation. There did not seem to be the suggestion that one might imagine the borderless world and then, having done so, decide it was a problem to be corrected.
I wonder a lot, these days, if we're not deep into a Chesterton's Fence situation, where we have to rediscover the hard way the older wisdom about having separate spheres with separate hard constraints and boundaries on behaviors, communities, and communication pathways to facilitate all sorts of important activities that simply don't happen otherwise - something like borders and boundaries as a crucial social technology, specifically about directing attention productively. Phones and tablets are, in their own Turing complete way, portals to a borderless world that pierces the older intentional classroom boundaries.
Look at the prior art. The key prior patent cited is BAESECKE, (2,134,850). This has a "wobbler" to change the transmitting frequency, which is fine, but that patent is vague as to how the receiving end stays in sync. Markey gets that syncing up transmitter and receiver is the real problem, and talks of it as a solved problem from television. (This is 1941, experimental TV exists.) Some mechanisms are discussed, borrowed from multiple player piano synchronization, which is where this technology came from.
Voice cryptosystems of the pre-WWII period included the Western Electric A-3 scrambler [1] There's one on eBay! [2] That split audio into a number of frequency bands, which were shifted and reassembled. At the receiving end, the process was reversed. The shift pattern changed periodically, on the order of tens of seconds. That was slow enough that keeping the thing in sync was possible with clockwork of the period.
Note that this is working on the audio, not the RF; it's a scrambler, not a frequency hopper. This is what AT&T used for transatlantic commercial voice. Worked OK, mediocre security. Only 6 different frequency shift patterns were in use at a time. The Germans cracked it.
This is not the better known SIGSALY. That's a similar concept, but with a lot more audio channels, much more frequent changes, a one-time key for the changes, and more hardware than some mainframe computers.
The A-3 was a desk-side wooden box. Neither system does frequency-hopping of the RF signal.
Hedy Lamarr's first husband was an arms manufacturer, and she apparently paid attention when visiting the factories. Hence the radio-controlled torpedo, which is close to Tesla's radio-controlled boat.
MARKEY (2,292,387) describes a cross between a radio-controlled boat and a set of synchronized player pianos. There's some handwaving around the sync problem. The trouble with syncing a frequency hopper is that you have trouble even finding the signal to get started. But if you're launching a torpedo or a bomb from a larger craft, both ends of the connection can be started in sync and will probably stay in sync long enough for the bombing run. Doing this with a player piano roll reader, vacuum pump and all, is probably not the right approach.
Trying to get things to sync up reliably has a long history. Edison's first useful invention was a way to get stock tickers to sync. There's a long history of clunky mechanisms, early ones involving flywheels or big tuning forks, and later electronic ones with too many screwdriver adjustments.
Not until the invention of phase locked loops did it really Just Work. (I'm into restoring early Teletype machines, and I'm way too aware of the early days of sync problems.)
Markey was just too early. Reasonable idea, but not practical at the time due to lack of supporting technology.
To the prompt “write a limerick about a dog,” GPT-2 wrote:
“Dog, reached for me
Next thought I tried to chew
Then I bit and it turned Sunday
Where are the squirrels down there, doing their bits
But all they want is human skin to lick”
While obviously not a limerick, I thought this was actually a decent poem, with some turns of phrase that conveyed a kind of curious and unusual feeling.
This reminded me how back then I got a lot of joy and surprise out of the mercurial genius of the early GPT models.
The first comment on that article caught my attention:
> I am a grad student in the philosophy department at SFSU and teach critical thinking. This semester I pivoted my entire class toward a course design that feels more like running an obstacle course with AI than engaging with Plato. And my students are into it.
It would be interesting to take a class of students and set them an assignment to come up with assignments for their fellow students that could not be completed using ChatGPT.
About ten years ago I went to a BarCamp conference where one of the events was a team quiz where the questions were designed to be difficult to solve using Google - questions like "What island is this?" where all you got was the outline of the island. It was really fun. Designing "ChatGPT-proof assignments" feels to me like a similar level of intellectual challenge.
It is always interesting when random touchstones from my life appear on Hacker News: books like the Aubrey-Maturin (master and commander) series, Ursula Le Guin’s works, Dante, John Le Carre’s George Smiley novels, Tolstoy... and now Charles Baudelaire, at the top of the page no less.
Baudelaire was a dark misanthrope and the poetry is very bleak. His life was not happy and he died at 46. You probably need to have at least a little of the same darkness in your soul to get something out of it.
It’s worth remembering too, how strange and controversial this work was when it first came out, using traditional verse forms but with a relentlessly modern subject, poetry from the gutter of the 19th-century city. Modernism in literature has had 150 years to settle but this is the raw beginnings.
Some good ones: The Albatross, Invitation to the Voyage, Evening Harmony, and the Epilogue (“Le coeur content, je suis monté sur la montagne”). And many others.
The main thesis here seems to be that LLMs behave like almost all other machine learning models, in that they are doing pattern matching on their input data, and short circuiting to a statistically likely result. Chain of thought reasoning is still bound by this basic property of reflexive pattern matching, except the LLM is forced to go through a process of iteratively refining the domain it does matching on.
Chain of thought is interesting, because you can combine it with reinforcement learning to get models to solve (seemingly) arbitrarily hard problems. This comes with the caveat that you need some reward model for all RL. This means you need a clear definition of success, and some way of rewarding being closer to success, to actually solve those problems.
Framing transformer based models as pattern matchers makes all the sense in the world. Pattern matching is obviously vital to human problem solving skills too. Interesting to think about what structures human intelligence has that these models don't. For one, humans can integrate absolutely gargantuan amounts of information extremely efficiently.
[Former member of that world, roommates with one of Ziz's friends for a while, so I feel reasonably qualified to speak on this.]
The problem with rationalists/EA as a group has never been the rationality, but the people practicing it and the cultural norms they endorse as a community.
As relevant here:
1) While following logical threads to their conclusions is a useful exercise, each logical step often involves some degree of rounding or unknown-unknowns. A -> B and B -> C means A -> C in a formal sense, but A -almostcertainly-> B and B -almostcertainly-> C does not mean A -almostcertainly-> C. Rationalists, by tending to overly formalist approaches, tend to lose the thread of the messiness of the real world and follow these lossy implications as though they are lossless. That leads to...
2) Precision errors in utility calculations that are numerically-unstable. Any small chance of harm times infinity equals infinity. This framing shows up a lot in the context of AI risk, but it works in other settings too: infinity times a speck of dust in your eye >>> 1 times murder, so murder is "justified" to prevent a speck of dust in the eye of eternity. When the thing you're trying to create is infinitely good or the thing you're trying to prevent is infinitely bad, anything is justified to bring it about/prevent it respectively.
3) Its leadership - or some of it, anyway - is extremely egotistical and borderline cult-like to begin with. I think even people who like e.g. Eliezer would agree that he is not a humble man by any stretch of the imagination (the guy makes Neil deGrasse Tyson look like a monk). They have, in the past, responded to criticism with statements to the effect of "anyone who would criticize us for any reason is a bad person who is lying to cause us harm". That kind of framing can't help but get culty.
4) The nature of being a "freethinker" is that you're at the mercy of your own neural circuitry. If there is a feedback loop in your brain, you'll get stuck in it, because there's no external "drag" or forcing functions to pull you back to reality. That can lead you to be a genius who sees what others cannot. It can also lead you into schizophrenia really easily. So you've got a culty environment that is particularly susceptible to internally-consistent madness, and finally:
5) It's a bunch of very weird people who have nowhere else they feel at home. I totally get this. I'd never felt like I was in a room with people so like me, and ripping myself away from that world was not easy. (There's some folks down the thread wondering why trans people are overrepresented in this particular group: well, take your standard weird nerd, and then make two-thirds of the world hate your guts more than anything else, you might be pretty vulnerable to whoever will give you the time of day, too.)
TLDR: isolation, very strong in-group defenses, logical "doctrine" that is formally valid and leaks in hard-to-notice ways, apocalyptic utility-scale, and being a very appealing environment for the kind of person who goes super nuts -> pretty much perfect conditions for a cult. Or multiple cults, really. Ziz's group is only one of several.
There might be some papers or other guides out there, but their advice will be based on whatever tools happened to be available at the time they were written and on the particular types of translations the authors cared about. The technology is advancing so rapidly that you might be better off just experimenting with various LLMs and prompts for texts and language pairs you are interested in.
I started using LLMs for translation after GPT-4 came out in March 2023—not that long ago! At first, the biggest problem was the context window: it wasn’t possible to translate more than a couple of pages at a time. Also, prompt writing was in its infancy, and a lot of techniques that have since emerged were not yet widely known. Even now, I still do a lot of trial and error with my prompts, and I cannot say with confidence that my current prompting methods are the best.
But, for what it’s worth, here are some strategies I currently use when translating with LLMs:
- In the prompt, I explain where the source text came from, how the translation will be used, and how I want it to be translated. Below is a (fictional) example, prepared through some metaprompting experiments with Claude:
- I run the prompt and source text through several LLMs and glance at the results. If they are generally in the style I want, I start compiling my own translation based on them, choosing the sentences and paragraphs I like most from each. As I go along, I also make my own adjustments to the translation as I see fit.
- After I have finished compiling my draft based on the LLM versions, I check it paragraph by paragraph against the original Japanese (since I can read Japanese) to make sure that nothing is missing or mistranslated. I also continue polishing the English.
- When I am unable to think of a good English version for a particular sentence, I give the Japanese and English versions of the paragraph it is contained in to an LLM (usually, these days, Claude) and ask for ten suggestions for translations of the problematic sentence. Usually one or two of the suggestions work fine; if not, I ask for ten more. (Using an LLM as a sentence-level thesaurus on steroids is particularly wonderful.)
- I give the full original Japanese text and my polished version to one of the LLMs and ask it to compare them sentence by sentence and suggest corrections and improvements to the translation. (I have a separate prompt for this step.) I don’t adopt most of the LLM’s suggestions, but there are usually some that I agree would make the translation better. I update the translation accordingly. I then repeat this step with the updated translation and another LLM, starting a new chat each time. Often I cycle through ChatGPT --> Claude --> Gemini several times before I stop getting suggestions that I feel are worth adopting.
- I then put my final translation through a TTS engine—usually OpenAI’s—and listen to it read aloud. I often catch minor awkwardnesses that I would overlook if reading silently.
This particular workflow works for me because I am using LLMs to translate in the same language direction I did manually for many years. If I had to translate to or from a language I don’t know, I would add extra steps to have LLMs check and double-check the accuracy of the translation and the naturalness of the output.
I was asked recently by some academics I work with about how to use LLMs to translate documents related to their research into Japanese, a language they don’t know. It’s an interesting problem, and I am planning to spend some time thinking about it soon.
Please note that my translation process above is focused on quality, not on speed. If I needed to translate a large volume of text more quickly, I would write a program to do the translation, checking, and rechecking through API calls, accepting the fact that I would not be able to check and polish the translation manually as I do now.
If anyone here would like to brainstorm together about how to use LLMs for translation, please feel free to email me. My website, with my email address on the Contact page, is linked from my HN profile page.
I am only partially qualified in that I am not a professional archeologist, but I have done post-doctoral archeological studies and have read enough archeological studies to understand the larger academic context.
It is not possible to present all the data informing a judgment in such a short work. Even in a book, it would not be possible. Thus it is common in archeology for papers to be written as part of an ongoing conversation / debate with the community - which would be defined as the small handful of other archeologists doing serious research on the same specific subject matter.
Part of that context here is that these tombs are well-established to be the royal tombs of Alexander's family, spanning a few generations including his father and his son. This is one of the most heavily studied sites in Greece for obvious reasons, and that is not something anybody is trying to prove.
In that context, his arguments are trying to identify any body as one among millions, but as one among a small handful of under ten possibilities.
At the same time, the fact that he is not a native English speaker and general archeological style come into play. For example:
"the painter must have watched a Persian gazelle in Persia, since he painted it so naturalistically (contra Brecoulaki Citation2006). So the painter of Tomb II has to be Philoxenus of Eretria" sounds like a massive leap, and it is. He continues:
"... Tomb I (Tomb of Persephone) must have been painted hastily by Nicomachus of Thebes (Andronikos Citation1984; Borza Citation1987; Brecoulaki et al. Citation2023, 100), who was a very fast painter (Saatsoglou-Paliadeli Citation2011, 286) and was famous for painting the Rape of Persephone (Pliny, N. H. 35.108–109), perhaps that of Tomb I."
Another huge leap, both 'presented as conclusions'. However he then continues to indicate these are just hypotheses: "These hypotheses are consistent with the dates of the tombs..."
So his English language use of presenting things factually does not indicate certainty in the way the words would be used in everyday speech. He seems to perhaps misunderstand the force of the terms, but also appears to be working within the context of the conversation with other archeologists I mentioned to start: They all know every affirmation is as "probably", rarely anything more. So it is relatively common shorthand of the craft in that sense.
I believe you are overthinking his responses to other authors, although I understand the culture shock. It is an ongoing conversation and archeologists tend to be blunt in their assessments. Add Greek bluntness on top of this, and it does not seem to matter to the material.
As to your last question, is this legitimate research? The answer overall appears to be yes, although I could see several points (such as the identification of artists I quoted above, and various items I noticed), which I would never have put into ink the way he did. Still, most of his arguments are compelling. It is a shame that the aggressiveness of a few affirmations detract from the overall value of his work. Archeology is not code nor is it physics. It does not pursue universal truths that are more easy to verify through repeated experiments, but unique historical ones which necessarily attempt to interweave physical details and ancient historical records. Each field has its own level of certainty, and the fact that we cannot establish these details with the same certainty as we can establish the chemical formula for water does not make them useless, or pure inventions. Far from it.
>Other important literature that was published during this time was work by Watpole himself. His novel, Castle of Otranto, was reportedly inspired by a dream he had while living at Strawberry Hill. Set in a castle in the Middle Ages, the epic details a lord and his family living in a haunted mansion. “In the late 18th and 19th century, Gothic became associated with spookiness, which got wound into ideas of the exotic and sublime,” Dr. Bork says. “By the 20th century, you have movies and mass media that start using this.”
That's... not a lot of detail.
The narrative I like comes from Walt Hickey's You Are What you Watch. Basically, there was wealth in the 1870s and 1880s during the Gilded Age, and those people built homes in the Victorian/Gothic/Queen Anne style. Their kids grow up in those homes, and suddenly books are becoming movies (early successes like Dracula in 1897 as a book and eventually movies), and horror is a big hit, and the kids who grew up in those homes are writing things that take place there. Meanwhile, the stock market crashes, those homes are abandoned and unmaintained and derided. "When a boring colonial-style home deteriorates with age, it looks distinguishing. When a fantabulous, whimsical home deteriorates with age, it starts to look spooky."
I have toyed with a slightly absurd but factualy
correct history of the ancient world, told from the point of view of roofers.
One of the main objectives of many, many, many
conquests was to steal the lead roofing that was used to cover and seal/water proof the roofs of
countless civic and religious and religious buildings,another main target were.....are....
the collums used to support said roofs, or lions
or dragons,horses,whatever....
can't imagine the chaos unfolding in high school and college freshman English composition classes, where for the sake of grading ease, "fairness" and simplicity (and more than any of these, instructor complacency) students are encouraged and incentivizes to write in a very mechanical and formulaic style near perfectly emulated by the default behavior of ChatGPT and friends.
Training, rewarding (and thus selecting for opportunity in favor of) people who optimize for and pride themselves on mechanical precision and speed in their work is setting them up to be defeated by the actual machines they will be competing against in the market.
Even the largest LLM has had less "total information" than most humans take in through all of their senses over their lifetime. A single day for a baby is taking in a continuous stream of among other things high quality video and audio and does a large amount of processing on that. Much of that for very young babies is unsupervised learning (clustering), where baby learns that object A and object B are different despite knowing nothing else about their properties.
Humans can learn using every ML learning paradigm in ever modality: unsupervised, self-supervised, semi-supervised, supervised, active, reinforcement based, and anything else I might be missing. Current LLMs are stuck with "self-supervised" with the occasional reinforced (RLHF) or supervised (DPO) cherry on top at the end. non multi-modal LLMs operate with one modality. We are hardly scratching the surface on what's possible with multi-modal LLMs today. We are hardly scratching the surface for training data for these models.
The overwhelming majority of todays LLMs are vastly undertrained and exhibit behavior of undertrained systems.
The claim from the OP about scale not giving us further emergent properties flies in the face of all of what we know about this field. Expect further significant gains despite nay-sayers claiming it's impossible.
Yes, he's the water we're swimming in. Here's an example I've been wondering about for years:
Freud taught that early childhood experience has a profound impact on adult life. Did anyone have this perspective before Freud? Was there a precedent? If I think of literary depictions of childhood in the 19th century, for example in Dickens, there can be great pathos, but not a psychological awareness of the consequences.
This emphasis on early childhood experience is tectonic in Western culture and we still haven't caught up to it in practice. If it really was Freud's contribution, that alone makes him great.
Another dimension of Freud is that he was a taboo-breaker, and in that sense a heroic figure. Some of the steps he dared to take are still so out-there that they can't be openly discussed. That's pretty remarkable after a hundred years.
It’s fascinating how little people really understand how LLMs are impacting businesses. Almost any task that couldn’t be done because it was too expensive to hire thousands of humans to sift through things is being outsourced to LLMs. I find it remarkable that people don’t realize how amazing these things are at classification and structuring of complex stuff with some alacrity, often far superior to human agents. They scale arbitrarily, work 24x7, and have very low failure rates compared to poorly trained high turnover humans. They generally do a good job identifying when they can’t classify something and delegate to human review. Are they perfect? No, but they’re considerably less error prone than a staff of hundreds of humans. Is it lucrative? Absolutely. I’ve seen this now at five major megacorps and I have to believe it’s going on at most.
Now that this and other articles have established that the board's actions were based not on a single event but rather on general, long-term, underspecified concerns, no one captures the essence of what happened more perfectly than Matt Levine:
"Well, sure, but [the board not trusting Altman] is a fight about AI safety. It’s just a metaphorical fight about AI safety. I am sorry, I have made this joke before, but events keep sharpening it. The OpenAI board looked at Sam Altman and thought 'this guy is smarter than us, he can outmaneuver us in a pinch, and it makes us nervous. He’s done nothing wrong so far, but we can’t be sure what he’ll do next as his capabilities expand. We do not fully trust him, we cannot fully control him, and we do not have a model of how his mind works that we fully understand. Therefore we have to shut him down before he grows too powerful.'
I’m sorry! That is exactly the AI misalignment worry! If you spend your time managing AIs that are growing exponentially smarter, you might worry about losing control of them, and if you spend your time managing Sam Altman you might worry about losing control of him, and if you spend your time managing both of them you might get confused about which is which. Maybe Sam Altman will turn the old board members into paper clips."
I think the best way to try this out is with LLaVA, the text+image model (like GPT-4 Vision). Here are steps to do that on macOS (which should work the same on other platforms too, I haven't tried that yet though):
Interesting that they used Chain of Thought Prompting[1] for improved reasoning so soon after its publication. Also related to DeepMind AlphaCode which generates code and filters results by unit tests, while Chain of Thought Prompting filters by checking for correct answer at the end.
Seems like language models can generate more training data for language models in an iterative manner.
This is awesome. I've been speculating along similar lines, and it's great to see this fleshed out.
I think "correct the errors in this ChatGPT essay" is a short-term viable homework exercise, but those errors might be gone in GPT-5 so I don't think it's long-term viable. Soon the LLM will just produce perfect essays at college level and there won't be hallucinations for the student to correct.
However, the "simulate the historical environment" task is great and I think it has long-term potential. I think it can be taken further; rather than "spot the errors that ChatGPT made", you could flip the script and make it "survive 20 turns of conversation without making a historical error", so you'd need to know things like local traditions, perhaps the geography of the ancient settlement you're studying, contemporaneous history like "who is the emperor and what's the sentiment towards him" and so on.
I'm also envisioning that, since text-based exercises are extremely easy to game (just pipe your text prompt into ChatGPT), and since ChatGPT is soon going to be strictly superior to a high-school level student, we could get around this by having the homework as an in-person verbal role-play or Q&A session, like a viva voce; essentially you have a verbal discussion with ChatGPT and you need to really know your material as it can dig into any part of the curriculum. Then ChatGPT can summarize each student's interaction, and the teacher doesn't have to sit through each individual one start-to-finish (1:1 exams are too time-consuming to be viable).
This round-trip through verbal interaction would potentially make the task more interesting (lots of people simply hate writing essays), shifts the focus away from tasks that will become obsolete (writing essays) in favor of ones that will be more relevant (human synthesis of ideas, and interpersonal interaction), and helps to mitigate the issue of LLM-assisted cheating by constructing an assignment that LLMs can't trivially solve.
Helen Toner has done a huge amount of work slowing down capabilities research in China. That's why she lived in Beijing for a year, she is a big part of why there are a lot of Chinese researchers from various AI labs signed onto the CAIS statement, and it's what her relationship to the Pentagon is all about. I think she is probably the individual person who knows the most in the world about the relative AI capabilities of China and the US, and her career is about working with the Pentagon, AI companies in the USA, and AI companies in China to prevent an arms race scenario around AI. It's the sort of work that really, really doesn't benefit from a lot of publicity, so it's unfortunate that this whole situation has put her in the spotlight and means someone else will probably need to backchannel between the US and China on AI safety now.
I don't know why she chose to publicly execute Altman, there just isn't enough information to say for sure. It probably wasn't a specific, imminent safety concern like "Our new frontier model was way more capable than we were expecting and attempted a breakout that nearly succeeded during internal red teaming", according to the new CEO it wasn't anything like that. The new CEO has heard their reason, but is putting a lot of pressure on them to put that reason in writing for some reason. I don't know why, we just don't have enough information.
But basically, she is a very qualified person on the exact topic you are concerned about and has devoted her career to solving that problem. I wouldn't write off what she's doing or has done here as "She didn't consider that China exists".
Anyone who has iterated on trained models for long enough knows that feedback loops can be a serious problem. If your models are influencing the generation of data that they are later retrained on, it gets harder and harder to even maintain model performance. The article mentions one experiment in this direction: "With each generation, the quality of the model actually degraded." This happens whenever there aren't solid strategies to avoid feedback loop issues.
Given this, the problem isn't just that there's not enough new content. It's that an ever-increasing fraction of the content in the public sphere will be generated by these models. And can the models detect that they are ingesting their own output? If they get good enough, they probably can't. And then they'll get worse.
This could have a strange impact on human language / communication as well. As these models are increasingly trained on their own output, they'll start emulating their own mistakes and more of the content we consume will have these mistakes consistently used. You can imagine people, sometimes intentionally and sometimes not, starting to emulate these patterns and causing shifts in human languages. Interesting times ahead...
I think it was in the book Dream Machine, mentioned elsewhere in the thread, where it described that there were largely two different philosophies of the relationship between the human and the machine/computer. One was "augmenting the human intellect", where technology serves to extend human abilities. The other was - if I recall correctly - cybernetics, where the human serves as part of a larger mechanical system. I might be remembering it wrong though.
Just a few days ago I was reading "Thinking with demons," an excellent history of ideas about witchcraft in Europe, and was struck by the following passage about a 16th century drawing of three witches by Hans Baldung Grien:
Informing everything in the scene, and establishing iconographically that it is indeed a scene of witchcraft, is the gesture of the witch who, bent on one knee, stares backwards at the world through her own legs. According to a contemporary German proverb those who adopted the pose would be sure to catch sight of the devil. This is perhaps the reason why the motif is also found among the monsters and devils who populate two widely separated versions of that most demonological of picture subjects, the temptation of St Anthony—those of Hieronymus Bosch (c. 1490s) and Jacques Callot (1635).
So it was fascinating to learn from this news article that "There are also regions of Japan where the folklore says that one can see ghosts, the world of spirits and demons, or the future by looking upside down through one’s legs."
There's a fascinating book containing short first-person accounts of living through the industrial revolution called Pandaemonium: Coming of the Machine as Seen by Contemporary Observers edited and with notes by Humphrey Jennings. It covers 1660-1886, and it's one of the few books that genuinely blew my mind. It's extremely accessible because of the first-person nature of the accounts, and I think one of the reason it's so compelling is that you can feel the change sweeping through so many aspects of daily life. I find it a very interesting lens through which to look at the current digital revolution.
I actually was the sole developer who wrote all the software that “networked” the custom hardware together. This project was so ahead of its time and yet required some rather arcane programming knowledge. So much fun though. AMA!
I have been practicing traditional Asiatic archery. There is a bunch of stuff in that article that is misleading.
First, the composite bow design used by the Mongolians was widely use as far as the region of the Ottoman-Turks … but also the Manchus, Chinese, Koreans.
Second, the recurve and reflex shapes of the bow is not what gives the bow its power. It is siyahs at the tip, which do not bend, and acts as a lever on the bendable parts. There is a tradeoff. The larger the siyah, the more mass the bow has to overcome to fire these arrows, so the draw weight need to be significant enough to warrant the use of a siyah. Turkish and Chinese composite bows start outperforming longbows of the same draw weight at around 40-50 pounds, while the Manchu horn bows start outperforming longbows at 70-80 pounds. According to those who test it out, the performance gap is exponential.
Third, the most powerful of the Asiatic bows are probably the Manchu bows. They double-down on making long, aggressive (forward angle) siyahs, and even innovating string bridges to increase brace height and help prevent the bows from unstringing themselves. They were not designed to shoot arrows fast. Like most Asiatic composites, these were designed to shoot very heavy, long arrows with great penetration power. Manchus also love their horses, archery, and the hunt.
A hundred years after the Manchus conquered China and founded the Qing dynasty, their archery traditions were in decline. The Qing emperors recruited Mongolians to be trained in the Manchu ways of archery to keep their traditions alive.
There is a lot more to it as well — such as the thumb draw and use of thumb rings; the variations of khatra, etc; how modern Mongolian archers competing in the Mongolian Games are switching away from thumb draw.
Lastly, a note on the Chinese. With early access to metallurgy and ideas on standardized parts, the Chinese “superweapon” was the crossbow. At no point in the Chinese history were there armor that could protect someone from a crossbow, and its effectiveness was likely a factor in the slow adoption of firearms.
There's lots of very exciting work going on around the fully mapped fruit fly connectome. For example, I'm a CTO of a stealth startup that aims to do for utilitarianism what carbon credits did for environmentalism. We are selling 'utility credits' which translates directly into us simulating trillions and trillions of fruit fly brains in a state of constant orgasmic bliss, which you can then buy to offset any actions your company has undertaken that damage global happiness or well-being. We've seen a lot of interest from some pretty large industry players.
Like, imagine if you watched a bunch of GenAI videos of cars sliding on ice from the driver’s perspective. The physics is wrong, and surely it’s going to make you a worse driver because you are feeding your internal prediction engine incorrect training data. It’s less likely that you’ll make the right prediction in real life when it counts.