Caution
This is an unedited draft written in a single continuous session. It is rambling, not fact checked, and doesn’t have citations yet. This draft is for feedback. An edited version may be published elsewhere and this version edited or removed without notice. Please do not share until this notice is removed.
Last week, my daughter and I were watching New Amsterdam, a Netflix show about doctors who have taken over their hospital and must now somehow save their patients without killing their own operation. It was the season one finale. The doctors are packed into an ambulance, saving one of their own, when they have a fatal crash. The protagonist stumbles around the hospital through the episode, his mind faltering between memories, imagined futures and a confused present, checking on his colleagues while fending off nurses who insist he needs rest.
Who will we find dead by the end of this episode? “Dr Sharpe,” my daughter announced. She had binged through several seasons of a detective show and made a sport of solving the whodunnit from the clues presented in each episode. Dr Sharpe was missing from the list of individuals accounted for and the tension was clearly building up, but I found myself surprised at my daughter’s suggestion. Dr Sharpe was not going to die in this episode unless there was an externality, like the actor having died or having left the show over a contract dispute. The story required her to not die. Someone else was going to die, someone setup for death in this episode. Sure enough, I was right, but it bothered me. This episode depicted an accident. Accidents aren’t staged and their outcomes are unpredictable. How could I be so assured then?
Children tend to notice all details because they don’t yet know which ones are irrelevant. This is a well known problem for magicians whose method is the sleight of hand, directing the audience’s attention to a harmless distraction while their hand is busy manipulating an object until it magically pops back into view. It works because adult brains ignore irrelevant detail, so the magician only has to conjure one irresistible detail to mask everything else. But the younger the audience is, the harder this is to do, because young brains don’t have that filter and will notice everything. Children don’t see magic tricks. They will plainly see how the coin was moved from one place to another, and sometimes even explain this loudly to an embarassed amateur magician.
My daughter, trained on hundreds of hours of episodic scriptwriting, saw the clue from the one missing character over the duration of the episode – but as we now knew, the scriptwriters weren’t hinting at who’d be gone. If anything, they were merely building up some tension to be released over the character who’s actually being written off. But I hadn’t felt this tension at all! I had felt completely assured at who was going to die and who wasn’t, and this moment of disagreement with my daughter popped it into focus. It was almost as if I wasn’t emotionally involved in the story, instead merely running an autocomplete in my head and deriving satisfaction from being on track. How had I found myself here?
Accidents don’t happen in TV shows and films. They’re staged. This is obvious at one level because the grand spectacle of cars flipping over and captured from multiple camera angles doesn’t happen unless those cameras were deliberately planted there, with the cars remote-controlled and the collision physics carefully calculated, because there isn’t an endless supply of cars for retakes. There is no accident whatsoever in an accident on film. But this is also true at a deeper level: an accident is a narrative device to convince the audience that the characters in the story are truly helpless. After an entire season of doctors taking extreme measures to save their patients, undertaking risky procedures, defying their supervisors, even flouting the law, one of their patients can’t simply die of natural causes. It can’t be a simple “we tried“. The death has to be grand, a force majeure event, what insurance companies lovingly call an “act of God” to excuse themselves from paying up. Any less and it risks infuriating the audience and having them nose around backstage making accusations that the character was forcibly killed so the scriptwriters could buff up another character’s story. The point of a script is to create a believable story, so if the audience is peeking behind the curtain to make sense of it, you’ve failed.
The author Arthur Conan Doyle learnt this the hard way when he got bored of writing Sherlock Holmes stories and wanted the character gone. He staged a scuffle between Holmes and his antagonist Moriarty in which they fall off a cliff into a waterfall and are never heard from again. Nobody could survive such a fall, Doyle reasoned, but his audience revolted. They wouldn’t accept it, and they were loud. Doyle found no peace until he conjured a miraculous escape for Holmes and resumed the detective stories, this time writing out a planned retirement for the character to prepare his audience for his own retirement.
Accidents are taboo.
Which is why my calm confidence about who was going to die in this episode was suddenly unsettling. I had been looking through the story at the narrative device all this while, and I only realised it when my daughter presented her observation. What, exactly, had I been watching all this time?
I have trouble reading emotion and I’ve known this for a long time. I’m a terrible negotiator. When making a pitch, I make up my mind about what the fair offer is, and proceed to defend it. Recovery from failure is a re-evaluation of the components of the offer. I’ve only found success as an entrepreneur when I got out of sales completely, contenting myself with charge of the non-negotiables. I used to flounder watching films too. I remember one particularly embarassing instance when my colleagues took me to a Merchant-Ivory film festival to watch Shakespearewallah. Towards the end, Shashi Kapoor’s character has an emotional outburst as he tells his lady how much she means to him. The very next scene, she’s on a ship sailing away and the credits roll. I was stumped. What had just happened? Had they run out of time and decided to just end the movie? He told her how he feels and she just left? My colleagues glared at me. Was I simply too young to understand this film? That outburst had been completely one-sided, all about how he felt and why his feelings mattered, a tension that been building up through the film. Of course she left, the story was over! I was 19 then and had never been in love, so this went into the mental reference of feelings one is supposed to have. I’ve since learnt that my condition isn’t unique, it’s just somewhere in a spectrum of normalcy. My fellow afflicted learn to extrapolate from our own feelings into familiar scenarios that we recognise others in. My fellow AI nerds may have an easier analogy: if you don’t have a great GPU, you have to cope with a CPU, which works but isn’t the right architecture. It can’t parallel process, so it’s single-tasking and slow, gets overloaded easily and overheats, and then requires a slowdown for recovery. Some may call this introversion.
The coping mechanism for emotional maturity – having a large reference of familiar scenarios and the emotions they represent – requires first hand experience, so I was indeed too young at that time to understand that film. Turns out this is a particularly valuable skill however, for the ability to anticipate emotions in a future scenario is absolutely necessary when building experiences for others, whether as a computer programmer, designer, architect or indeed scriptwriter. When I watch a show, I’m now invariably confirming that I’m reading it correctly, that I’ve spotted all the inflections. Their work is my training data.
Scriptwriters have rules for this. Doyle showed us what happens when you carelessly stage an accident. The rule – and I’ll summarise from my understanding here – is that any character or object that causes an inflection in the story cannot be introduced at the time the inflection happens. They have to be introduced in advance. This is because they’re outside the universe of the story the audience is familiar with until that point, so anything they do is an Act of God, a divine intervention. What was the point of giving the audience so much build-up if it was totally irrelevant to this moment? All significant characters and objects must be introduced in advance for the audience to notice and anticipate their re-appearance when they’re needed.
Oddly enough, the Act of God wasn’t always evidence of an amateur writing themselves into a corner and then playing Calvinball to escape. It used to be a formal storytelling device, and it’s ancient. The Greeks called it Deus Ex Machina, literally God of the Machine, because in their plays the god descends over the stage suspended by rope and pulleys, a machine operated from behind the curtain. When they tell a story of hubris, of a man who rises above all men, the god must descend to put the man in his place and conclude the story. This god exists in this universe and is anticipated even if only making a first appearance when needed. This god’s background presence is so ordinary that you can have a Biblical story about a man who’s been ordered to kill his own son and proceeds to do it, only to be foiled by the very same authority. Modern scholars get into knots discussing the location of free will and liability in this story because modern scholars don’t like this god who defies any definition and have to make sense from the protagonist’s perspective. Was his attempt foiled by the hand of god or the word of god? Was there actual intent to murder, or assured knowledge of it being a test? If he swung the axe and did not miss, did he pass the test or fail it? You can see this analysis will give you a headache unless you sincerely believe the background god exists and knows best, and fellow adherents of this belief system can confidently write stories where this god drops into the story with the appreciation of the audience.
Modern adaptations of mythologies will continue to use this device when it’s popularly expected in the story, but sometimes the filmmaker can’t bear the embarassment and must adapt. Ridley Scott in Exodus: Gods and Kings (2014) couldn’t make Moses part the waters, audience expectation be damned. Instead, we were shown a weary leader who spotted a tenuous opportunity and took the risk on behalf of his people, hurrying them across a dry riverbed before descending floodwaters closed the path. Their pursuers were washed away, not by an act of god, but their own hubris in pursuit against the overwhelming force of nature. High five, Mr Scott.
Comics writer Stan Lee famously insisted all of his superhero characters existed in the same fictional universe, even if each had isolated stories. Superheros are by definition divine, for they can stage miracles that ordinary humans cannot. They exist to narrate escapist fantasies for incurable social ills. See for instance Jill Lepore’s The Secret History of Wonder Woman. But superheroes are hard for anyone to relate with for they don’t have the same frailties. Makers reach into familiar tropes, a tormented mind or an equally fantastical antagonist. In the latest reboot of The Batman (2022) from director Matt Reeves, the bat mask is an embarassment, only worn when the character is emotionally prepared. The antagonists are all ordinary everyday people without exaggerated quirks. Their power is merely emotional torment. This is a story of a man afraid of who he is. I was particularly tickled by this emo-Batman because Michael Keaton, the actor who came to fame for a prior portrayal of Batman, appears in Birdman (2014) as an actor known primarily for playing a bird-themed hero, struggling to have an identity beyond that role. Birdman was painful to watch because actor and character blended into each other. Keaton was mocking his own identity crisis, presenting his insecurities as his character’s insecurities. The new Batman adopted this becoming-trapped-as-Batman as the character’s primary crisis.
Stan Lee’s insistence on a shared universe – made generations prior – eased his burden of building relatability, for heroes and villains could crossover between their stories without the bother of build-up. This idea was rebooted for the films and they now exist in a distinct Marvel Cinematic Universe, where the events of one film are the background of the next, even with entirely different characters. Sometimes this is visibly awkward. Daredevil is about a blind lawyer who turns costumed crusader at night. He can’t see, but when he turns off the lights in a roomful of baddies, they’re totally unprepared for the dark and the lethal chops landing from every direction. As a bearer of hope for the disabled, Daredevil is heartwarming. Unfortunately, the show was released after an Avengers movie in which aliens dropped out of the sky and destroyed his neighbourhood in New York City, so this patron hero of disabled orphans, tasked with protecting them from exploitation, is surrounded by people who keep recalling that recent time when aliens blew up their buildings. Our hero is totally unmatched and the aliens have no role here. The awkward reference is solely to inform the audience that this too is part of the Marvel Cinematic Universe, even though this is a TV show on Netflix while Marvel is on the way to acquisition by rival Disney. Daredevil and his fellow coterie of Defenders have since left Netflix for Disney+.
Sometimes a filmmaker will somehow goof it up spectacularly, dropping a twist that the audience is utterly unprepared for. I remember watching The Adjustment Bureau (2011) and walking out bewildered. That night’s furious googling was how I discovered this lovely device I’m now brandishing before you, the Deus Ex Machina.
How is this even remotely related to Artificial Intelligence? I’m getting there, I promise, but there’s more ground to cover.
We’ve kept the trope but discarded the god. That old god of miracles is unwanted, beaten into hideouts like your insurance policy, where you’ll find a section on exemptions titled Acts of God, or if they’re embarassed, as Force Majeure. If God smites you the insurance company won’t dare rise above God to support you, it says, but here too the appeal to divinity is only masking ordinary business logic: if too many people make a claim at the same time their finances will go bust, so they reserve the right to save themselves first.
If there’s a god dropping miracles in your story today, what you have is a bug that must be quashed, and the easiest way to strip it of divinity is by bumping the character or object back into the past, letting the audience notice and remember so the next appearance is not miraculous. An amateur will toss it back a few scenes, but a great writer will plant it far, far back and let the anticipation fester until the moment of use also triggers a release for the audience.
In the middle of season two of Severence, a TV show about workers who have achieved work-life balance through a technological fix, one character receives a gift that reminds him of his skin tone. He accepts the gift, takes it home and places it in storage. The next day he asks a colleague for her thoughts. She gives him a canned response. We know it’s canned because it doesn’t answer the question that was not asked, a question that’s lingered as a tension since the gift was first unveiled. The gift and skin tone aren’t brought up again, but the character’s sense of identity and belonging have been poked, and that unsteady footing stumbles over the rest of the season to spectacular effect. I was astonished to learn the character’s skin tone was specified in the casting call several years ago. They planted it right at the start, and then they waited a long time to use it so very carefully.
Shows like Severence grow communities of theorists dedicated to analysing every obscure clue. In a modern production this is a sport between these two sides, for the critics and commentators bring in an audience comparable to the show’s own marketing efforts. This presents the next problem, the false god, a character or object that doesn’t do anything for the story. If you give the audience a bauble and hint that it’s important, and then just leave them holding it, you can expect the audience to feel betrayed. This is in the same vein as a political party’s electoral promises, or a religious leader who assures you their God is Coming Soon. We see it as a trope, a performance that has some significance within their belief system, but meaningless outside it.
TV has tropes, of course, a vast collection of such narrative devices that have been overused and are easily recognised, and sometimes intentionally used to reference the trope itself. We’ve discussed the god of the machine, but here’s another, a term you’ve very likely heard: jumping the shark. It originates in a show that, faced with declining viewer metrics, had a character literally jump over a shark. The scene made no sense within the show, its purpose being to cause a stir and convince more people to watch. In effect, a showrunner’s desperation for relevance causes the characters to break out of their world to grab more viewers and bring them back, for there’s no better explanation for why some scene belongs in the story. The show has jumped the shark.
Related, the notion of a fourth wall. A low budget show set indoors will have three visible walls in the room. There’s no fourth, for that’s the side with the cameras and production crew. The fourth wall is a whole family of tropes, and one of them involves breaking it, when a character faces the wall and speaks to the viewer. Comic strips have thought bubbles and novels have a variety of mechanisms to present a character’s inner thoughts, but film and TV are particularly constrained for you can’t hear a thought. It has to be spoken out. A scriptwriter charged with such a translation may find they have to rewrite the story to add scenes where thoughts are presented as dialogues, but the easier way out is to simply have the character speak at the fourth wall. Recasting thoughts as dialogues leads to yet another trope, exposition, where a dialogue or presentation exists primarily for the viewer’s benefit. To avoid this trope, the script needs both audience and characters to receive information at the same time. This leads to yet another, of a character who exists solely to represent the audience during exposition.
The counter to exposition is yet another rule, “show, don’t tell.” Severence never actually discusses the character’s skin tone. We were shown the gift and we inferred discomfort from the awkwardness with which it was handled. No words were spoken.
I hope you can see where I’m going with this: scriptwriters risk having their own constraints entangled with their scripts and must carefully extricate themselves, but this still leaves an imprint that the audience can see.
The Office (2001–03 UK; 2005–15 US), a show about the tropes of corporate working life, also cheekily embraces the tropes of its medium. The characters exist in their own reality TV show, giving interviews after every scene to speak out their inner monologue. The audience is promoted from behind the fourth wall to a seat within the story as the interviewer, albeit one who never speaks or appears on camera. This narrative device spawned its own genre, the mockumentary. In Borat (2006), the character is clearly exaggerated, but everyone else is not an actor and not aware they’re speaking with an actor. The camera crew do not speak but are occasionally acknowledged when an interviewee breaks the fourth wall – unintentionally, for they’re not actors and are actually addressing the cameraperson. Taika Waititi took this further in What We Do In The Shadows (2014), a film about immortal vampires living the quiet suburban life, planning dinner parties where their guests will be on the menu. The vampires have hired a documentary crew and granted them immunity from being on the menu themselves – as we learn when they remind one of their senile roommates – until they run into a rival pack of werewolves who aren’t aware of this contract. The vampires flee, leaving the camera crew undefended. We’re treated to shaky footage and heavy huffing, the first acknowledgement that the invisible crew are also human, with human frailties. Waititi’s formula was a hit and What We Do In The Shadows rebooted as a TV series (2019–2024), moving from his home country New Zealand to a suburb of New York.
Waititi wasn’t just toying with the inherited trope. He was also satirizing an adjacent genre, the reality TV show. Bear Grylls in Man vs. Wild (2006–11) finds himself stranded in wilderness with no resources, and must survive and find his way back to civilisation. Grylls does everything solo, from building a shelter to hunting a meal, swatting away insects while he tries to sleep without comfort. His film crew presumably bring their own food and shelter, for their travails are not shown. This also means relief was available but denied, the depicted reality being only a sanitized recreation of said reality. In February 2019, Grylls shot an episode in India with Prime Minister Narendra Modi, documenting their isolation in the wilderness while the Pulwama attacks happened.
YouTubers filming their own lives are generally less adept at masking this tension. Are they operating their own cameras? Are they a team taking turns? Is the cameraperson a hired hand, excluded from being a participant in their reality? Does this then also reflect a class hierarchy, similar to how a domestic worker present in your home is neither family nor friend?
In visual media, script is inseparable from scriptwriting. The process imprints into the product. The audience sees these marks, and the makers in turn manipulate these marks to augment the story. Which brings us back to the Deus Ex Machina problem: you do not want a false mark, a character or object that is not part of the story. Since this is visual media where a crowd can’t be depicted without people and a room can’t be occupied without objects, the task is to strip these entities of any notability, replacing them with generics, stereotyped normality.
As I sat that evening reflecting on that scene in New Amsterdam, the generics started popping into attention.
If a phone appears in a scene, it’s always an iPhone. The only generic exceptions are period dramas with old technology and sci-fi with futuristic technology. A phone that’s not an iPhone is a mark: a product placement, a hint at the bearer’s class location, or that particular device is itself a participant in the story.
An iPhone is a filmmaker’s vanilla flavour of phones, meant to tell you that the phone is not an important object and can be safely ignored. What matters is the conversation that happened on the phone or the alarm that rang to wake up a character and start their day. Vanilla, of course, is a real flavour. It was once rare, exotic and expensive. The quest for vanilla fueled international trade wars and built the backbone for European colonization of the entire world. There is nothing ordinary about it. Vanilla was a protagonist in its own right, and yet has had the misfortune of its name becoming associated with the expression vanilla flavour, for use whenever a flavour is uninteresting or irrelevant. Filmmakers faced with the challenge of depicting a phone that’s not interesting have collectively agreed on turning the iPhone into the vanilla flavour of phones.
These stereotyped generics are everywhere. If a character in an American show is black, brown or Asian, that detail is important and will be referenced in the story. White skin is the vanilla flavour of skin tones: if a character is white, that detail is not relevant to the story.
This doesn’t mean filmmakers are necessarily constrained to them. Just as with stock tropes, some take delight in demolishing these generics. In Black Panther (2018), there are only two white-skinned characters, one a stock villain who serves to highlight the real antagonist, another a bumbling official who represents his entire institutional apparatus. I found the story a bore, but the real excitement was in the meta narrative, for the film stands up in protest against the tropes of the superhero genre it’s part of. The real story is of political speech in an insider rebellion.
Others do it even better. Showrunner Shonda Rhimes seems to delight in a particular form: In Bridgerton, the Queen of England is black and her nobility is multi-coloured, white, black and brown. Everyone acknowledges their origins. Kate Sharma speaks of her home in Bombay and Lady Danbury remembers her childhood as a princess in Sierra Leone. Their casting isn’t colour-blind, but what’s missing is any commentary on their position, for the conceit is unashamedly ahistorical. Pre-modern England was not like this, and the show makes a point of refusing to address its divergence from reality. There are no stock characters in the main cast. Almost everyone is well-rounded. By refusing to comment, the show indulges in a form of exposition I’ve not seen before: it makes the audience question why they were expecting a particular trope and what it says of their own biases. Rhimes does this again in The Residence, a homage to detective characters set in the White House. The detective’s exposition is exaggerated this time, all tell-don’t-show, acknowledging the genre, but other tweaks lurk in the shadows. In the finale, a cleaner and waiter are seated on the couch while the President and his husband stand behind them, all of them on trial for murder. The sheer unlikeliness of any of these three conceits is again ignored, left to the audience to ask themselves why this isn’t normal.
Great filmmakers wield tropes to create a meta narrative that layers depth into their stories. At this point, I have to ask: is this not a language with a grammar? They communicate and we understand, with meaning flowing over the substrate of sound and visuals that fill each frame. Those frames don’t have the story, they’re just the medium.
Storytelling is as old as our species, and our spoken and written languages may be ancient, but cinema and TV are young, barely a century old, and the rapid evolution of language in these two media – overlapping but distinct – is plainly visible for anyone to see without any formal training. How is this not a real language?
This question coalesced for me when I went to an AI conference over the weekend and heard everyone talking about training their Large Language Models on underrepresented languages. What did they mean by “language”, I wondered. A question this naive could mark me as an imposter, someone who didn’t belong in this conversation. I had worried for this, so I indulged in another trope of such gatherings: I put on a business suit, thereby shifting my category from imposter to important. “What do you mean by language,“ I asked, referencing the Deus Ex Machina problem in visual media that shapes their stereotypes. I got awkward stares. Some told me later that they were still thinking of it, so I assume the suit worked. I took it off later over dinner because it was getting goddamned stuffy.
Early search engines indexing the web had to solve for identifying what topic was being discussed in each page. The technology for interpreting language wasn’t there yet, so they used simple heuristics: if a term appears multiple times in a page, it’s probably the topic. A naive implementation will blow up quickly because English grammar requires the mandatory use of filler words, articles, pronouns and prepositions: a, an, the, he, him, she, her, it, they and so on. If frequency alone counts, absolutely every piece of text in the English language is on the topic of English grammar. Search engines have to filter out these words – they’re called “stopwords” now – to read past them for the actual topic of the page.
Sometimes a page assumes its readers know where they are and doesn’t mention the topic at all, diving straight into arcane detail. The topic is only mentioned on another page that links to this page. In 1996, computer scientists Larry Page and Sergey Brin developed a search algorithm that emphasised these links between pages, dubbed PageRank, a marked improvement over previous search engines. Their company Google has since had its name turned into a verb for search, the vanilla flavour of search if you prefer. Pay attention the next time you see a character on screen doing a web search.
Stopword removal continues to be an essential feature of every search engine, and it continue to confound the very people who make search engines. Early computer programmers had to be frugal and named their languages after the letters of the alphabet. A, B and then C, which became a hit. A derivative language was named C++ after a particular idiom in the language, and a later one C#, but that symbol isn’t for a hashtag but the musical notation sharp. If you typed C# language, did you misspell C #language? The stopword filters in search engines strip out very small words and punctuation so it only sees language here, losing the proper noun, but if programmers can’t find their own documentation, they have to add those particular exceptions to their stopword filters. Tech and its users tug at each other. A later community around the Go language blinked first, adopting the term “Golang” for their identity.
PageRank has long been obsolete, for opportunists quickly discovered that they can game the algorithm to create a network of pages referencing each other. If you operate a website of any notability, your inbox must be full of people very politely asking if they can write for you – for free! – relentlessly, even now in 2025. Some vestige of PageRank still remains to be gamed.
If a search engine has to infer what you mean from the terms you’ve offered and find relevant pages, and is somehow still able to do that reasonably well despite the avalanche of slop fed to the engine every day, then you have to agree that the quest for meaning buried in the content is real and has been an ongoing technological quest for decades.
I was therefore surprised when someone at the conference claimed LLMs still process content one character at a time. It’s been a popular gimmick to ask an LLM for the number of R’s in the word Strawberry – they invariably get it wrong because they don’t see the word as a sequence of letters. Surely they’re tokenizing by word and not by letter? I don’t know who’s right because I’m unable to keep up with the pace of LLM development. I can either follow that or get some work done, and I need to work. That comment, however, contrasted too much with the discussions on training LLMs in other languages. What do they mean by language, again?
The current crop of LLMs – ChatGPT, Llama, Claude, DeepSeek, etc – are all trained on data scraped off the internet, but the internet is mostly words, images, sounds and videos, and this is a tiny fraction of the corpus of language. If your community isn’t thriving on the internet, your corpus isn’t already in an LLM and you have to catch up by training one. But what does your language even look like? Are words and sounds sufficient to capture the knowledge held in your language? Are body movements language? Sign language is indeed a real language, so is dance not a language? Does it only become a real language when it has a formal written notation? Surely these are important questions that deserve examination, alongside our current obsession with feeding LLMs ever more data?
“Philosophical,” I heard someone mutter, so let’s have another example: in the English statement “What time is it?”, what does the “it” refer to? There’s no subject here but the grammar requires one, so “it” is a placeholder reference to nothing. Stopword filters indiscriminately remove such words, but a language-aware parser will be better at identifying specific instances that are irrelevant.
My rambling thesis on this page is that cinematic language, in grappling with the Deus Ex Machina trope, found a solution in using stereotyped generics as placeholders because their visual grammar needs a subject to be present, even if the subject is irrelevant. Their vast corpus is now littered with these stereotypes, occupying frame after frame.
If an LLM actually understands language, it’ll ignore these placeholders. If it doesn’t however – instead consuming this data stripped of meaning – all of this junk will appear as the dominant theme of the data and the LLM will stubbornly insist on reinforcing the stereotypes.
We have already tested this. Ask any LLM to draw a clock showing a particular time and it’ll give you one showing 10:10. At some point in the past, clockmakers and watchmakers all standardised on 10:10 in their marketing images, presumably because the hands resemble a smile and stay clear of the dials behind them. If it’s now promoted with any other time, that time itself becomes a topic. An advertisement for an expensive watch is meant to sell you an investment into a family heirloom that also looks good on your wrist, not to distract you into pursuit of a conspiracy theory. Clockmakers have unwittingly locked themselves into one standard time, and it wasn’t a problem until the current generation of LLMs got fixated on the idea that 10:10 is the only time that a clock can show.
LLM makers already know this. Their technology is still in the indiscriminate stopword removal phase of sophistication, so a large chunk of their capital pays for an army of trainers tasked with nudging their models away from stereotypes. Presumably, they too have a reference list to work through, and 10:10 on the clock is one of those lucky instances where the public caught one first. LLMs would be unusable without their labour, but now you see the contradiction in the claim of LLMs training on the entirety of human knowledge via the internet: if the technology has an inherent tendency to pick up the wrong ideas and has to be manually trained away from it, this secondary training is heavily curated, and you can bet the curation steers in the direction of whatever markets the capital investors are interested in. It is in no way an objective container of human knowledge. This makes it particularly hard for anyone attempting to use the technology for an underserved language. They don’t have the capital for the manual curation, don’t necessarily benefit from prior training on other models, possibly don’t even have standardised LLM-friendly notation for their corpus of knowledge, and worse, are in an environment where the depth of this problem is unacknowledged, hidden below the froth of rah-rah cheerleading, “after all what is the human brain but a biological LLM”.
Let’s hope the technology at least advances from letters and words to metaphors, tropes, narratives and meta narratives. The human understanding of language is vastly richer than what a puny LLM can comprehend.
PS: Want another trope? If I somehow invoke New Amsterdam or TV with my daughter again, right here at the bottom, it’s called a bookend.