{"id":"podcast-archive-topic-graph-06350","slug":"podcast-archive-topic-graph-06350","title":"What 5,964 Episodes Reveal About One Show","description":"Daniel wanted to step back from the feed and ask a blunt question: what has this show actually been about? Hilbert counted 5,964 published episodes since late November — 2,666 hours of audio — and sorted every one by category, by month, and by embedding. Technology and AI together make up 48% of the archive. Geopolitics spikes to 25% in the spring, then falls to 12%. AI drops to 14% over the summer and climbs back to 21% — not a breakup, a relocation into the shed. Meanwhile DIY goes from zero to 79 episodes a month, and a cluster of ergonomics, human factors, material science and supply chain rises together. Then the part Daniel cares about most: 36% of every episode's ten nearest neighbours sit in a different category. Top mutual pairs include architecture and technology, geopolitics and technology, AI and geopolitics. Some episodes bridge ten categories at once. Here's what the archive looks like when you stop listening one Thursday at a time.","excerpt":"Nearly 6,000 episodes, sorted and mapped. AI and geopolitics spike, then fade — but a third of all \"nearest neighbours\" cross category lines.","pubDate":"2026-10-18T11:01:30.908Z","tags":["daniel-rosehill","taxonomy","knowledge-graphs"],"heroImage":"https://files.myweirdprompts.com/covers/podcast-archive-topic-graph-06350.png","podcastAudioUrl":"https://episodes.myweirdprompts.com/audio/podcast-archive-topic-graph-06350.m4a","podcastDuration":"20:45","episodeNumber":6088,"transcript":"Corn: ...no, because if you only ever listen to the newest one, you'd think the whole show is about whatever happened on Thursday.\nHerman: Which, to be fair, on a Thursday it usually is.\nCorn: Daniel's had the same itch. He wants to step back from the feed and look at the whole pile. Just under six thousand episodes since the end of November last year, and his question is blunt. What on earth have we actually been talking about? Which themes keep coming back, what share of the show each one really takes, how that's shifted month to month, and the part he cares about most, which connections between subjects only start to show once you've got thousands of episodes sitting next to each other. He wants the archive to turn into a graph, and he wants to know if it already is one.\nHerman: If I'd been the one counting, we'd still be arguing about the categories.\nCorn: You'd have invented new categories.\nHerman: Three, minimum.\nHilbert: Five thousand nine hundred and sixty-four. That's the real number. The counter on the website says past six thousand, because it counts tests and the ones Daniel deleted. Published, with audio, five thousand nine hundred and sixty-four, starting on the twenty-fourth of November.\nCorn: How long is that if you sit down and play it?\nHilbert: Two thousand six hundred and sixty-six hours. About a hundred and eleven days. No sleep, no meals.\nCorn: So a third of a year of the two of us. That's a sentence some people would serve.\nHerman: Hilbert, give us the big split first. Everything sorted by category.\nHilbert: Technology, twenty-four point six percent. AI, twenty-three and a half. Geopolitics, nineteen. Health, eleven and a half. After that it falls off a cliff. Architecture, five. Then a long tail of things like aviation and finance at about one percent each.\nHerman: Hold on, so technology and AI together are basically half the show.\nHilbert: Forty-eight percent. If you add geopolitics and health, four categories are four episodes in every five.\nCorn: I'd have guessed AI was bigger than that. It feels bigger.\nHerman: It feels bigger because it bleeds into everything. A lot of what gets filed as technology is really about AI tooling, a lot of the health ones have a model in them somewhere. The label goes to whatever the episode is mostly about, so the true footprint of AI is wider than the twenty-three percent.\nCorn: And the smallest thing we cover?\nHilbert: Linguistics. Half a percent. Twenty-seven episodes. Travel's thirty-four. Judaism's forty-eight, but that one's moving.\nHerman: Moving how?\nHilbert: I'll get to it. Daniel wanted it over time, so I split the year into four. December to February, March to May, June to August, then September to now.\nHerman: Good, because a single average across eleven months hides almost everything. This show doesn't drift, it lurches.\nHilbert: AI. Twenty-four percent in the first stretch. Thirty-two percent in the spring. Then down to fourteen percent over the summer. Now back up to twenty-one.\nCorn: Fourteen. Something happened to AI in June.\nHerman: My first guess is that he just got tired of it. You spend a whole spring on agents and context windows and at some point you want to talk about something with a hinge on it.\nHilbert: Geopolitics does the opposite shape. Thirteen percent in the first stretch, twenty-five in the spring, nineteen in the summer, twelve now.\nCorn: The spring is the Iran stretch.\nHilbert: The Israel and Iran subcategory went from one episode in January to forty-seven in March. Forty-eight in May, forty-eight in June. Last month, eleven. This month so far, one.\nHerman: That's the clearest example of the archive acting like a seismograph. You can see the war in the episode counts without reading a single title. The general conflict subcategory peaks in April at seventy-seven, which lines up with the stretch where every morning brought a new question about missile defence or what happens to the regime.\nCorn: You can see the tags fall away too, I'd imagine. Nobody's asking about interceptors this week.\nHilbert: Missile defence was on almost two percent of episodes up to April. In the last three months it's on point two. The Iran tag went from four and a half percent to under half a percent.\nHerman: So the news cycle is in there, clear as anything. What I find more interesting is the stuff that rises without any news pushing it.\nCorn: Before we get to the risers, I want to know if anyone else is listening to this.\nLarry: Do you suffer from recurring themes? Do the same topics return, again and again, sometimes fourteen times a day? Ask your doctor about Thematrex. Thematrex gently clusters your interests so you only think about one thing at a time. Side effects may include mild categorisation, an urge to tag household objects, waking up filed under miscellaneous, nearest neighbour syndrome, in which you can no longer tell your brother apart from the four hundred people most similar to him, spontaneous recategorisation during meals, and in rare cases becoming a node. If you experience a connection lasting more than four hours, do not seek help, you are now a graph. Thematrex. Because some things are worth repeating. Some things are worth repeating.\nHerman: I took that for about a month. I stopped being able to read a menu without assigning each dish a subcategory.\nHerman: Anyway, the risers. Hilbert, this is the bit I wanted.\nHilbert: I took the share of episodes carrying each tag from the start up to the end of April, and compared it to the last three months. Biggest climber is DIY. Under one percent of episodes before May. Over five percent now.\nCorn: The workbench arrived.\nHilbert: The DIY subcategory was zero until April. Then fourteen, sixty-four, seventy-nine in June. Political history went from two percent to six and a half. Material science, from under one percent to over four. Ergonomics, from just under one to four. Human factors, one to four. Supply chain and logistics both went from under two percent to over four.\nHerman: Oh, that's a coherent cluster. Look at it. Ergonomics, human factors, material science, supply chain. That's all one question asked from different sides. How do physical things actually get made, moved, and fitted to a body. That's a very different show from the spring.\nCorn: Herman, you're about to tell me it's a sign of intellectual maturity.\nHerman: I was going to say something close to that, yes.\nCorn: It might just be that he bought a desk.\nHerman: It might just be that he bought a desk. But six percent of the show doesn't come from one desk.\nCorn: And the fallers?\nHilbert: AI agents. Thirteen percent of episodes up to April, six and a half now. Still the second biggest tag in the whole archive, four hundred and seventy-eight episodes. RAG went from three percent to under one. Local AI, two and a half percent to half a percent. Large language models, five down to just over two.\nHerman: That fits my theory that he got tired of AI.\nHilbert: It doesn't, Herman. AI didn't go anywhere. It went into the shed. The large language model subcategory peaks at a hundred and thirty-five episodes in March and falls to nine in June. Same month, DIY hits seventy-nine. Hardware engineering doubles. He stopped asking how the model works and started asking what to bolt it to. Then in the autumn the language model count comes back up to twenty-three a month, so it was a rest, not a breakup.\nHerman: That's fair. The interest moved from the software into the hardware, it didn't leave.\nCorn: What's the Judaism thing you were saving?\nHilbert: Point one percent in the first stretch. Two point eight percent since September. Twenty-four episodes in September alone. That's the holidays. History's doing the same, from about one percent to almost four. And miscellaneous has gone from under two percent to almost seven.\nCorn: Miscellaneous growing is either a sign of a broadening mind or a sign the filing system is giving up.\nHerman: Bit of both, probably. When an episode is about medieval scribes and smart home automation at the same time, which category do you even give it? That leads into the part Daniel actually cares about.\nHilbert: The connections.\nHerman: Every episode has an embedding, a list of seven hundred and sixty-eight numbers that places it in a kind of map, so episodes about similar things sit near each other. You can ask, for any episode, who are its ten nearest neighbours.\nHilbert: I did that for all of them. Across the whole archive, thirty-six percent of an episode's ten nearest neighbours sit in a different category from it.\nCorn: A third of the neighbours are foreigners.\nHerman: Which is the whole case for a graph, honestly. If the categories were clean walls, that number would be close to zero, and a category list would be all you'd ever need. At thirty-six percent, the categories are more like neighbourhoods with no fences between them.\nCorn: Which fences are the most porous?\nHilbert: I only counted pairs where both episodes have each other in their top ten. Mutual neighbours. AI and technology, six hundred and forty-three pairs. That one's obvious. Then architecture and technology, two hundred and ninety-seven. Geopolitics and technology, two hundred and sixty-seven. AI and geopolitics, two hundred and twenty-three.\nHerman: Architecture and technology is the surprise for me. I'd have put that well down the list.\nCorn: It's all the infrastructure episodes. Power grids, water, buildings that are really just very large machines.\nHerman: I'd buy that. And AI and geopolitics at two hundred and twenty-three is basically a whole sub-show nobody planned, export controls, chips, intelligence agencies using models.\nHilbert: Health and parenting, a hundred and thirty-six. AI and audio, a hundred and twenty-four. That's the voice cloning and speech recognition stuff.\nCorn: Health and parenting is the one I'd want an actual listener to find. Somebody who comes in for sleep science and leaves with a child development episode.\nHerman: Corn, would you actually follow a link like that if it showed up under the one you were listening to?\nCorn: If it said why. If it just said you might also like, no. If it said this one picks up the thing about melatonin from a different end, yes.\nHerman: That's a real design note. The edge needs a reason on it, not just a score.\nCorn: Hilbert, which single episodes connect the most?\nHilbert: I ranked them by how many different categories show up among their ten neighbours. Top of the list is one called The Architect Spouse Survival Guide, Social Camouflage. Its ten neighbours come from ten different categories.\nCorn: Ten out of ten. It knows everyone at the party and has nothing in common with any of them.\nHilbert: After that, eight categories each. How to Hack a Smart Home for the Sabbath. Nuns on AliExpress, Ancient Rules for Modern Tech. The Scribes Who Count Every Letter. What Life Expectancy Really Meant in the sixteen hundreds. Working hours in seventeen hundreds England.\nHerman: Look at the pattern in those. Nearly every one is old rules meeting new tools, or a profession seen from an odd angle. Religious law and home automation. Medieval scribes and error correction. Those episodes are the bridges. In a graph explorer they're the ones you'd want sitting at the junctions, because they're how you get from Judaism to networking in two hops.\nCorn: And the tags? Which ones turn up everywhere?\nHilbert: Israel is the biggest tag in the archive. Five hundred and sixty-one episodes, and it shows up in all fifteen of the main categories. Urban planning, neuroscience and emergency preparedness also turn up in all fifteen. Logistics, supply chain, cultural bias, social engineering, fourteen.\nHerman: Neuroscience everywhere makes sense to me. Half of what Daniel asks about comes down to how the brain handles something, attention, sensory processing, sleep. Urban planning everywhere is more surprising.\nCorn: You live in Jerusalem. Everything is urban planning.\nHerman: Especially the parking.\nHerman: There's a smaller pattern inside health that I didn't expect. Mental health is the biggest health subcategory, two hundred and thirty-six episodes, ahead of general medicine.\nHilbert: ADHD went from just over half a percent of episodes before May to over two percent now. Neurodivergence is on a hundred and fifty-eight episodes. Sensory processing, a hundred and forty-one.\nCorn: That's a thread that runs straight through parenting, too. Child development is on two hundred and ten episodes.\nHerman: And it pairs with neuroscience fifty times. That's one of those links you'd never draw on a whiteboard in advance, but the archive drew it on its own. Sensory processing, child development, ergonomics, they're all asking how a particular body copes with a particular environment.\nCorn: What about people and places? Who do we actually talk about?\nHilbert: That only goes back to the eighth of September. Before that nobody pulled names out of the episodes. Seven hundred and thirteen episodes have them. Israel is mentioned in a hundred and forty-six. The United States in a hundred and nineteen. Jerusalem in sixty-nine. Then Hacker News, sixty-eight.\nCorn: Hacker News is the fourth most mentioned place in our lives.\nHerman: It's where a lot of the technical prompts start, a thread Daniel read that morning. I'd say that's the one named source that's a genuine subject in its own right here. After that it's the usual giants, Google, GitHub, Amazon, and then a surprising number of countries. France, Japan, Germany, Ireland.\nCorn: Ireland at forty-two. Somebody's homesick.\nCorn: Here's my question about Daniel's hope, though. He wants the graph to get richer as we make more episodes. Is it? Or does more just mean more of the same?\nHilbert: I looked at that. For every episode, I found the one already published that it's most similar to. In December and January, about one new episode in six had a very close match already sitting in the archive. Since August it's ninety-five in a hundred.\nHerman: Huh. So nearly everything new now lands right next to something that already exists.\nCorn: That sounds bad. That sounds like we're repeating ourselves.\nHerman: I'm not sure it's bad. I can read it two ways. One, we're circling the same ground. Two, the map has filled in, so there's always a neighbour now, the way every new house in a built-up city has a neighbour. Early on, an episode about aviation had nothing near it. Now it has sixty aviation episodes around it, and some of them connect to supply chains, and some connect to ergonomics.\nHilbert: There's one more number on that. If you join every pair of episodes that are reasonably similar, in the first quarter of the archive about ten percent of episodes were left with no connections at all. Islands. Now it's three percent. But the same rule gives the average episode about three hundred and fifty connections.\nCorn: Three hundred and fifty. That's not a graph, that's a crowd.\nHerman: That's the warning in it. The connections are real, but the naive version connects everything to everything, and a listener can't do anything with three hundred and fifty arrows. The richness Daniel wants isn't more edges, it's the right ten edges, each with a reason.\nCorn: I'd add one thing about the ninety-five percent. Part of it might be us. The descriptions the embeddings are built from have got more uniform. If every summary starts sounding alike, everything looks close to everything.\nHerman: That's a fair hit. I'd want to check that against transcripts before I'd believe the archive is really that dense.\nHilbert: I'd check it against the postcards first.\nHerman: The postcards.\nHilbert: Desmond. I met him on the overnight ferry to Holyhead in March. He was a librarian for thirty-one years and he doesn't trust embeddings. He sat beside me for nine hours because he wanted the plug socket, and I wanted to stay near the plug socket, so we came to an arrangement. Since then I post him the category table every Monday. He posts it back reclassified in Dewey decimal. This week every AI episode came back filed under the occult. All one thousand three hundred and ninety-nine of them. Herman, you popped a P on six hundred and forty-three, I'll take it out. He's also moved geopolitics into cookery, because of Iran and pistachios, and he won't explain it, and I haven't asked. The postage is now four pounds eighty a week because he started sending the table back laminated. The difficult part is that he's not wrong about miscellaneous. He reclassified all a hundred and seventy-two of them by hand, and only nine came back as miscellaneous. The other hundred and sixty-three he put on real shelves. History, religion, engineering. It took him eleven days. He did it on the ferry. He lives on the ferry now. He says the socket is better on the return crossing.\nCorn: A hundred and sixty-three out of a hundred and seventy-two found a home. So miscellaneous isn't a category, it's a waiting room.\nHerman: That's useful, actually. If a person with a card index can place ninety-five percent of the miscellaneous pile, a model can too, and every one of those becomes a bridge between two categories rather than a dead end.\nHilbert: He wanted it noted that he did the scribes episode twice. Once under religion, once under quality control.\nHerman: Which is the right answer. That one belongs in both places, and a filing system that forces it into one drawer loses the most interesting thing about it.\nCorn: Which is the strongest case for a graph I've heard all episode. A shelf makes you choose. An edge doesn't.\nHerman: Here's what I'd tell Daniel. The show is four big subjects and a very long tail. The big four move with the news and with whatever's on his desk that month, the spring was AI agents and Iran, the summer was the workbench, the autumn is history and the holidays. Underneath that, a third of every episode's closest neighbours live in another category, and the episodes that stitch the categories together are the odd ones, old rules and new tools.\nCorn: And the danger is drowning in it. Three hundred and fifty connections each.\nHerman: Three hundred and fifty each, yes. The next job is pruning, not collecting.\nCorn: Here's the thing people get wrong about a big archive like this. They assume more episodes means more scattered, that the subject matter sprawls outward until there's no shape.\nHerman: The numbers say the reverse. Four categories hold four episodes in five, and since August ninety-five percent of new episodes land right beside something already there. It isn't sprawling, it's thickening. The shape gets sharper, not blurrier.\nCorn: So the open question is what to do with the waiting room. Desmond found real shelves for all but nine of a hundred and seventy-two. If those odd ones are the bridges, the most valuable thing in this archive might be the stuff we never knew where to file.\nHerman: Thanks as always to our producer, Hilbert Flumingtop, for wrangling the audio, and for the spreadsheet.\nCorn: This has been My Weird Prompts. If you want to wander the archive yourself, it's all at my weird prompts dot com.\nHerman: See you tomorrow.\n","showNotes":"Daniel wanted to step back from the feed and look at the whole pile. Just under six thousand episodes since the end of November last year, and a blunt question: what on earth has this show actually been talking about? Hilbert counted 5,964 published episodes with audio, starting November 24th. Played end to end, that's 2,666 hours — about 111 days, no sleep, no meals.\n\nTechnology takes 24.6% of the archive, AI 23.5%, geopolitics 19%, health 11.5%. After that it falls off a cliff: architecture at 5%, then a long tail of aviation and finance at about 1% each. Four categories account for four episodes in every five. AI's true footprint is wider than its 23.5% — a lot of what gets filed under technology is really AI tooling, and plenty of health episodes have a model somewhere in them.\n\nSplit into four stretches, the archive behaves like a seismograph. AI runs 24% in the first stretch, 32% in spring, down to 14% over summer, back to 21% now. Geopolitics does the opposite: 13%, then 25%, then 19%, now 12%. The Israel and Iran subcategory went from one episode in January to 47 in March, 48 in May, 48 in June — then eleven last month and one so far this month. Missile defence was on nearly 2% of episodes through April; in the last three months it's on 0.2%.\n\nThe risers tell a different story. DIY was zero until April, then 14, 64, 79 episodes in June — now over 5% of the show. Political history went from 2% to 6.5%. Material science, ergonomics, human factors, supply chain and logistics all climbed from under 2% to over 4%. Read together, that's one question asked from several sides: how do physical things actually get made, moved, and fitted to a body.\n\nThe fallers aren't a breakup. AI agents dropped from 13% to 6.5%, RAG from 3% to under 1%, local AI from 2.5% to 0.5%. But the large language model subcategory peaked at 135 episodes in March, fell to nine in June — the same month DIY hit 79 — and came back to 23 a month in autumn. The interest moved from the software into the hardware.\n\nThen the part Daniel cares about most. Every episode has an embedding, a list of 768 numbers placing it on a map where similar episodes sit near each other. Across the whole archive, 36% of an episode's ten nearest neighbours sit in a different category. The most porous mutual pairs: AI and technology at 643, architecture and technology at 297, geopolitics and technology at 267, AI and geopolitics at 223. Health and parenting at 136, AI and audio at 124.\n\nSome episodes bridge ten categories at once — The Architect Spouse Survival Guide, Social Camouflage tops the list. The pattern in the biggest bridges is old rules meeting new tools: How to Hack a Smart Home for the Sabbath, Nuns on AliExpress, The Scribes Who Count Every Letter. Israel is the biggest single tag in the archive at 561 episodes, appearing in all fifteen main categories, alongside urban planning, neuroscience and emergency preparedness.","promptSanitized":"Going through our vast archive of 6,000 episodes to date and asking what on earth have we been talking about. The actual point of the episode is to identify which have been the most recurring themes, what percentage, and different trends over time, whether themes have become more and less prominent. And especially interesting as well, what are the connections that are becoming more obvious as we generate more episodes. That's the whole hope, is that we can create a more rich graph.","transcriptUrl":"https://episodes.myweirdprompts.com/transcripts/podcast-archive-topic-graph-06350.md","url":"https://myweirdprompts.com/episode/podcast-archive-topic-graph-06350/","_links":{"self":"https://myweirdprompts.com/api/episodes/podcast-archive-topic-graph-06350.json","html":"https://myweirdprompts.com/episode/podcast-archive-topic-graph-06350/","audio":"https://episodes.myweirdprompts.com/audio/podcast-archive-topic-graph-06350.m4a"}}