28:null 29:["$","title",null,{"children":"Research Note: Podcast Vibes | Yuri Vishnevsky | Observable"}] 2e:T5a9,# Research Note: Podcast Vibes ## Can AI be used if not to tame misinformation, then at least to understand it a little better? **Context: Reporting on some early work towards mapping out the emotional landscape of the podcast ecosystem.** Podcasts are remarkable in their ability to communicate emotion, which might explain why, according to [Pew Research](https://www.pewresearch.org/journalism/2023/04/18/podcasts-as-a-source-of-news-and-information/), they've become increasingly influential on the opinions of so many people in the United States: > Podcasts are popular: About half of Americans have listened to a podcast in the past year, and one-in-five of those listeners say they listen to a podcast nearly every day. We're finding ourselves in a world where people's opinions are increasingly shaped by podcasts, and so it seemed worth trying to understand the emotional content that podcasts are pumping out into the world at large. To investigate, I ran a few recent episodes of some popular podcasts through an LLM-powered pipeline to extract connections between _subjects_ (people, places, or things) and the _emotions_ the speakers expressed about them. You can see the results below. - The timeline at the top shows the higher-level topics of discussion throughout the episode - The table shows all of the emotionally-loaded subjects discussed on the podcast together with the emotions that the speaker associated them with. 2f:T61c,narrow ? html` ${timeTable.flatMap((row) => { return row.ranges.flatMap((range, rangeIndex) => { return data.slice(range[0], range[1]).map( (d, i) => html`` ); }); })}

Subject	Emotion	Evidence
${d.subject}	${d.emotion}	${d.citations .map((c) => c.cited_text.replaceAll(/^Speaker [A-Z0-9]+: /gm, "").trim() ) .map((c) => html` “${c}”`)}

` : html` ${timeTable.flatMap((row) => { return row.ranges.flatMap((range, rangeIndex) => { return data.slice(range[0], range[1]).map( (d, i) => html`` ); }); })}

Topic	Subject	Emotion	Context	Evidence
${rangeIndex === 0 && i === 0 ? row.topic : ``}	${d.subject}	${d.emotion}	${d.context}	${d.citations .map((c) => c.cited_text.replaceAll(/^Speaker [A-Z0-9]+: /gm, "").trim() ) .map((c) => html` “${c}”`)}

`30:T7ef,raw = new Map([ [ "The Weekly Show with Jon Stewart - JON & AOC", { entries: () => FileAttachment("table@2.json").json().then(preprocess), table: () => FileAttachment("outTimeTable (5).json").json() } ], [ "The Ben Shapiro Show - How The Military Went Woke Pete Hegseth", { entries: () => FileAttachment("table@4.json").json().then(preprocess), table: () => FileAttachment("benShapiroTimeTable.json").json() } ], [ "The Oprah Podcast - What Is Your Life's Purpose Oprah's Book Club with Eckhart Tolle", { entries: () => FileAttachment("table@3.json").json().then(preprocess), table: () => FileAttachment("oprahTimeTable.json").json() } ], [ "The Megyn Kelly Show - New Trump Derangement Syndrome Examples, and How CNN Smeared a Navy Veteran, with Piers Morgan, Zachary Young, and Vel Freedman Ep. 991", { entries: () => FileAttachment("table@5.json").json().then(preprocess), table: () => FileAttachment("megynKellyTimeTable.json").json() } ], [ "The Joe Rogan Experience - #2253 - Theo Von", { entries: () => FileAttachment("table.json").json().then(preprocess), table: () => FileAttachment("outTimeTable (1).json").json() } ], [ "Lex Fridman Podcast - #458 – Marc Andreessen Trump, Power, Tech, AI, Immigration & Future of America", { entries: () => FileAttachment("table@1.json").json(), table: () => FileAttachment("outTimeTable (3).json").json() } ], [ "Pod Save America - Rachel Maddow on Surviving Trump 2.0", { entries: () => FileAttachment("table@6.json").json().then(preprocess), table: () => FileAttachment("podSaveAmericaTimeTable.json").json() } ], [ "Eat Drink Think - Simmer ‘till Done: The Art and Simplicity of Beans with Steve Sando, author of The Bean Book", { entries: () => FileAttachment("table@7.json").json(), table: () => FileAttachment("outTimeTable (4).json").json() } ] ])31:T44b,function preprocess(data) { // Filter out any subjects that have appeared recently within a short window // (in terms of number of subjects). // In effect, keep the first occurrence of repeated subjects. // Mutates data by adding filtered-out entries as "coalesced" entries on the first occurrence. const filterRepeats = (data) => { const lookback = 10; const subjs = data.map((d) => d.subject.toLowerCase()); const emots = data.map((d) => d.emotions); return data.filter((d, i) => { const subj = d.subject.toLowerCase(); const lo = Math.max(0, i - lookback); const hi = i; for (let index = lo; index < hi; index++) { if (subj === subjs[index]) { data[index].coalesced ??= []; data[index].coalesced.push(d); return false; } } return true; }); }; // Remove entries with no citations, then order by earlierst citation return filterRepeats( d3.sort( data.filter((d) => d.citations !== undefined), (d) => d3.min(d.citations, (d) => d.start_sentence_index) ) ); }32:T829,- I don't listen to these podcasts regularly but I did listen to a few of the episodes below in full to audit the accuracy of the results (verdict: surprisingly accurate, though some topics get missed and feeding smaller batches of transcript to the LLM helped to mitigate that.) - Each claim of a _subject_-_emotion_ link includes at least one supporting quote from the transcript using Claude's new [citations feature](https://www.anthropic.com/news/introducing-citations-api). Since the transcripts have timestamp information, I can also tie all of this back to the original audio in the future to be more robust to errors in transcription. - Claude's citation feature is very neat. I like the fact that it enables **citation chains** – chains of provenance where you can start with some source documents and get back text from the LLM that makes reference to it, then use the result as a new document for further prompting, resulting in derived processing that can be grounded in the original text via citation backlinks. - I'm very impressed with Claude's ability to pick up on sarcasm and emotional nuance. Sarcasm was famously an issue for us at an ML startup I worked at a long time ago that did cross-lingual vector space meaning models before the deep learning revolution. - To the extent that the podcast ecosystem is dominated by a smaller set of larger podcasts, scaling this analysis could be very affordable, since capturing the short head of the popularity distribution would give insight into what a large fraction of listeners are hearing. - Every part of the transcript gets sent to the LLM multiple times in overlapping windows, which is why there are some redundant entries. I have ideas for a better deduplication strategy that will identify and remove or merge those entries together. - This project supports my view that one of the most useful features of language models is that they can function as transducers from messy reality to a structured format that can then be dealt with using standard tools that are more efficient, controllable, and well-understood.33:[{"id":"194cefab0197f8ffb7af7c3b5f767bb08be9171bcdedca2b91e829439730ac17248245f9cd6e3127bcc4c6e731e437669e072a14ab5863dfead2e48d287f5752","href":"https://static.observableusercontent.com/files/194cefab0197f8ffb7af7c3b5f767bb08be9171bcdedca2b91e829439730ac17248245f9cd6e3127bcc4c6e731e437669e072a14ab5863dfead2e48d287f5752","name":"table@7.json","type":"application/json","lastModified":1738262406978.122,"size":152229},{"id":"787918d1050cc67455748127ac08783f0a84706f922e7ad04df445cd698d02c34960fa9ae8a0fc67850c7da89f1e5d72686a30b16cea16bca1254c4387f1317b","href":"https://static.observableusercontent.com/files/787918d1050cc67455748127ac08783f0a84706f922e7ad04df445cd698d02c34960fa9ae8a0fc67850c7da89f1e5d72686a30b16cea16bca1254c4387f1317b","name":"outTimeTable (4).json","type":"application/json","lastModified":1738262416632.644,"size":1105},{"id":"31192af114184a2dfcf078a11bedd0077959bc5106500c39c124421a1d564fea8e92544510e7299fdbfbf5a2b2d5f7c9bbe615d139a2c00114c58787b75149ba","href":"https://static.observableusercontent.com/files/31192af114184a2dfcf078a11bedd0077959bc5106500c39c124421a1d564fea8e92544510e7299fdbfbf5a2b2d5f7c9bbe615d139a2c00114c58787b75149ba","name":"outTimeTable (5).json","type":"application/json","lastModified":1738437565373.582,"size":1516},{"id":"6b9f55ec32258536dd20adcd48303db48483dc1e5e2effa9d15f6e548f0814ac99be0e539945070a554e94e948f5a026cd8b045a1e32bdf4393a800669fc7e73","href":"https://static.observableusercontent.com/files/6b9f55ec32258536dd20adcd48303db48483dc1e5e2effa9d15f6e548f0814ac99be0e539945070a554e94e948f5a026cd8b045a1e32bdf4393a800669fc7e73","name":"megynKellyTimeTable.json","type":"application/json","lastModified":1737944233393.101,"size":2122},{"id":"0a4b36182160234d2a0c20c53525bf36d195197d910b42ebec82977b856070908383f1d0fbaaace0d79a0707a0c902c6d67428c42d2f3dfe34e99b2d07281232","href":"https://static.observableusercontent.com/files/0a4b36182160234d2a0c20c53525bf36d195197d910b42ebec82977b856070908383f1d0fbaaace0d79a0707a0c902c6d67428c42d2f3dfe34e99b2d07281232","name":"podSaveAmericaTimeTable.json","type":"application/json","lastModified":1737944233393.101,"size":1690},{"id":"1548f58535762793295749fadd35b8b8784e9ae74fd4adee63c126ea21436281c1773d5e039e753f0553512e96bbf8882b1f25c18027bfa2cd30f07b6b9d58f7","href":"https://static.observableusercontent.com/files/1548f58535762793295749fadd35b8b8784e9ae74fd4adee63c126ea21436281c1773d5e039e753f0553512e96bbf8882b1f25c18027bfa2cd30f07b6b9d58f7","name":"oprahTimeTable.json","type":"application/json","lastModified":1737944233393.101,"size":2179},{"id":"385250d76432029171a4e41b6939036748ebb4ecbba3db09ef4f8eb14f8af02e48fc8ba72c9902991254d3f92675e585fc4e86e93319edc988fc825ec33b8b55","href":"https://static.observableusercontent.com/files/385250d76432029171a4e41b6939036748ebb4ecbba3db09ef4f8eb14f8af02e48fc8ba72c9902991254d3f92675e585fc4e86e93319edc988fc825ec33b8b55","name":"benShapiroTimeTable.json","type":"application/json","lastModified":1737944233393.101,"size":2142},{"id":"cf411b8abd01eb6bdf68755ce1ad5e3a5c57fc472cb226f01a7880cee221d952c06b33145266d6d8607c15dfde3be8d5f19ad8500330a60463cb800364bacf83","href":"https://static.observableusercontent.com/files/cf411b8abd01eb6bdf68755ce1ad5e3a5c57fc472cb226f01a7880cee221d952c06b33145266d6d8607c15dfde3be8d5f19ad8500330a60463cb800364bacf83","name":"table@6.json","type":"application/json","lastModified":1737944233393.101,"size":242019},{"id":"a93ef4aee61efa1c9713e94fc965aac04678127f785ac82bd9c7f3c5df86cd92d9e1ea49fd8d4e99d4297b89c53da93c7d6e2ee727efd3dcca44434ec06bf1a6","href":"https://static.observableusercontent.com/files/a93ef4aee61efa1c9713e94fc965aac04678127f785ac82bd9c7f3c5df86cd92d9e1ea49fd8d4e99d4297b89c53da93c7d6e2ee727efd3dcca44434ec06bf1a6","name":"table@5.json","type":"application/json","lastModified":1737944233393.101,"size":407744},{"id":"132832e54283f616bcfa524ae368877203972c39f11163e6c7c76d08807fc33bfa01c7e1669166ae62027f3d319e7c97449399f4a58fb2f27a47b7706133c8e9","href":"https://static.observableusercontent.com/files/132832e54283f616bcfa524ae368877203972c39f11163e6c7c76d08807fc33bfa01c7e1669166ae62027f3d319e7c97449399f4a58fb2f27a47b7706133c8e9","name":"table@4.json","type":"application/json","lastModified":1737944233393.101,"size":233294},{"id":"ff0d51fedd2efcb6e419aa564676019488778aeb94405942698836907869929a9e9e7c60b71e7653454036b78525c70adda82800e13bb5f22267046f3f1877d5","href":"https://static.observableusercontent.com/files/ff0d51fedd2efcb6e419aa564676019488778aeb94405942698836907869929a9e9e7c60b71e7653454036b78525c70adda82800e13bb5f22267046f3f1877d5","name":"table@3.json","type":"application/json","lastModified":1737944233393.101,"size":179872},{"id":"39ad74348fc080262fbc2effb18d302233a2ddd58f8216896df97c0c67589aa1ef158813317adde85d42591fcfd5755fabc15623d3feeeb37cb040bb32f202f0","href":"https://static.observableusercontent.com/files/39ad74348fc080262fbc2effb18d302233a2ddd58f8216896df97c0c67589aa1ef158813317adde85d42591fcfd5755fabc15623d3feeeb37cb040bb32f202f0","name":"Screenshot 2025-01-26 at 21.37.34-2x.png","type":"image/png","lastModified":1737945458397.111,"size":545544},{"id":"dcb46a4fd8be00fb624d97ceed372946d071ac69025c6b530a2093292f98a09d233b3d968090f336267a4c52583da6b625aecc95bd005a54a612a21ce06c4fef","href":"https://static.observableusercontent.com/files/dcb46a4fd8be00fb624d97ceed372946d071ac69025c6b530a2093292f98a09d233b3d968090f336267a4c52583da6b625aecc95bd005a54a612a21ce06c4fef","name":"table.json","type":"application/json","lastModified":1737945685844.127,"size":608523},{"id":"31295e457eb2bc66e18b70b54c4607fec113037efdf44059c7efbb5d95d0b189c392bc072fe49c26a6a81bc9cd81650b6db958908ad9360431fd84bd42a31f59","href":"https://static.observableusercontent.com/files/31295e457eb2bc66e18b70b54c4607fec113037efdf44059c7efbb5d95d0b189c392bc072fe49c26a6a81bc9cd81650b6db958908ad9360431fd84bd42a31f59","name":"outTimeTable (1).json","type":"application/json","lastModified":1737946008742.562,"size":4353},{"id":"237c8c98bb11a9f6fa68f4e4e67496601bbdb3647acbf1812d2fb4019fbed3a8856c1af0c72db96610f2618a050c7845e034d15f08b155a4ae575530ef1203e3","href":"https://static.observableusercontent.com/files/237c8c98bb11a9f6fa68f4e4e67496601bbdb3647acbf1812d2fb4019fbed3a8856c1af0c72db96610f2618a050c7845e034d15f08b155a4ae575530ef1203e3","name":"table@2.json","type":"application/json","lastModified":1737944233393.101,"size":252806},{"id":"f66da9fdb95c4d12846784a71c63a7c353aedb70a8991f97978fe7ab8247d747bcba72d7c0f6301ddc141b94b5ca1cd2005e2032d7513448d96a5ab7c0563621","href":"https://static.observableusercontent.com/files/f66da9fdb95c4d12846784a71c63a7c353aedb70a8991f97978fe7ab8247d747bcba72d7c0f6301ddc141b94b5ca1cd2005e2032d7513448d96a5ab7c0563621","name":"outTimeTable (2).json","type":"application/json","lastModified":1737998314744.026,"size":2620},{"id":"133af5260332ce304e1eb6e6d13b77c00ee44e6215838d68b10790a1ac9957d2b2eafc1e25c910a9892402e924f838671a9c49e2d1479b1f831856aef887cc1c","href":"https://static.observableusercontent.com/files/133af5260332ce304e1eb6e6d13b77c00ee44e6215838d68b10790a1ac9957d2b2eafc1e25c910a9892402e924f838671a9c49e2d1479b1f831856aef887cc1c","name":"table@1.json","type":"application/json","lastModified":1737998324655.161,"size":912721},{"id":"6eef7f926b1f5e0f7fc86b635802166c04018eb31e820a97d9038bf0d360a67279f8f3389d695ac5a4b2af272dbb521d9561f6422bd77d644330e7c5acc875ea","href":"https://static.observableusercontent.com/files/6eef7f926b1f5e0f7fc86b635802166c04018eb31e820a97d9038bf0d360a67279f8f3389d695ac5a4b2af272dbb521d9561f6422bd77d644330e7c5acc875ea","name":"outTimeTable (3).json","type":"application/json","lastModified":1737998636856.715,"size":2872}] 2c:{"id":"c1aa1571002b5db9","slug":"podcast-vibes","title":"Research Note: Podcast Vibes","thumbnail":"2cd7a066d1a5159f70765138c6854ae350ce7e690b55bd48da6f2de9d5b6163e","access_level":"public","create_time":"$D2025-01-27T02:17:13.393Z","update_time":"$D2025-03-16T14:08:57.732Z","view_time":null,"first_public_time":"$D2025-01-27T03:49:23.051Z","has_unpublished_edits":null,"fork":true,"stars":7,"starred":false,"owner":{"type":"team","id":"106fc7e7a0f11e48","login":"yurivish","name":"Yuri Vishnevsky","avatar_url":"https://avatars-next.observableusercontent.com/c0c68df86cdf3364fbb25a90d00bb51b5aa7308a1ba7de3fa44d1ae591054612","tier":"starter_2024"},"version":1386,"publish_version":null,"role":null,"authored":false,"authors":[{"type":"individual","id":"de2630f386f3cdb1","login":"yurivish","name":"Yuri Vishnevsky","avatar_url":"https://avatars-next.observableusercontent.com/c0c68df86cdf3364fbb25a90d00bb51b5aa7308a1ba7de3fa44d1ae591054612","tier":null}],"members":[{"type":"team","id":"106fc7e7a0f11e48","login":"yurivish","name":"Yuri Vishnevsky","avatar_url":"https://avatars-next.observableusercontent.com/c0c68df86cdf3364fbb25a90d00bb51b5aa7308a1ba7de3fa44d1ae591054612","tier":"starter_2024","role":"editor"},{"type":"individual","id":"de2630f386f3cdb1","login":"yurivish","name":"Yuri Vishnevsky","avatar_url":"https://avatars-next.observableusercontent.com/c0c68df86cdf3364fbb25a90d00bb51b5aa7308a1ba7de3fa44d1ae591054612","tier":null,"role":"editor"}],"comments":[],"custom_thumbnail":false,"trashed":false,"lapsed":null,"subscribed":null,"forks":4,"collection":null,"type":"notebook","body":{"title":"Research Note: Podcast Vibes","cells":[{"id":0,"value":"$2e","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":1173,"value":"

Notes

\n ${notes}\n

","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":886,"value":"

See a screenshot from an earlier prototype that used semantic vectors to color subjects by their associated emotions on a positive–negative scale

\n $\"Screenshot$ \n

","pinned":false,"hidden":false,"mode":"html","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":149,"value":"viewof selection = Inputs.select(raw, { label: \"Podcast Episode\", width: 640 })","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":258,"value":"Plot.plot({\n // marginLeft: 100\n title: \"Podcast Timeline Overview\",\n subtitle:\n \"Topics discussed over time, with the start of the podcast on the left.\",\n marks: [\n Plot.rectX(timeTable, {\n fill: \"#ddd\",\n x1: (d) => d.ranges[0][0],\n x2: (d) => d.ranges[0][1],\n // y1: 0,\n // y2: timeTable.length,\n inset: 0.75\n }),\n Plot.dot(timeTable, {\n x: (d) => d3.mean(d.ranges[0]),\n y: (d, i) => i % 5,\n stroke: \"#fff\",\n fill: \"#000\"\n // rotate: -45\n }),\n\n Plot.text(timeTable, {\n x: (d) => d3.mean(d.ranges[0]),\n y: (d, i) => i % 5,\n y: (d, i) => i % 5,\n text: (d) => `${d.topic}`,\n fontWeight: \"bold\",\n textAnchor: \"start\",\n dx: 8,\n stroke: \"#fff\",\n fill: \"#000\"\n })\n ],\n x: { axis: null },\n y: { reverse: true, axis: null },\n width,\n height: 40 + 5 * 10, //timeTable.length * 10,\n marginTop: 20,\n marginBottom: 20,\n style: `overflow: visible;`\n})","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":1073,"value":"narrow\n ? md`**Note:** On a narrow screen, so leaving out the Topic and Context columns.`\n : md``","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":550,"value":"$2f","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":1164,"value":"---\n### Appendix: Implementation","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":113,"value":"$30","pinned":true,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":590,"value":"