The claim

A short video by kylertalks, “are you a polymath?”, argues that the people who connect many fields will lead the next decade. The strongest reply is that polymath is a fancier name for an interdisciplinary expert and that renaming adds nothing. Two further questions come with the claim. Is a person’s own set of connections between fields different from the connections a system picks for them, the question the tasteslop essay asked about taste? And can those connections be read off a record of the person?

We took it as four questions: whether the word adds anything, whether connecting fields pays, how far the connection should reach, and whose connection it is. The last one we tested on Framework’s Ideas page.

Is a polymath more than an interdisciplinary expert?

Partly not. The sources define the word three ways. Root-Bernstein and Root-Bernstein use polymath and multidisciplinarian for the same thing, “active engagement in multiple diverse interests, vocational or avocational.” Peter Burke’s history of polymaths counts, in a reviewer’s words, people who made “original contributions to several (if not ‘many’) disciplines”. He lists about 500 people, which is a catalogue and not a measurement. A preprint posted in October 2026, not yet peer reviewed, defines polymathy as breadth, depth and integration together.

On the other side, the National Academies’ definition of interdisciplinary research, quoted by Porter and Rafols, is “a mode of research by teams or individuals that integrates perspectives/concepts/theories and/or tools/techniques and/or information/data from two or more bodies of knowledge or research practice.” That covers one person. A commentary on environmental research treats polymaths as a contributor to interdisciplinary work and not a rival to it (abstract only).

So the reply holds on definitions: the polymath is the one-person end of the range of interdisciplinary work, and no source we read measures a line between the two. The word adds two things the other term leaves open. It says the joining happens in one head, and it counts hobbies. Nobel laureates had on average 0.94 arts and crafts avocations, against 0.33 for Sigma Xi members surveyed in 1936 and 0.59 for Royal Society fellows, an odds ratio of 31.79 against Sigma Xi and 10.88 against the Royal Society (Root-Bernstein and colleagues). The groups come from different eras, nations and surveys, and the authors say the data cannot show the quality of the arts and crafts.

Counting fields is also a weaker measure than it looks. Porter and Rafols followed six research fields from 1975 to 2005. Articles cited about 50% more disciplines, yet an index of how far apart the cited fields are rose by about 5%, because citations stay in neighbouring fields. They write that “interdisciplinarity is not the opposite of specialization.”

Does connecting fields pay?

The best-known result is a pairing. Uzzi and colleagues scored 17.9 million papers from 1950 to 2000 on how unusual the pairs of journals each paper cites are. Papers whose references were mostly conventional pairs with a few unusual ones were the likeliest to land in the top 5% by citations: 9.11 in 100, against 5.82 for conventional only, 5.33 for unusual only and 2.05 for neither. Only 6.7% of papers sat in that best cell.

Breadth alone pays less cleanly. Fleming’s study of 17,264 US patents found that new combinations of technology components “lead to less useful inventions on average, but it also implies an increase in the variability that can result in both failure and breakthrough.” Leahey and colleagues, studying scientists at research centres, report “both penalties (fewer papers published) and benefits (increased citations)” from interdisciplinary work, “a high-risk, high-reward endeavor” (abstract only). Arts and Fleming followed 396,336 inventors of US patents from 1975 to 2002. When an inventor moved into a technology field they had not patented in, the new patent held 6.7% more combinations of technology subclasses never combined before and was cited 1.0% less; having an expert co-inventor muted the penalty. Hill and colleagues report a “pivot penalty”, in which impact “steeply declines the further a researcher moves from their previous work”, and find that larger pivots engage weakly with established mixtures of prior knowledge (abstract only).

The same shape appears outside science. Films that target more genres drew larger audiences and were rated lower by them (949 US films, 2000 to 2003; Hsu). Among 18,476 Australian research proposals, the more interdisciplinary the proposal, the lower the chance of funding (Bromham and colleagues, abstract; the data alone do not settle whether that is bias or weaker proposals). And among papers, those that cite almost only their own field or almost only other fields had low impact, with no clear correlation across all disciplines combined (750,743 papers from 2000; Larivière and Gingras).

Who does the connecting matters. Taylor and Greve followed 4,485 US comics from 1972 to 1996, and found that “individuals were able to combine knowledge diversity more effectively than teams”, though teams varied more in value. This is the nearest result we found to one person joining fields, and it is one industry. Wang, Veugelers and Stephan found that papers pairing journals never before cited together had 57% higher odds of reaching the top 1% by citations, but did worse over windows shorter than three years and better from the fourth year (661,643 papers from 2001; 89% make no new pair). The payoff is slow.

Pace decides who wins. In a natural experiment around the end of the Soviet Union, after which the pace of change differed between areas of mathematics, specialists made about 22% fewer citation-weighted publications than generalists in slow-paced areas and gained about 83% relative to generalists in fast-paced ones (Teodoridis and colleagues; here a generalist is a mathematician whose papers spread over many subfields of one discipline). Burke’s history, as the review reports it, is one of decline from the “monsters of erudition” of the 17th century to specialisation, “a kind of defence mechanism, a dyke against the deluge of information”, with new disciplines briefly making room for polymaths because their first teachers were trained in something else.

Read together: connecting fields pays for the few whose work stays rooted in a conventional base, it pays in the spread of outcomes and with delay, it costs volume and funding, and in one study the edge went to breadth when change was slow and to specialists when it was fast. Whether the next decade is slow or fast for a given field is the open part. The studies do not say it will be the polymaths’.

How far should the connection reach?

Working scientists’ own analogies are mostly near. Dunbar recorded 16 meetings in four molecular biology labs and counted 99 analogies: 40 to the same organism, 57 to another organism and 2 to a non-biological source, and both distant ones explained an idea and did not build a hypothesis. Far analogies do not arrive unprompted either. In Gick and Holyoak’s experiments, undergraduates who had read a fortress story solved the radiation problem 92% of the time (11 of 12) when told the story was relevant and 20% (3 of 15) when not. And there is a point past which far stops helping: Fu and colleagues report that far stimuli are more likely to lead to innovative solutions, but that “there is such a thing as too far” (abstract only; near and far are relative).

Machines can find links across fields that no one has made. Swanson found that the literatures on Raynaud’s syndrome (about 2,000 articles) and fish oil (about 1,000) had no article in common and had never been cited together, and proposed fish oil as a treatment. A placebo-controlled trial of 32 patients later found that in the fish-oil group the median time before an induced attack rose from 31.3 to 46.5 minutes at six weeks (p = 0.04), with the benefit in primary and not in secondary Raynaud’s. Embeddings trained on 3.3 million materials abstracts ranked unstudied materials, and the top 50 were on average eight times more likely than a random unstudied material to be studied as thermoelectrics within five years (Tshitoyan and colleagues; “studied” means appearing in the corpus’s abstracts, not tested). Sourati and Evans built models that simulate what experts could reach next and found predictions up to 400% better than content-only models, most where the literature was sparse. When their model was set to avoid what experts would reach, at a moderate setting its hypotheses were very unlikely to come from the scientific community and likely to succeed, and at a stronger one none of its COVID-19 predictions was found within the window.

So the connection that works is usually near, a far one usually needs a hint or a base to land on, and machines that connect fields do best when they model what the people in the field could reach.

Is a person’s own connection the opposite of a system’s pick?

Studies of what systems pick point one way. On Spotify, across more than 100 million users and about 70 billion streams, listening the users chose themselves was almost always more diverse than the listening the algorithm drove, and users who grew more diverse over a year cut their programmed streams (Anderson and colleagues; within one person’s listening). In an iTunes plug-in study with 858 matched users per group, recommendations widened each user’s range, 9,709 to 13,741 unique artists against 9,928 to 9,281 without, and raised the overlap between users: “Personalization … helps users widen their interests, which in turn creates commonality with others” (Hosanagar and colleagues; one recommender, not randomised, about half the users uninstalled it).

Language models show the same form for fields. Told to draw on other fields for a problem, three models still returned about 5% unique domains over 50 tries, and the cross-domain instruction “still collapses” (Shen, Druckmann and Zou, preprint; repeated samples of one problem). Ten models asked for analogies landed in about two domain clusters per ten generations, even when told to be diverse (Shen and colleagues, preprint). Writers using an instruction-tuned model wrote essays more like one another’s than writers working alone (38 writers; Padmakumar and He).

That is the tasteslop split again. What a system picks is shared across people, what a person picks is theirs. None of these studies shows that a person’s own connections make better work. For that we tested the one place in Framework where the lab picks connections for a person.

The test

For a named person, Framework’s Ideas page writes five sets of drafts for an everyday brief. One set, Far, is the person’s method meeting outside material. The outside material is five things drawn at random from a fixed list of thirty everyday things, the same list for every person (a library card that lists everyone who borrowed the book before you, a baseball scorecard kept by hand, a pit crew that changes four tires in under three seconds). The writers are not shown the page’s record of the fields the person drew on.

We tested three changes to where the material comes from, with the plan and the bar for changing the page fixed before any run. Far was written four ways: with the fixed list as shipped (Fixed list); with ten things from that list that a model picked as able to change how this brief is solved, five of them drawn (For the brief); with no material given, so the model chooses the fields (Model picks, a reference and not a change that could ship); and with five things from the person’s own recorded links to other fields, each link counted only if its quote is found in the person’s page word for word (Own links). Stretch, which has no outside material, is the anchor. The runs: Steve Jobs, Leonardo da Vinci, Buckminster Fuller and Ada Lovelace, whose pages name fields outside their main work; three briefs (coffee shop, 40th birthday, rental living room); two seeds; 24 runs, five drafts per set. A separate model scores fit and unusualness without knowing the set, 0 to 100. Another call lists the three ideas most people would guess the person would propose, and a coder blind to the set marks each draft against them. The coder also marks whether the idea works the way the outside thing works (its mechanism is borrowed, 0 to 2) or only uses it as a theme, a look or a name. Last, a model was given one draft from each of the four people and asked to match them to the names; chance is 25%.

SetFitUnusualExpected
Stretch69.329.241%
Fixed list62.841.929%
For the brief66.237.234%
Model picks67.141.236%
Own links68.835.544%

Expected is the share of drafts at the move most people would guess for the person.

SetMechanismNamed
Stretch1.1649%
Fixed list1.5124%
For the brief1.4726%
Model picks1.6643%
Own links1.1238%

Mechanism runs from 0 to 2, 2 being an idea that works the way its outside thing works. Named is how often a reader matched the draft to the right person, where chance is 25%.

Neither change met its bar, and the page is unchanged. Choosing the ten for the brief raised fit by 3.4 points, interval −0.1 to +6.9 (the bar was 4 or more with an interval above 0), and lowered unusual by 4.7 (the bar allowed 4). Using the person’s own links raised fit by 5.9 and made a reader’s match to the person 14 points more frequent, interval −1 to +31 (the bar was 10 points with an interval above 0). It also raised the share of drafts at the expected move by 15 points (the bar allowed 5) and lowered unusual by 6.4 (the bar allowed 5).

The same data shows four other things. First, Far as shipped cannot be told apart by person: 24% right, 17 of 72 matches, against 25% by chance and 49% for Stretch. Five unrelated everyday things overwhelm what is the person’s. Second, where the outside thing was chosen with the person in mind, part came back, 38% from their own links (interval 26 to 50) and 43% when the model chose the fields (32 to 56). Left to itself the model took a field that was one of the person’s recorded links in 8 of 120 drafts, 6 of them for Steve Jobs.

Third, the person’s own links made the drafts more theirs and also more expected, the trade the tasteslop test found from the other side. There, steering away from the expected move cost the match to the person; here, steering toward the person’s material raised the match and the expected move together. The own links were also borrowed for mechanism less often than the list’s things, 1.12 against 1.51, a difference of −0.39 (−0.60 to −0.18). That is partly a difference of kind: the list’s things are devices with a way of working, and the links are fields and people, Zen, Emerson, Jacquard’s loom, which the drafts used more as a theme or a name.

Fourth, the one place the own links look like the opposite of a system’s pick is in where the drafts borrow from. In 16% of drafts from the person’s own links, a draft by another person in the same run took from the same field, against 52% when the model picked the fields, 34% from the fixed list and 78% when the ten were picked for the brief (the same ten for all four people, so high by construction). After the plan was written we also saw that the model choosing the fields did as well as the fixed list on unusual (−0.7, −5.8 to +4.4) and better on fit (+4.2, +1.0 to +7.5) and on matching to the person (+19 points, +8 to +31). That arm was a reference in the plan and not a candidate, so it did not enter the decision, and its fields are the common ones: theater, a repair cafe, oral history, citizen science.

These are four people with well-documented links to other fields, everyday briefs, and drafts nobody rated. One model family writes, scores, codes and matches, so the coder and the matcher may share the writers’ idea of the expected move, and the matcher may be recognising a name such as Jacquard’s loom and not the person.

For the four people, one call read the page and returned the person’s main field and up to six links to other fields, each with a quote of at most twenty words. A link counted only if its quote was found in the page word for word. All 37 links returned over two runs passed, and 18 of 19 agreed between the two runs. Steve Jobs gave four (liberal arts and the humanities, Zen, Braun’s design, Polaroid), Leonardo six (optics, anatomy, water, perspective, machines, Euclid), Buckminster Fuller six (Transcendentalism, a Unitarian family, naval service, Froebel’s blocks, general semantics, geometry in nature) and Ada Lovelace three (mechanical calculation, writing and translation, weaving).

On 30 pages across three tiers, 27 gave two or more checked links, and 8 pages had at least one link whose quote was not in the page, which the check catches. Of the 1,823 people’s pages in Framework, 1,209 have a Lineage section, where such links are recorded, and 614 are short pages with documented positions and nothing else, so for a third of them there is nothing to read.

What comes out is what the page records, not how the person thinks. By Porter and Rafols’ point, a count of fields overstates the distance: Ada Lovelace’s page names mathematics as her home field, and one of her links, mechanical calculation, sits beside it.

What it says

The reply to the video is half right. Polymath is the one-person end of interdisciplinary work, and none of the studies sets one apart from the other by measurement. What they support is a base and a few links: the best results pair a conventional body of work with a few unusual ties, and a move far from one’s own work costs. Whether that leads a decade depends on how fast the field changes.

For a framework, the useful record is the person’s own links to other fields, each tied to a quote and marked near or far, and not a count of fields. In this test, outside material from a fixed list erased the person, and material from their own links brought part of them back and made the drafts more expected, not less. Both are readings from four well-documented people and one model family, and nobody rated the drafts.

Sources

  1. Root-Bernstein and Root-Bernstein, Polymathy Among Nobel Laureates As a Creative Strategy: The Qualitative and Phenomenological Evidence, Creativity Research Journal 35(1), 2022 (author manuscript).
  2. Root-Bernstein, Allen, Beach and colleagues, Arts Foster Scientific Success: Avocations of Nobel, National Academy, Royal Society, and Sigma Xi Members, Journal of Psychology of Science and Technology 1(2), 2008.
  3. Levitin, Masters of None?, a review of Burke, The Polymath: A Cultural History from Leonardo da Vinci to Susan Sontag, Yale University Press, 2020, Literary Review 491, 2020. The book was not read.
  4. Araki, Berg, Ivcevic and Maertz, Toward a Modern Science of Polymathy: From Historical Traditions of Wide-Ranging Learning to a Consolidated Breadth–Depth–Integration Triadic Approach, Preprints.org, October 2026 (abstract only; not peer reviewed).
  5. Young and Marzano, Embodied interdisciplinarity: what is the role of polymaths in environmental research?, Environmental Conservation 37(4), 2010 (abstract only).
  6. Porter and Rafols, Is science becoming more interdisciplinary? Measuring and mapping six research fields over time, Scientometrics 81(3), 2009.
  7. Uzzi, Mukherjee, Stringer and Jones, Atypical Combinations and Scientific Impact, Science 342(6157), 2013.
  8. Fleming, Recombinant Uncertainty in Technological Search, Management Science 47(1), 2001.
  9. Leahey, Beckman and Stanko, Prominent but Less Productive: The Impact of Interdisciplinarity on Scientists’ Research, Administrative Science Quarterly 62(1), 2017 (abstract only).
  10. Arts and Fleming, Paradise of Novelty—Or Loss of Human Capital? Exploring New Fields and Inventive Output, Organization Science 29(6), 2018.
  11. Hill, Yin, Stein, Wang, Wang and Jones, The pivot penalty in research, Nature 642, 2025 (abstract only).
  12. Hsu, Jacks of All Trades and Masters of None: Audiences’ Reactions to Spanning Genres in Feature Film Production, Administrative Science Quarterly 51(3), 2006.
  13. Bromham, Dinnage and Hua, Interdisciplinary research has consistently lower funding success, Nature 534, 2016 (abstract and supplement).
  14. Larivière and Gingras, On the relationship between interdisciplinarity and scientific impact, Journal of the American Society for Information Science and Technology 61(1), 2010.
  15. Taylor and Greve, Superman or the Fantastic Four? Knowledge Combination and Experience in Innovative Teams, Academy of Management Journal 49(4), 2006.
  16. Wang, Veugelers and Stephan, Bias against novelty in science: A cautionary tale for users of bibliometric indicators, Research Policy 46(8), 2017 (accepted manuscript).
  17. Teodoridis, Bikard and Vakili, Creativity at the Knowledge Frontier: The Impact of Specialization in Fast- and Slow-paced Domains, Administrative Science Quarterly 64(4), 2019 (submitted version).
  18. Dunbar, How scientists think: On-line creativity and conceptual change in science, in Ward, Smith and Vaid (eds.), Creative Thought, American Psychological Association, 1997.
  19. Gick and Holyoak, Analogical problem solving, Cognitive Psychology 12(3), 1980.
  20. Fu, Chan, Cagan, Kotovsky, Schunn and Wood, The Meaning of “Near” and “Far”: The Impact of Structuring Design Databases and the Effect of Distance of Analogy on Design Output, Journal of Mechanical Design 135(2), 2013 (abstract only).
  21. Swanson, Medical literature as a potential source of new knowledge, Bulletin of the Medical Library Association 78(1), 1990; DiGiacomo, Kremer and Shah, Fish-oil dietary supplementation in patients with Raynaud’s phenomenon: A double-blind, controlled, prospective study, American Journal of Medicine 86(2), 1989 (abstract).
  22. Tshitoyan, Dagdelen, Weston and colleagues, Unsupervised word embeddings capture latent knowledge from materials science literature, Nature 571, 2019.
  23. Sourati and Evans, Accelerating science with human-aware artificial intelligence, Nature Human Behaviour 7(10), 2023.
  24. Anderson, Maystre, Anderson, Mehrotra and Lalmas, Algorithmic Effects on the Diversity of Consumption on Spotify, Proceedings of The Web Conference 2020.
  25. Hosanagar, Fleder, Lee and Buja, Will the Global Village Fracture Into Tribes? Recommender Systems and Their Effects on Consumer Fragmentation, Management Science 60(4), 2014 (author final draft).
  26. Shen, Druckmann and Zou, Unlocking LLM Creativity in Science through Analogical Reasoning, arXiv 2605.11258, 2026; Shen, de Sousa, Barcelos, Guo and Zhu, On the Diversity of Analogy Making in Large Language Models, arXiv 2608.03233, 2026.
  27. Padmakumar and He, Does Writing with Language Models Reduce Content Diversity?, arXiv 2309.05196, ICLR 2024.
  28. Countless Lab, What tasteslop is made of.
  29. Countless Lab, polymath test on Framework’s Ideas page, November 18, 2026: 24 runs, 377 new model calls and 13 more from cache, plan written before the first run.