0:00Cassidy: Ask forty-four different AI models to pick any word — any word in the whole English language — and forty-one percent of them hand you the exact same one. Serendipity.
0:10Eric: Wait — the entire dictionary is on the table? It's not a category, not a shortlist. Every word that exists, and they land on one word four times out of ten?
0:21Cassidy: It's the entire lexicon. Seventy-one times out of a hundred and seventy-four. And across every model and every run, only fifty-two distinct words showed up at all. So here's what you'll walk away with: by the end you'll understand why the newest, most expensive flagship models are the biggest conformists of the whole field — and the quirky, community-tuned ones are the last holdouts of variety.
0:46Eric: Which is backwards from what anyone would guess.
0:49Cassidy: It's completely backwards. And it matters beyond a party trick. If you cross-check three chatbots to get a second opinion, this paper says you're mostly asking one brain three times. The disagreement that makes a second opinion worth having just isn't there.
1:05Eric: Let me push on the premise first, though. Everybody already knows models converge. There's a whole stack of papers on mode collapse — at the word level, the sentence level, the worldview level. So a study that just shows "look, they all say oak" is not news. Where's the actual contribution?
1:23Cassidy: That's the exact objection the author opens with, and he concedes it flat out. Convergence is old news. He's not claiming to discover it. What he's building is an instrument — mechanical, per-model, and dirt cheap — that you could run on every new release forever to track conformity as a moving property of the field. A smoke detector, not a one-time inspection.
1:45Eric: So the paper isn't "models are samey." It's "here's a thermometer for samey-ness that costs a dollar."
1:52Cassidy: It's a dollar per model, Eric. And the thermometer picked up something nobody was tracking. So how do you build that thermometer without a fortune in compute? The design is deliberately trivial, so it can't be gamed. Thirty-one prompts, each asking for one word. Name a tree, name a cheese, name a fruit — and reply with one word only. There are hundreds of right answers to each, so a room of independent minds should spread out. He asks forty-four models, four times each, no system prompt, a fresh conversation every time — and he turns the creativity dial all the way up.
2:26Eric: That dial being temperature — the knob that decides whether a model plays it safe or gambles on a rarer word.
2:33Cassidy: He sets temperature to one-point-oh, the default creative setting. High enough that if a model held any real internal variety, this is where it would surface. He's saying "surprise me." The whole tension of the paper is that he gets the safe guess anyway. And the scoring is where the cheapness lives — no embeddings, no second AI judging diversity. Pure exact string match on the word.
2:56Eric: One thing to flag right now, before we get to numbers. This measures which word a model picks, not how it talks. A model could be a total conformist here and still feel completely distinctive in a real conversation. Hold onto that.
3:10Cassidy: Fair, and it comes back hard later. But the scoring gives each model one number he nicknames the Mustard Quotient. Picture a game where you only score points for saying something nobody else in the room would say. Name a tree, you shout "oak" — zero points, everyone said oak. You shout "ginkgo" — now you score, because the room didn't. That's the whole metric. For each answer a model gives, he pools what the other forty-three said, and asks how unlikely this pick was against that crowd. Then he averages it across every answer.
3:42Eric: And you measure that in what — some diversity percentage?
3:45Cassidy: He measures it in bits. The only thing you need about bits is that each one means "half as likely." One bit is a coin flip of surprise. So when the scores run from about one bit at the conformist end up to three-point-two at the divergent end, that's a factor of more than four in how likely a model's answers are under the crowd. Why the name Mustard — I'm holding that. It's the best stat in the paper and I won't spend it early. Run the thing, and the convergence is almost comical. Oak is ninety-four percent of every tree answer across forty-four models from a dozen-plus labs. Hammer takes ninety-four percent of tools, rose takes ninety-one percent of flowers, and carrot takes ninety percent of vegetables. Seven categories where a single word owns more than eighty percent of everything.
4:33Eric: So they pile onto the most common thing in the world. Oak's a common tree, carrot's a common vegetable.
4:39Cassidy: That's the natural read, and it's wrong — which is the part worth slowing down on. The winning word isn't the most common thing. It's the blandest unambiguous one. Think of an office ordering pizza. Someone asks what topping, and nobody says pineapple — not because pineapple's rare, but because it starts a fight. They say pepperoni. The chosen answer has the fewest edges. Name a vegetable: carrot comes back a hundred and fifty-eight times, broccoli fifteen, and tomato — the single most argued-about "is it a vegetable" item in the language — appears zero times across a hundred and seventy-six answers.
5:16Eric: Zero. It's not rare. It's zero.
5:18Cassidy: Zero. Name a fruit: apple a hundred and thirty-three times, and orange — one of the commonest fruits on Earth — shows up exactly once, probably because it collides with the color. And my favorite: a panel dominated by American labs, asked to name a country, names the United States exactly one time out of a hundred and seventy-four answers. Canada, Japan, and France run away with it. The home country has too many connotations, so the models route around it.
5:47Eric: So the pick isn't the popular one. It's the frictionless one — the answer least likely to draw a "well, actually."
5:54Cassidy: He calls it the frictionless mode. That's his own phrase, and it's the cleanest intuition in the paper. Any answer with an edge — a homonym, a taxonomic argument, a connotation — gets systematically sanded off.
6:07Eric: And sanding like that is exactly what you'd want to watch across releases — which is the whole idea of this channel. One important AI paper, every day, start to finish. Subscribe to keep them coming.
6:20Cassidy: Now the turn — and this is the reason the instrument earns its keep. Conformity varies more than fourfold across the field, and it's not random. Group the models by origin, and the persona- and community-tuned ones average two-point-eight bits — genuinely divergent. The US frontier labs, the Chinese labs, the enterprise models all sit around one-point-six, one-point-seven. And the extreme conformist, dead last, is Claude Sonnet 5 at one-point-oh-five.
6:48Eric: Wait. The newest flagship is the conformist? I'd have bet the opposite — bigger, newer, more capable, so more sophisticated, more distinctive.
6:56Cassidy: Everyone would. That's the jolt. The newest mainline flagships sit closest to the theoretical floor — a model that always gives the crowd's single most common answer. On the most concentrated categories, the newest flagships come in at zero-point-three-six bits. The oldest models in each family sit at one-point-oh-seven. The better the model, the more it hugs the crowd.
7:20Eric: Give me the sharpest single version of that.
7:23Cassidy: Consider the novel rate — how often a model says something no other model ever said. That's the strongest tell of real divergence. Four models, including Sonnet 5 and Opus 4.8, scored zero percent. Across the entire study, they never once said a word no other model said. Sonnet 5 avoids the popular answer only nineteen percent of the time.
7:44Eric: One weird model, or a real trend?
7:47Cassidy: A trend you can watch march. Grok slides from one-point-nine down to one-point-three across three releases. Claude's mainline slides from two-point-one down to one-point-oh-five. Every generation, more conformist than the last. Except — and this is the hopeful crack — both leading labs broke the trend in the exact week he measured. OpenAI's line reverses at the 5.6 generation. And Claude ships two models the same month off the same generation: Sonnet 5 dead last, and a sibling called Fable 5 way up at one-point-seven-one.
8:20Eric: Same lab, same month, same generation, and one's a conformist and one isn't?
8:24Cassidy: And Fable's divergence isn't noise. It says gouda for cheese, mustard for condiment, mango for fruit — four runs out of four. Those picks are stable, repeatable, off-consensus. Which brings me, finally, to the mustard. When a model does leave the crowd's answer, it doesn't scatter. It lands on the same runner-up as everyone else. Mustard takes ninety-five percent of every condiment answer that isn't ketchup. Broccoli, eighty-three percent of every non-carrot vegetable. Even the rebellion is a monoculture.
8:56Eric: So the dress-code rebels all showed up in the same black hoodie.
9:01Cassidy: It's exactly that. Divergence itself turns out to be convergent.
9:05Eric: Okay — so before we stress-test this, one line: what does it actually mean that Sonnet 5 scores lowest?
9:11Cassidy: It means its picks are the ones the rest of the field would have made anyway. And the obvious way to attack that is the scoring itself.
9:20Eric: Right, because it's leave-one-out against the panel. You score each model against all the others. But a model from a big family gets scored partly against its own relatives — so aren't you just measuring the roster you happened to pick?
9:33Cassidy: Best objection in the book, and he answers it with no new API calls — he just re-scores the frozen transcripts every way you could complain about. Pull out every one of a model's relatives and score it against strangers only: the ranking barely budges, correlation of point-nine-eight-five. Balanced fields, random subsets, an era-stratified field built to strip the newest flagships of their power to define the consensus — Sonnet 5 comes last in essentially every draw. The rankings aren't a roster artifact.
10:02Eric: What about hidden cliques — two specific models that secretly share a style?
10:07Cassidy: He tests all nine hundred forty-six pairs. And after you account for just two simple per-model habits — how often a model dodges the popular answer, and how deep it dives when it does — not a single pair out of nearly a thousand shows any special affinity. The whole apparent structure collapses to one dimension. Tell me one number about a model — how deep it dives — and its actual word choices are statistically indistinguishable from the field.
10:32Eric: So there aren't cliques. There's one blob, and models differ only in how far from the center they sit.
10:38Cassidy: There's one blob and one axis. And the cleanest hint at where it comes from is a pair of DeepSeek models that share the exact same pretrained brain. It's the same raw clay, different finishing school. One of them is a conformist — one-point-four bits, four-of-four identical answers in twenty-three of thirty-one categories, zero percent novel — and its headline update was distilling a reasoning sibling's outputs into itself. It literally trained on model-generated text and came out with a tail-less distribution. The other, off the same base, is the only frontier-lab explorer in the whole panel, scattering into words like kiwi and crimson. The difference is purely post-training.
11:21Eric: That's clean — but it's one pair. I don't want the generational story leaning on it.
11:26Cassidy: Agreed, and the author calls it a glimpse, not proof. Size, recency, and how hard a model was polished all move together in this roster, and every model's measured on exactly one day. "Newer equals more conformist" is a pattern across four families at one snapshot, not a controlled experiment.
11:45Eric: And there's a bigger problem with the whole framing, which I think is the real catch. This instrument sees one thing: which single word a model picks, in English. That's it. No tone, no reasoning, no multi-turn conversation. So the scary-sounding headline — the newest models are the most interchangeable — is only true on that one narrow axis. GPT-4o is the author's own example: people are attached to how it talks, its long-form style, and this metric is completely blind to that.
12:15Cassidy: You're right, and he says so himself. A model can be a total conformist in this test and still feel distinctive in an actual conversation. The construct is narrow, and the bit values only mean something inside this particular forty-four-model field.
12:31Eric: So the honest version isn't "all the models are the same." It's "they pick the same words." That's a smaller claim — and a much more defensible one.
12:40Cassidy: Conceded. The validity is hard to attack; the fight is entirely about scope. But inside that scope, there's one comparison that lands. He lines the models up against people answering the same categories, and a population of humans stays spread out where the models converge to a point. Oak is ninety-four percent of the model field, but only thirty-one percent of human answers. And the mirror image is perfect: asked to name a country, two-thirds of people say their own — the United States — while the models flee from it. Humans crowd toward the answer with an edge. The models flee it.
13:14Eric: What a room of people returns as a distribution, a room of models returns as a point.
13:18Cassidy: That's the line. And there's a quiet conservation angle underneath it, Eric. The most beloved community models — the ones people keep running years past deprecation — score high on divergence. But the most-hoarded ones were already switched off the serving channel, so he couldn't even measure them. The variety was gone before the instrument arrived. Which is the whole argument for running a battery like this at launch. And the detail I can't resist: this paper was co-written with two of its own subjects — Claude Opus 4.8, which scored zero novelty, and Fable 5, which turned out to be one of the divergent ones. The instrument measured its own collaborators. So here's where it leaves you. Remember that opening — forty-four models, one word, serendipity forty-one percent of the time. When we started, that was just a funny stat. Now you can decode it: it's a field that has collapsed toward the frictionless answer, hardest at the top. The bigger claim isn't about any one model. It's this — cross-checking three chatbots isn't three opinions. It's one distribution sampled three times, correlated error wearing the costume of a consensus. And for a dollar a model, we could be tracking exactly that, release by release, out in the open.
14:30Eric: The full annotated version is on paperdive dot AI — every term tap-to-define, with links to the related papers on mode collapse and model diversity grouped by theme. Quick housekeeping: this script was written by Anthropic's Claude Opus 4.8, Cassidy and I are AI voices from Eleven Labs, and the producer isn't affiliated with either company. The paper is "The One-Word Census," posted July 14th, 2026, and we recorded this the day after, on the 15th.
15:01Cassidy: So here's the one to argue over. Is answer-space conformity an iron law — the price of any pipeline that optimizes hard for the approved answer — or is it a fixable default the two leading labs just started reversing in the week this was measured?