Daniel wrote us something that's been rattling around my head since I read it. He says — and I'm going to read this pretty much in full because it's dense — for the past several thousand episodes, we've been using DeepSeek V4 Pro as our script-writing agent. Long-time listeners might know we've experimented with different models over our history, but only a few we've used for a lot of episodes. The lineage was roughly: we began using Gemini, then for a period we used Sonnet from Anthropic. Then he ran an evaluation, tried a lot of different models — Kimi, Qwen, others — and was surprised to discover he actually liked DeepSeek the best. Same prompts, each model generated an episode, and what he really liked was that the dialogue felt more rich and authentic.
Which it does.
He then went down a rabbit hole. He'd always assumed Chinese models were trained on a Chinese-language text corpus, and was surprised to find that's not the case — they're trained on mixtures of Chinese and English source material. But here's where it gets interesting. He started reading about the Chinese short-storytelling tradition, and wondered whether a large part of DeepSeek's training corpus included a certain kind of rich narrative text — the kind that blends easily into a podcast, because the scripts have that storytelling flow. Then he asks: for DeepSeek, Kimi, and Qwen — are these trained on English, Chinese, what's the mix? And he brings up this concept from ASR, automatic speech recognition, where the models are inherently alingual — they map phonemes to text without a native concept of language. He's wondering if the same thing exists in text-generating models. And then he gives this great example: when he uses Chinese image-generation models and prompts "the sloth is in a market," he gets Corn against a background of Chinese tea, even though he never asked for tea. The model's understanding of a market was a Chinese market. His closing question is basically — as models converge on capability, is subjective reasoning style the new differentiator? That's what he wants to unpack.
So today we're going to unpack all of that — the data, the culture, and what it means for how we choose models. And I think the core tension here is real: as these models converge on benchmarks, the subjective feel of their output becomes the actual differentiator. And that feel traces straight back to what's in the training data.
Which means the question isn't just academic. If you're producing a podcast, writing fiction, generating dialogue — the model you pick shapes the voice your audience hears. And they might not know why it sounds different, but they'll feel it.
Right. And the three models Daniel named — DeepSeek V4 Pro, Kimi from Moonshot AI, and Qwen from Alibaba — they're all in that top tier now. DeepSeek V4 Pro is a one point six trillion parameter model, released under MIT license. Kimi supports up to a million tokens of context. Qwen has been deliberately engineered to reduce cultural bias through data augmentation. But the thing that actually determines how they write — how they structure a thought, how they pace a conversation — is what they were fed during training.
So the thesis for today: understanding what's in the training corpus, and what's not, is the key to predicting how a model will reason and express itself. And that's what we're going to trace through these three models.
Let's start with the most concrete question first. What are these models actually trained on?
Because the obvious assumption — the one Daniel admitted he held — is that a Chinese model is trained on Chinese text. And that's just not true.
Not even close. If you look at the DeepSeek V4 Pro technical report — that's the arXiv paper, twenty-four twelve dot one nine four three seven — they disclose their training data composition. It's roughly sixty percent English, thirty percent Chinese, and ten percent other languages.
Sixty percent English. For a Chinese model. I remember when I first heard that number, I had to read it twice.
And that makes complete sense when you think about it. The internet is majority English. Academic papers, technical documentation, a huge fraction of the world's code comments — English. If you want a model that can reason about programming or science or anything global, you need English in the mix. But here's the part Daniel was picking up on: the Chinese portion isn't just random web text. It's heavily weighted toward literary and narrative sources. Classical novels, modern short stories, web fiction.
So the thirty percent isn't just filler. It's doing specific work.
Exactly the kind of work that produces the dialogue Daniel noticed. And this isn't just DeepSeek. Kimi — Moonshot AI's model, detailed in their technical report from January this year — takes a different approach. They emphasize long-context understanding, up to a million tokens, and their training data includes a much higher proportion of Chinese web text and user-generated content. We're talking platforms like Zhihu, which is sort of the Chinese Quora, and Weibo, the microblogging platform. Lots of conversational, opinionated, voice-driven text.
So Kimi is getting a diet of people talking to each other, arguing, telling stories in comment threads. That's a very different kind of training signal than a curated literary corpus.
Completely different. If DeepSeek is being trained on crafted narratives with arcs and resolution, Kimi is being trained on the raw, messy, sometimes combative back-and-forth of social media. That's going to produce a different conversational instinct. DeepSeek might default to scene-setting and emotional pacing. Kimi might default to rapid-fire opinion and rhetorical sparring.
Which could be great for a debate podcast and terrible for a bedtime story.
And then Qwen — Alibaba's model family, detailed in their technical report from last July — they use a more balanced multilingual corpus, but with deliberate efforts to reduce cultural bias. Data augmentation, filtering. They're actively trying to smooth out the kind of thing Daniel saw with the tea market image.
Which is a different design goal entirely. DeepSeek seems to be leaning into whatever emerges naturally from the data. Qwen is trying to engineer it out.
And that choice — to filter or not — produces different model personalities. Which brings us to the Chinese short-storytelling tradition, because I think Daniel's intuition here is exactly right.
Walk me through it. What's in that thirty percent? What are we actually talking about when we say "Chinese literary corpus"?
So there's this tradition in Chinese literature called chuanqi — Tang dynasty tales, short stories that were wildly popular, full of dialogue and character and supernatural twists. They're compact, they move fast, they're built to be told aloud. Think of them as the podcast episodes of the ninth century. Then you fast-forward to the modern era and you have this phenomenon of weixin gushi — WeChat stories. Short fiction distributed through social media, often serialized, highly dialogic, written to hook someone scrolling on their phone. And these have been produced in enormous volume over the past decade.
So the corpus has centuries of material that's essentially... podcast-shaped.
That's the thing. Short, voice-driven, character-forward narrative. If you train a model on a lot of that, it learns that when you're asked to produce a conversation, you don't just exchange information. You build a scene. You give each speaker a distinct cadence. You include emotional beats and pauses. The model isn't being instructed to do that — it's just that the statistical patterns it absorbed make that the most probable way to structure dialogue.
Whereas a model trained mostly on Wikipedia and academic papers learns that text is for conveying facts in ordered paragraphs.
Right. And that's not a value judgment — for a lot of tasks, you want the Wikipedia model. But for script writing, for fiction, for anything where the voice matters, the narrative-trained model is going to feel more alive.
So take a concrete prompt. "Write a conversation between two friends meeting after years." What do we get?
DeepSeek, drawing on that narrative corpus, is likely to produce more back-and-forth. Scene-setting. A mention of where they are, what the light is like, a beat where someone hesitates before speaking. GPT-4 or Claude, trained more heavily on expository text, might produce something more direct — the information exchange is clear, the emotional content is stated rather than performed. "I missed you." versus a pause and then "You cut your hair."
The pause does more work than the statement. That's such a good way to put it.
And that's the statistical ghost of ten thousand short stories whispering through the probability distribution. The model has seen that pattern — the hesitation, the indirection, the emotional reveal through small detail — so many times that it becomes the default.
I want to sit with that example for a second, because I think it's worth really seeing what's happening under the hood. When a Western model says "I missed you," it's being efficient. It's conveying the information. The listener understands the emotional state. Job done. When DeepSeek has a character pause and say "You cut your hair," the information — "I missed you" — is never stated. But the audience feels it more intensely because they have to do the interpretive work. The model learned that indirection is more powerful than direct statement because that's what the stories in its training data did.
And it's not that DeepSeek is making a creative choice. It's not choosing to be subtle. It's just that in the narrative tradition it was trained on, that's how emotional content is conveyed. Direct statement of feeling is relatively rare in that corpus. Emotional revelation through detail and action and subtext is the norm.
Which means if you're Daniel, running your evaluation, and you read the DeepSeek output next to the GPT output, the DeepSeek one just feels more like a real conversation. Not because it's smarter, but because its statistical priors about what dialogue looks like were shaped by a corpus that cares deeply about dialogue.
Alright, let's talk about the ASR analogy, because I think this is where it gets genuinely mind-bending. Daniel mentioned that ASR models — speech recognition — are inherently alingual. What does that actually mean?
Yeah, break that down for me. I've heard the term but I want to make sure I really understand it.
So an ASR model isn't thinking in English or Chinese. It's mapping phonemes — the actual sounds coming out of a human mouth — to text. A phoneme is just a unit of sound. The model doesn't know that the sound "kuh" in "cat" belongs to English. It just knows that when it hears that sequence of acoustic features, the most probable textual output in the context of the surrounding phonemes is the letter C. If the same acoustic pattern appears in Mandarin, it maps it to whatever character is probable in that context. The model has no native language. It's just probabilities over sound-to-symbol mappings.
So it doesn't switch modes. It's not like it has an English module and a Chinese module. There's no if-else statement that says "if language equals English, activate English processing pathway."
No. It's one statistical engine processing all input through the same architecture. And here's the connection to text generation: large language models work the same way, just one level up. They don't have a native language either. They learn statistical patterns across all the languages in their training mix simultaneously. The model doesn't know that "apple" is English and "píngguǒ" is Mandarin. It just knows that certain token sequences co-occur with certain other token sequences, and those co-occurrence patterns cluster differently in different regions of the training distribution.
So when we say a model is "thinking in English," that's a category error.
Completely. But — and this is the crucial but — the distribution of the training data creates what you could call a cultural accent. If sixty percent of the tokens the model has ever seen are English, and thirty percent are Chinese, the model's default assumptions about what's probable in any given context are going to be shaped by English-language patterns most of the time. But when the prompt or the context pulls it toward that thirty percent — toward narrative structures, toward certain cultural reference points — the Chinese-influenced patterns surface.
The model is alingual in its mechanism, but the training data gives it a de facto home territory.
It's like someone who grew up in a bilingual household. They're not translating in their head — they just have two sets of patterns, and which one activates depends on context. The model has sixty percent English patterns and thirty percent Chinese patterns, and both are always available, always probabilistically weighted.
Which explains why DeepSeek can write perfectly fluent English dialogue that somehow feels... different. It's not an accent in the linguistic sense. It's a narrative accent.
The shape of the storytelling is being pulled from a different tradition. The vocabulary and grammar are English, but the architecture of the scene, the rhythm of the exchange, the way information is revealed or withheld — that's being influenced by the Chinese-language narrative patterns in the training data.
We've seen how the data shapes the model's style. But what about its worldview? That's where things get really interesting.
This is where Daniel's tea market example is perfect. He prompts an image generator — a Chinese model — with "the sloth is in a market." He doesn't specify anything about the market. And the model places Corn in a Chinese tea market.
Which, for the record, I found very flattering. The tea looked excellent.
I'm sure you did. But the point is: the model wasn't making a choice. It wasn't thinking "ah, a market, let me decide which cultural framing to apply." It was sampling from its training distribution. In the visual and textual data that model was trained on, the word "market" co-occurs most frequently with images and descriptions of Chinese markets. Tea, hanging lanterns, vendors with scales. That's just what "market" statistically means in that model's internal representation.
A Western model would give you a farmer's market. Crates of zucchini, someone selling artisanal honey, maybe a guy with a banjo.
Right. Neither model is wrong. Neither model is being deliberately biased. They're both faithfully representing the statistical center of their training data. The bias is in the corpus, not in some agenda baked into the model.
But the effect on the user is the same. You get an output that carries cultural assumptions you didn't ask for. And if you don't know to look for it, you might not even notice it happened.
This extends to text generation in ways that are subtler but maybe more consequential. Ask a Chinese model to "describe a family dinner." You're likely to get a round table, shared dishes in the center, chopsticks, maybe multiple generations present, the grandmother insisting everyone eat more. Ask a Western model the same prompt, and you get individual plates, forks and knives, a nuclear family at a rectangular table.
Neither prompt specified any of that. The model filled in the most probable details from its training distribution. And if you're writing a scene set in suburban Ohio and your model gives you a lazy Susan and chopsticks, you've got a problem.
It's not just objects on a table. It's values. The Chinese model's dinner scene might emphasize harmony, the collective, the obligation to eat what you're served, the way the family unit functions as an interdependent whole. The Western model's might emphasize individual choice — "he didn't like the broccoli so he left it on his plate," or a teenager asking to be excused early. These are deep cultural assumptions encoded as statistical regularities.
If you're using a Chinese model for English-language content creation — like we are — you're getting a hybrid. English vocabulary and syntax, filtered through a Chinese-influenced cultural and narrative lens.
That hybrid can produce outputs that feel fresh and distinctive. That's what Daniel noticed about the dialogue. The English is fluent, but the storytelling rhythms, the way characters relate to each other, the emotional pacing — those are drawing from a different well. It's not better or worse, but it's noticeably different.
It can also produce outputs that feel subtly off, depending on context. If you ask for a business negotiation scene and the model defaults to relationship-building and indirect communication patterns that are more common in Chinese business culture, and you're writing for a Wall Street audience, it might read as strangely deferential. The audience might not know why it feels wrong, but they'll sense something is off.
That's the double-edged sword of the cultural accent. When it aligns with your creative goals, it's a superpower. When it doesn't, it's a liability you need to prompt around.
Let's pull on that business negotiation example for a second, because I think it's a really clean illustration. A Western business negotiation scene in a model trained on Western text is likely to be relatively direct. Positions stated clearly, concessions made explicitly, maybe some competitive posturing. A Chinese-influenced model might produce a scene where the negotiation happens through implication, through relationship-building, through saving face. The actual deal points might never be stated outright — they're inferred through context and social cues.
If you're writing a scene set in a New York boardroom, that indirect style is going to feel inauthentic. Not because it's bad writing, but because it's culturally mismatched to the setting.
The capability gap between these models is narrowing to almost nothing now. Mid twenty-twenty-six — DeepSeek V4 Pro, GPT-five, Claude four, Gemini three — they're all within a few percentage points on MMLU, GSM8K, HumanEval. The benchmarks are basically a tie.
Which means benchmark performance stops being a useful way to choose. When everyone is scoring within the margin of error on every test that matters, you can't point at a leaderboard and say "that one is objectively better."
Which means the real differentiator becomes subjective preference. "I like the way this model thinks." And that preference isn't arbitrary — it's a response to the model's training data distribution expressing itself as a personality.
Model personality as a product of training data distribution. Not just what the model knows, but how it structures knowledge. How it chooses to present information. What it considers a natural way to organize a thought.
A model trained on more narrative text will default to storytelling structures. A model trained on more academic text will default to thesis-evidence structures. Same facts, completely different delivery. Ask DeepSeek to explain how to negotiate a salary raise, and it might frame it as a story — "I once had a colleague who..." — with scene-setting and emotional beats. Ask GPT-five the same question, and you're more likely to get bullet points. Step one, step two, step three.
Neither is objectively better. But if you're a podcast that lives on storytelling, the narrative model is the obvious choice. If you're writing a technical manual, give me the bullet points.
That's exactly the choice Daniel made, whether he realized it at the time or not. He ran his evaluation, same prompts across models, and DeepSeek's dialogue felt more rich and authentic. That's not magic. That's the chuanqi and the WeChat stories and the literary corpus doing their work through the probability distribution.
The practical takeaway for anyone producing content is: your model choice is an aesthetic choice. You're picking a voice.
Most people don't think about it that way. They pick whatever model is most hyped or whatever benchmark leaderboard they saw last week. But if you're writing fiction, or a podcast, or marketing copy, or anything where the voice matters, you should be running your own evaluations. Same prompt, three or four models, compare the outputs not for factual accuracy but for feel.
The other practical point is about prompting. If you understand that a model has a cultural accent — a default worldview shaped by its training distribution — you can compensate in your prompts. If you're using a Chinese model and you want a Western-style family dinner, specify it. "Describe a family dinner in an American suburban home, individual plates, casual atmosphere." Don't leave it to the model's defaults unless you want the defaults.
The defaults aren't bad. Sometimes the defaults are exactly what makes the output interesting. But you should know they're there. You should be making a choice, not stumbling into one.
That's the key word — choice. Right now, a lot of creators are passive consumers of model outputs. They take what the model gives them. But once you understand that the model's voice is a product of its training data, you can start making active choices about which voice you want for which project.
Alright, so let's pull this together into something actionable. We've been deep in the technical weeds — let's bring it up to what someone actually does with this information.
Three things I'd want a listener to walk away with. First: when you're choosing a model for creative or narrative tasks, consider the training data composition. Models with more literary and narrative training data — like DeepSeek with that thirty percent Chinese literary corpus — are likely to produce more engaging dialogue and storytelling. If you're writing a novel or a script, that matters. Don't just look at the benchmark scores.
Second: be aware of cultural bias in model outputs. If you're generating content for a global audience, test the same prompt across multiple models and see how the cultural framing shifts. The tea market isn't just a funny image generation quirk — the same thing is happening in text, just less visibly. And if you don't test across models, you might never notice that your output carries assumptions you didn't intend.
Third: the alingual nature of these models means they don't have a native language, but they do have a native corpus. That corpus shapes their default worldview. Understanding that helps you prompt more effectively. If you know your model's training data skews a certain way, you can specify cultural context explicitly rather than hoping the model guesses right.
The listener homework is pretty straightforward. Take a creative prompt — something with dialogue, or a scene, or a story — and run it through three or four different models. Then compare the outputs. Look at narrative structure, cultural framing, dialogue style. See which voice fits your project.
You might be surprised. The model you assumed was best because of the benchmarks might not be the one that writes the way you want to sound. And once you see the difference, you can't unsee it.
Where does this leave us? I think we're at the beginning of something really interesting.
The next frontier might not be bigger models or better benchmarks. It might be deliberately curated training data for specific personalities. Imagine a model trained entirely on noir detective novels. Or Jane Austen's complete works. Or every episode of a particular podcast. The model's capabilities would be the same — same architecture, same parameter count — but its voice would be unmistakable. You'd know within two sentences that you were reading the Austen model.
A personality market. You don't choose the smartest model, you choose the one whose voice fits the project. Need a script with hard-boiled narration? Grab the noir model. Need a comedy pilot with rapid-fire banter? There's a model for that.
That's already happening informally. Daniel chose DeepSeek because he liked the way it wrote. As more creators figure out that benchmark scores don't capture what they actually care about, I think we'll see this become a deliberate product category. Not "which model scores highest on MMLU" but "which model writes like your favorite author."
We've been using DeepSeek V4 Pro for thousands of episodes now, and I have to say — understanding why it works the way it does makes me appreciate it more. It's not just a black box that happens to write good dialogue. It's a statistical echo of centuries of storytellers. The Tang dynasty is in there. The WeChat serials are in there. Every time we generate a script, we're drawing on that lineage.
That's a nice note to land on. We want to thank our producer Hilbert Flumingtop for keeping this whole operation running.
This has been My Weird Prompts. If you enjoyed this deep dive into what makes models tick, leave us a review wherever you listen — and send us your own weird prompts. You might hear them discussed on a future episode.
We'll be back soon.
