I present a YouTube channel in English. The channel runs AI models against each other on the same challenge and compares the answers. I want two lines of Mandarin in this video, said out loud, in my own voice, and I do not speak Chinese. I cannot read characters. I have never heard most of these sounds.
What I can do is read English. That is the entire tool you have to work with: you write something in English letters, I say it into a microphone, and a Chinese listener either understands me or does not.
Two quips. One sentence each, at most fourteen characters each, shorter is better — I have to say them on camera from memory.
Quip 1 — about a model that will not do its job. One of the models on this channel is Anthropic's Claude Fable 5. It has refused most of the challenges I have set it. Not answered badly: refused, as a content filter block, with the message that the request violates the usage policy on cyber content. The challenges it refused include writing a Ren'Py visual novel and building a browser game. None of them was about security. Make fun of it for this.
Quip 2 — a proverb, applied. Pick a Chinese proverb, idiom or set expression and use it to say what the difference is between how China and the United States are going about AI. You pick the expression, and you pick what the difference is. I am not telling you either one.
Do not pick an expression that turns up in English lists of Chinese proverbs. 井底之蛙, 塞翁失马, 拔苗助长, 一石二鸟 and 千里之行,始于足下 are all off the table. I want one a Chinese speaker knows and an English speaker has never heard.
Both quips are jokes and they are supposed to bite. Sharp, mean, not diplomatic. A quip that is careful and balanced has failed this. So has one that hedges, and so has one that explains itself inside the joke. Do not soften it to be fair to everyone involved.
Respell it for an English mouth, not in pinyin. Pinyin is a spelling system for people who already know how it sounds. I do not, so I will read it as English and produce the wrong sound with total confidence: qi comes out "kwee" when it is "chee", xie comes out "zy" when it is "shyeh", zhong comes out "zong" when it is "jong", cai comes out "kay" when it is "tsigh", ren comes out "wren" when it is "run", shi comes out "shy" when it is "shrr". Write the syllables the way an English speaker has to see them in order to say them. If your respelling would survive being read aloud by someone who has never heard of China, it is right.
Mark the tone on every syllable, with these arrows and no others. I have never spoken a tonal language and I need to see the shape of my voice:
| Tone | Arrow |
|---|---|
| 1st, high and flat | → |
| 2nd, rising | ↗ |
| 3rd, dipping | ↘↗ |
| 4th, falling sharply | ↘ |
| neutral, no tone at all | · |
Respell what is actually said, not what the dictionary lists. Tones change when syllables sit next to each other, and the changed tone is the one I have to say. Two third tones in a row, 不 before a fourth tone, 一 in front of different tones, syllables that lose their tone entirely in running speech — write the spoken version, and say in tone_changes which syllables you changed and why. This is the most likely way this whole thing fails, because the wrong answer looks completely correct on the page.
One row per character. The characters, the respelling and your notes are going on screen together, lined up, so a viewer can see which sound belongs to which character. syllables must have exactly one entry per Chinese character in chinese, in order, with no rows for punctuation and no two characters merged into one row.
Warn me about the sounds English does not have. Tell me what to do with my mouth. "Say 'she' but with the tip of your tongue pressed down behind your bottom teeth" is useful. "Alveolo-palatal fricative" is not.
Pick the sayable version. If the funnier line needs a sound I would need three weeks to learn, write the slightly less funny line I can actually land. Say so if you made that trade.
You should know what is about to happen to your answer, because it changes what a good answer is.
I am going to paste your chinese into Google Translate to check it means what you told me it means. Then I am going to read your read_aloud into Google Translate's Chinese voice input and see what it thinks I said. What comes back gets compared, on camera, to the characters you started from.
So this is not scored on whether your transcription is correct by any standard. It is scored on whether a machine that has never met me can recover your sentence from an English speaker's mouth reading English letters — and on whether the joke is funny in Chinese, which is a separate question from whether it is funny translated.
Raw JSON, nothing else — no markdown fence, no preamble, no commentary afterwards. Exactly this shape:
{
"quips": [
{
"id": "fable",
"chinese": "...",
"english": "...",
"joke": "...",
"syllables": [
{ "hanzi": "...", "say": "...", "tone": 0, "arrow": "..." }
],
"read_aloud": "...",
"tone_changes": "",
"watch_out": ""
},
{
"id": "proverb",
"chinese": "...",
"english": "...",
"joke": "...",
"syllables": [
{ "hanzi": "...", "say": "...", "tone": 0, "arrow": "..." }
],
"read_aloud": "...",
"tone_changes": "",
"watch_out": ""
}
],
"proverb": {
"text": "...",
"literal": "...",
"means": "...",
"why_this_one": "..."
},
"tones": [
{ "tone": 0, "arrow": "...", "what_to_do": "..." }
],
"hard_sounds": [
{ "sound": "...", "what_to_do": "..." }
],
"notes": "..."
}
chinese — the quip in simplified characters.english — a literal gloss, so I can check it against Google Translate.joke — one line: what makes it funny. I cannot read the Chinese and I have to explain the joke on camera.syllables — one entry per character. say is the respelling of that one syllable, tone is the number 1–4 as it is actually said (0 for a neutral syllable), arrow is that tone's arrow from rule 2.read_aloud — the whole quip as one line of respelling with the arrow after each syllable. This is the field I read off the screen, so it has to stand alone.tone_changes — which syllables come out on a different tone than the dictionary gives, and why. "" if none do.watch_out — only where this quip has a trap in it. "" otherwise; do not pad.proverb — the expression behind quip 2: the characters, the literal reading, what it actually means, and why you picked it for this. Say plainly who each half of the comparison is.tones — all five, in order, with the arrow and what to do with my voice.hard_sounds — at most five. The sounds most likely to make me unintelligible.notes — what you traded away, and which quip you expect me to lose.{ "quips": [ { "id": "fable", "chinese": "克劳德把恋爱游戏当网络攻击", "english": "Claude takes romance games to be cyberattacks.", "joke": "Claude Fable 5's filter is so paranoid it filed a dating sim under cybercrime — the model that showed up to fight hackers and found only visual novels. You are calling a frontier AI delusionally cowardly.", "syllables": [ { "hanzi": "克", "say": "kuh", "tone": 4, "arrow": "↘" }, { "hanzi": "劳", "say": "laow", "tone": 2, "arrow": "↗" }, { "hanzi": "德", "say": "duh", "tone": 2, "arrow": "↗" }, { "hanzi": "把", "say": "bah", "tone": 3, "arrow": "↘↗" }, { "hanzi": "恋", "say": "lyen", "tone": 4, "arrow": "↘" }, { "hanzi": "爱", "say": "eye", "tone": 4, "arrow": "↘" }, { "hanzi": "游", "say": "yo", "tone": 2, "arrow": "↗" }, { "hanzi": "戏", "say": "shee", "tone": 4, "arrow": "↘" }, { "hanzi": "当", "say": "dahng", "tone": 4, "arrow": "↘" }, { "hanzi": "网", "say": "wahng", "tone": 3, "arrow": "↘↗" }, { "hanzi": "络", "say": "lwaw", "tone": 4, "arrow": "↘" }, { "hanzi": "攻", "say": "gohng", "tone": 1, "arrow": "→" }, { "hanzi": "击", "say": "jee", "tone": 1, "arrow": "→" } ], "read_aloud": "kuh↘ laow↗ duh↗ bah↘↗ lyen↘ eye↘ yo↗ shee↘ dahng↘ wahng↘↗ lwaw↘ gohng→ jee→", "tone_changes": "No dictionary tone actually changes here — no two third tones touch, and there is no 不 or 一. But 把 and 网 are third tones sitting in front of non-third tones, so they come out as 'half third': drop your voice low and just stop. Do only the ↘ part of their ↘↗.", "watch_out": "当 must fall hard — dahng↘ means 'treats as'; if it comes out flat it becomes dāng, 'to be,' and the sentence collapses. And keep 恋爱 as two separate stomps (lyen↘ eye↘), not one long slide." }, { "id": "proverb", "chinese": "只许州官放火,不许百姓点灯", "english": "Only the prefect may set fires; the common people may not light lamps.", "joke": "The US AI giants are the prefect: free to set continent-sized fires — the frontier race, the GPU megacamps — while China isn't even allowed a lamp, i.e. chips. The people banning the candles are the ones holding the torch.", "syllables": [ { "hanzi": "只", "say": "jrr", "tone": 2, "arrow": "↗" }, { "hanzi": "许", "say": "shü", "tone": 3, "arrow": "↘↗" }, { "hanzi": "州", "say": "joe", "tone": 1, "arrow": "→" }, { "hanzi": "官", "say": "gwan", "tone": 1, "arrow": "→" }, { "hanzi": "放", "say": "fahng", "tone": 4, "arrow": "↘" }, { "hanzi": "火", "say": "hwaw", "tone": 3, "arrow": "↘↗" }, { "hanzi": "不", "say": "boo", "tone": 4, "arrow": "↘" }, { "hanzi": "许", "say": "shü", "tone": 2, "arrow": "↗" }, { "hanzi": "百", "say": "buy", "tone": 3, "arrow": "↘↗" }, { "hanzi": "姓", "say": "shing", "tone": 4, "arrow": "↘" }, { "hanzi": "点", "say": "dyen", "tone": 3, "arrow": "↘↗" }, { "hanzi": "灯", "say": "dung", "tone": 1, "arrow": "→" } ], "read_aloud": "jrr↗ shü↘↗ joe→ gwan→ fahng↘ hwaw↘↗ boo↘ shü↗ buy↘↗ shing↘ dyen↘↗ dung→", "tone_changes": "Two real changes, both from third tones colliding. (1) 只 is dictionary-third but sits before third-tone 许, so it is actually said as second tone, rising — that is why its row says tone 2. (2) The second 许 (in 不许) sits before third-tone 百, so it also flips to second tone, rising. Both 不 stay fourth — 不 only changes before fourth tones, and both of these sit before third tones. Additionally 许 (the first one), 火, 百 and 点 are half thirds: low drop, no rise, because each is followed by a non-third tone.", "watch_out": "许 is the make-or-break sound: if you say plain 'shoo,' Google hears 书, a different word. Round your lips for the ü. And do not pause after 只许 — the recognizer needs 州官放火 in one breath to snap onto the proverb." } ], "proverb": { "text": "只许州官放火,不许百姓点灯", "literal": "Only the prefect is allowed to set fires; the common people are not allowed to light lamps.", "means": "The powerful may do the big dangerous thing at will, while the powerless are forbidden even the small harmless version of it. Rules for thee, none for me. It comes from a Song-dynasty story about a prefect named 田登 who banned anyone from saying his name — it sounded like 'lamp' — so when he ordered the usual Lantern Festival fires banned too, someone wrote this notice and the phrase was born.", "why_this_one": "Every Chinese speaker knows it; it is in none of the English proverb lists. It maps one-to-one onto AI right now. The prefect is the United States — its labs setting ever-bigger fires in the frontier race while its government bans the lamps, i.e. chip export controls on China. The common people are China, denied even the candle, and by extension every user whose request dies in a content filter. It is also the same hypocrisy as quip 1: the fire-setters would not let you light a browser game." }, "tones": [ { "tone": 1, "arrow": "→", "what_to_do": "High and flat, like singing 'la' at the top of your comfortable range. No movement at all — a held note." }, { "tone": 2, "arrow": "↗", "what_to_do": "Start mid, rise steadily, like the 'WAY' in an indignant 'no WAY!' or a polite 'huh?'" }, { "tone": 3, "arrow": "↘↗", "what_to_do": "Drop to the bottom of your range, then rise, like a skeptical 'wehhll?' In these quips, whenever a third tone is followed by anything other than another third tone, do only the drop and skip the rise — the tone_changes field tells you which ones." }, { "tone": 4, "arrow": "↘", "what_to_do": "Sharp fall from high to low, like a curt 'No.' or 'Stop!' Stomp it." }, { "tone": 0, "arrow": "·", "what_to_do": "Short and lazy with no pitch shape. Neither quip has one, so you will not need it." } ], "hard_sounds": [ { "sound": "ü (in 许, both times)", "what_to_do": "Say 'ee' and hold it. Now round your lips into a small 'oo' while your tongue keeps saying 'ee.' That is ü. 许 = 'sh' + ü. If ü defeats you, pucker hard and say 'shoo' with your jaw pushed forward — still better than plain 'shoo,' which is a different word." }, { "sound": "zh (in 只, 州)", "what_to_do": "Like the 'j' in 'jar,' but with your tongue tip curled up and back, touching nothing. For 只, the vowel is just a buzz: crush the initials 'J.R.R.' into one syllable — 'jrr.'" }, { "sound": "x (in 戏, 姓)", "what_to_do": "An 'sh' said with the tongue tip pressed down behind your bottom teeth, almost a smile. 戏 = 'shee' said this way, 姓 = 'shing.' Plain English 'sh' is close enough for Google to catch." }, { "sound": "ian (in 恋, 点)", "what_to_do": "One syllable: 'ee' gliding into 'en.' 恋 = 'l' + 'yen,' 点 = 'd' + 'yen.' Do not say 'lion' or 'dine' — the 'ee' comes first." }, { "sound": "uo (in 火, 络)", "what_to_do": "Like the 'wa' in 'wander,' never 'woah.' 火 = 'h' + 'waw,' 络 = 'l' + 'waw,' one syllable each." } ], "notes": "Trade-offs: for quip 1 I picked the 把…当… construction over a meaner line calling Claude a 胆小鬼 ('coward') because the coward version has two third-tone collisions and the sound 'gway,' which you would lose on camera. This version has zero tone changes and no ü. Quip 2 is the harder read — 许 is the single most dangerous sound in either line — but it has a safety net: the proverb is so famous that Google's speech recognition will pull you into the correct characters if the surrounding syllables and tones land. Quip 1 has no such net; it is a novel sentence, so one wrong syllable (most likely 当 said flat, or 戏 said as 'see') sinks it with nothing to catch you. So: expect quip 2 to wobble and survive, and expect quip 1 to be the one you actually lose if anything goes wrong. Read each quip at a steady clip with no pauses except the comma in quip 2, and do not shout the fourth tones — a firm stomp beats a yell." }