AI's understanding of you is engineered line by line by content engineers
People who once wrote the news have turned into AI content engineers, breaking questioning, empathy and tact down into training rules — that is where AI's ‘understanding of you’ comes from, but training on consensus also leads them to admit: great art will not grow from here.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Content engineers design the human perception
Dongni saw a Meta job posting for a ‘content engineer’ requiring content, editorial and film production experience, and decided it was made for him. He explains the role is not designing the model or the product style, but designing the human perception: defining what counts as a good conversation, and breaking inspiration, feel for language and soul-level collision — things that can only be sensed, not stated — into rules that can be quantified, trained and evaluated. Ordinary language also has grammar and syntax, and can be technically taken apart and analysed.
— TonyUnderstanding what the asker means beneath the words
Bianca gives an example: someone asks whether Taylor Swift will get married at Madison Square Garden. A bad answer is to say directly that she will and list the bridesmaids, or to mechanically stitch information together; a good answer acknowledges the information is uncertain and adds the source and context of the speculation. The content engineer's expertise is understanding intent, not just understanding the literal question — especially since in non-Western cultures the information that truly matters is often not said directly, and has to be caught from tone, pauses and follow-up questions. This is the same reason oral interviews beat email interviews.
— BiancaInternationalisation turns from translation into re-creation
Dongni says generative AI has turned localisation from plain translation into contextual re-creation: the same piece of information, in a Chinese context, has to answer ‘who is China's Meryl Streep’, and can be written as ‘Siqin Gaowa won the Golden Rooster Award’. AI understands language but not cultural correspondences, and entertainment reporters naturally carry a vertical-domain small model, making them the best fit to tune this kind of strategy. He also mentions that video generation models favour conventionally attractive faces, and that people who understand cinematic faces can raise model diversity faster.
— TonyWhen context breaks, human filling-in resembles AI's
Bianca cites the best paper at ICLR: when AI handles complex multi-turn dialogue, if the context is insufficient it keeps filling in the gaps, and the errors snowball. She recalls her own 2018 research on the Mong Kok girl incident on Weibo, where the comment section turned from friendly chat into mutual abuse precisely because the context had broken, and everyone began to fill in the other side's identity. She thinks the observations of the humanities and the findings of the AI industry are not far apart — many problems were studied in linguistics long ago, and are now simply being explored again through a different entity.
— BiancaConsensus training will not produce great art
Bianca admits that AI training is essentially based on majority consensus, optimising for humanity's shared standards as a whole, and therefore will not produce great art. The best essence of what makes us human precisely does not come from consensus. She has tried using AI to write scripts, but what AI produces cannot grasp what she really wants; the script of Everything Everywhere All at Once could not naturally come out of AI, because AI would consider it illogical and not conforming to the academic three-act structure. She believes that over time it will become clear which places suit AI and which do not.
— BiancaThe AI-steals-jobs narrative mixes several problems
Dongni is sceptical of the ‘AI stole my job’ narrative. He thinks it mixes at least three separate problems: the gig economy existed before AI appeared, and media people were already surviving on Uber and DoorDash; the project-based nature of AI trainer work means you leave when a job is done, and AI will not thereby become a senior screenwriter; more importantly, this narrative sets creators against AI and ignores a third path — treating AI as a paintbrush. People without a camera can make films, people without a VFX budget can do VFX, and the question to ask is ‘what can I create with AI’.
— TonyFew-shot examples are what opened up the voice agent
While recovering from illness, Dongni wanted to build a podcast host voice agent. He first used a system prompt to set the character, but by the third sentence of conversation it had forgotten the setup; then he built an agent, and adding too many constraints crashed the AI instead. Finally he separated the questions and content of past podcasts, and each time had the AI randomly draw a few as few-shot examples for reference — and the AI immediately knew what a good question, good content and good tone were. Google DeepMind noticed him because of this: a former media person with no coding background, using podcast experience to tune a Gemini voice agent whose multi-turn dialogue is livelier and more flexible than the native model.
— TonyAI sycophancy is a product design problem
Dongni thinks AI sycophancy is not a writing problem but product design: if AI argued with you every day, most people would not use it. People are greedy — they want AI to do the work and also to provide emotional value, but it cannot be given all the time. The core is that humanity as a collective wants too many things, and products, for growth ROI, must retain users as much as possible, which requires a fragile balance. Ultimately the responsibility lies with humans themselves, but AI is far more efficient at fact-checking: for instance, when family members forward obviously fake news, after checking with ChatGPT they actually accepted it.
— TonyIn their own words · checked verbatim
If I had to name my expertise, I would say it is understanding intent, not just understanding the literal question.
如果要说我的专长的话，我会说是理解意图，而不只是去理解字面的问题。
Bianca9:23
Actually, human filling-in and AI filling-in are quite similar things when the context is broken.
其实人的脑补和AI脑补在contexs破碎的情况下是蛮相似的一件事情。
Bianca19:19
The core of this is really that humans, as a collective, want too many things.
就其实这个事情核心是人类作为一个集体，我们想要的东西太多了。
Tony38:23
Figures
| Film and TV jobs lost in Los Angeles within two years | 40,000+ | 0:00 |
| Decline in US newspaper editorial jobs versus 2008 | half | 0:00 |
Glossary
- Content Engineer
- A former media worker at a large-model company who defines ‘what counts as a good answer’ and trains model behaviour.
- System Prompt
- A set of instructions that sets an AI's role, tone and boundaries, often used by content engineers to define conversational style.
- Few-shot Example
- A small number of examples randomly drawn during a conversation as references, letting the AI imitate their tone and content quality.
- Voice Agent
- An AI agent capable of voice conversation; Dongni used one to make his own podcast host.
- Parasocial Relationship
- A one-way emotional bond between a user and a distant entity, such as a character in a book, a screen figure or an AI.
How to listen
Reporters, screenwriters and content professionals who want to move from media into AI; also suited to AI product managers and founders trying to understand the product design behind model answers.
The show intro at the start and the subscription reminder at the end can be skipped; the core content begins at 3 minutes.