Agent Is Like a Gorilla Just Picking Up a Burning Stick; Token Consumption Rate Is the Key Metric
Ming Chaoping swapped the key metric for AI products from DAU to token consumption rate, then further to per token valuation—because high token consumption doesn't mean high value, just as the efficiency of converting oil into mechanical energy is what matters.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Battery life data and battery life experience are not the same thing
When Ming Chaoping worked on products at OnePlus, he found that the paper battery life data was the most beautiful, but user feedback was poor. The reason is that users are only sensitive to battery level in two ranges: the moment they unplug the phone at 95%-100%, and when they are most anxious at 5%-0%. Apple's last 5% of battery is particularly durable because it's not actually 5% of real data—it might be 10%, using enough extra battery to ease anxiety. Many companies design battery life as a linear straight line, but the ideal state should be an S-curve. Experience is both subjective and objective, not purely data-driven.
— Ming ChaopingByteDance's methodology erases flashes of genius
Ming Chaoping believes that ByteDance's data-driven system raises the overall average level, allowing new graduates to deliver 80-point results in a week, because AB experiments have already attached 20%, and no matter how stupid you are, you won't launch it. But what it erases is the flash of genius from the top 100-point talents—iPhone's multi-touch screen, Tesla, these disruptive innovations cannot be validated. He compares data to the rearview mirror when driving: you can see the road you've traveled, but it can't guide you forward.
— Ming ChaopingToken consumption rate is the key metric in the AI era
After the 2025 Spring Festival, Ming Chaoping asked himself while sleepless: if I were a VC looking at AI products, what should I look at? DAU comes from the mobile internet era, and today it may not still be the key metric. His conclusion was to look at the token consumption rate. Later, he saw an interview with Manus and found they coincided. But today he has iterated further: greater consumption is not necessarily better; taking it a step further is pursuing the value of each token, called per token valuation. For example, using a million tokens to write a website, many tokens may be wasted; a good agent should write the website with fewer tokens.
— Ming ChaopingThe shell is undervalued; it should be called container and environment
Ming Chaoping believes the word shell is undervalued by everyone; he prefers to call it container and environment. Environment is a reactor for people—the same person, eating barbecue at a street stall versus a high-end French restaurant, will behave, speak, and even sound different. Giving users only a prompt box is irresponsible, because users really don't know what to do with the product. Early Midjourney users typed short phrases like a dog, a cat; today's 500-word prompts evolved in Discord channels, where users see others' work, copy their prompts, and remix them. This process exchanges how everyone uses the product.
— Ming ChaopingCoding development is similar to Camera back then
Ming Chaoping compares the technical development of coding to the development of phone cameras. Early discussions about cameras involved people holding DSLRs, adjusting ISO and shutter speed, but after cameras appeared on phones, a new group of people called mobile photographers emerged. Today coding is the same: previously you couldn't code without knowing programming, but today web coders programming in natural language have appeared. The second variable is that the growth rate is very fast. When phone cameras appeared, the volume of photos produced per day went from 1 million to 10 million to 100 million, rising exponentially. Trend variables in the early days always involve a new group of people appearing, and the growth rate is very fast.
— Ming ChaopingThe agent ecosystem is like two countries: Singapore and the US
Ming Chaoping uses countries as a metaphor for the agent ecosystem: one is a very dictatorial and authoritarian country, with extremely high operational efficiency but strongman politics; the other is a democratic country, where Republicans and Democrats battle back and forth, and building a railway might take twenty years. Behind this are two philosophies—pursuing the upper limit and preserving the lower limit. Corresponding to agents, one is a single agent scheduling all other vertical agents, and the other is a bunch of vertical agents scheduling each other. He leans toward the latter, because if you turn PageRank into agent rank, the biggest problem is turning all 2C companies into 2B companies, which will face resistance.
— Ming ChaopingThe next OS should be alive, with interfaces generated in real time
Ming Chaoping believes that all interfaces in today's operating systems are fixed and dead, written in advance by programmers, and every process is just clicking a button to jump to that interface. But the next OS should be alive, with all interfaces generated in real time based on your task scenario at the time. For example, what that screen shows while driving is different from on a computer, and even while driving, it looks different when moving versus stopped. Why not generate an interface in real time during a meeting, where you can see subtitles, take notes, and see progress? This requires a coding engine; in the future, the two seconds of opening a website will include not only data transmission but also real-time rendering and generation of the interface.
— Ming ChaopingEntrepreneurship is like playing Go; the order of moves matters
Ming Chaoping compares entrepreneurship to playing Go. In Go, there is something called dapu—after a game, you review it by placing the game record in front of you and analyzing it. He emphasizes that what matters is not what the final game record looks like, but the order of moves. You can't just see that Douyin looks like this and WeChat looks like that today and make an identical thing, because if you make a wrong move at step fifty, there might be no step one hundred. Product prioritization is an important manifestation of order: what to do, what not to do, what can be done three years later, what must be done today.
— Ming ChaopingIn their own words · checked verbatim
You can't help but want to carve it, to add more scaffolding and more design. Then once these things are added, it goes from 80% AI content to 20% AI content, and you're going against the biggest variable of this era.
你就情不自禁的想去雕花 想去给它更多的脚手加更多的设计设计 然后然后这些东西一扎上去之后 它就从80%的AI含量变成20%的AI含量 你就跟这个时代最大的变量背道而驰了
Ming Chaoping1:11:59
What you should pursue is the value of each token, called per token valuation.
你要追求的是每个token的价值 叫per token valuation
Ming Chaoping1:14:59
99% of us are going to die. We must—our entrepreneurial mindset is to assume we will fail, and then find a way not to fail.
咱们99%是要死掉的 我们一定 我们的创业心态就是默认自己会挂掉 然后想办法让自己不要挂
Ming Chaoping1:46:15
You must always believe that the model will get better, and at the same time you must always believe that the model has nothing to do with you.
你永远相信model这个事情会变好 同时你要永远相信model跟你无关
Ming Chaoping1:47:15
What matters is not what the final game record looks like, but the order of moves.
重要的不是最终的棋谱 仗什么样的 重要的是它下棋的顺序
Ming Chaoping2:33:44
Figures
| Ming Chaoping's age | 29 years old | 4:17 |
| Company founding date | October 10, 2024 | 4:17 |
| Team size | 22 people | 2:12:35 |
| Sleep cycle | Wakes up every 1.5 hours | 2:37:44 |
| Human brain power consumption | 25 watts | 2:28:43 |
Glossary
- per token valuation
- When measuring AI products, look not only at total token consumption but at how much value each token generates.
- web coder
- A new group of people who can create software using natural language without learning programming.
- agent rank
- Analogous to PageRank, a ranking mechanism that determines which agent gets called when agents schedule each other in the future.
- the bitter lesson
- An idea in AI: methods that rely on computation and scale will ultimately outperform methods that rely on human knowledge.
How to listen
AI application founders, product managers, investors watching the agent ecosystem—especially those thinking about coding agents and OS opportunities.
The childhood and debate experience at the start can be fast-forwarded; jump directly to the battery life experience section at 30:28.