Once the gap is filled, FDE disappears
The one-person company doesn't hold up, small teams do; FDE isn't a lasting business, it's just this phase's process for filling the gap in how enterprises absorb AI — once the gap is filled, the function dies.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Model vendors' token margins are squeezing third-party agents
Active users of third-party agent apps may be down a third from their peak. The reason isn't complicated: model vendors have very high token margins, so they can supply tokens at very low cost, while third parties buy tokens from them and can't fight a price war. More importantly, the stronger the model, the smaller the difference a Harness makes — Manus used to pile on a lot of scaffolding to make it strong, and once the model upgraded, 80% of that scaffolding was dead. For ordinary tasks like writing slides, writing docs or doing research, a slightly better or worse Harness makes little difference.
OPC doesn't hold up — sell first, don't stockpile skills
A lot of people who lost their jobs call themselves OPC and ask whether the most important thing is building a Harness or something else. The answer is none of it matters; what matters is finding who pays you. Most OPCs are just pleasuring themselves out there, laughing at this and laughing at that, then conning a batch of people who also want to be OPC into giving them money — but nobody has money left, and a stream with no source can't last. Doing the work and turning the work into money are two different things; the hard part is the second one, and most of these people weren't in sales to begin with and have never sold anything directly in a market.
The information gap is bigger than you think: house calls to install Codex for bosses
A group of entrepreneurs think you're an AI expert, but the questions they ask are how to install Codex and how to store a credit card. Someone specifically found a kid who'd come back from the UK, covering Shanghai for house calls and everywhere else by remote video, to swipe cards and download software for them. Many companies that look extremely high-tech — their founders use Doubao day to day, and ask Doubao a lot of things at work. Traditional enterprises run on thin margins: a VP might make only 500k, an R&D lead might make only 800k, and even if he understands, he also has to be able to express it, and the boss doesn't believe him — the monk from outside is the one who gets believed.
FDE's fate: once the gap is filled, it dies
FDE probably isn't a lasting business; it's more like this phase's process for filling the gap in how enterprises absorb AI, and as the gap slowly gets filled, the FDE hand on the clock disappears. Because the stronger the model and the more complete the product, the more it should be a company's own employees holding an agent and just spraying it themselves. What every company has to solve isn't an AI problem, it's still a business problem — a restaurant needs customers to come and spend money eating its food. AI outsiders don't understand how a restaurant operates, restaurants don't understand AI, and if you want to make money you work desperately to understand the restaurant and slowly help fill its gap.
AI transformation is politics — the boss needs a knife from outside
Insiders know the company, so why bring in McKinsey to explain things? A lot of the time it's politics. AI revolutionises yourself: for employees it's there to take their jobs, for management it's there to take their power, and operating on yourself is hard. The boss looks very authoritative, sprays his subordinates like dogs, but he can no longer interfere in the business, and may not even dare fire that executive. The executive will say very directly to the HR lead pushing AI transformation: what's in this for me? The two know each other well, so he can't say there's nothing. So you borrow a knife to kill — it's not that the position can't do it, it's that the position can't do it.
Distillation is an extremely high-ROI way to get data
Distillation is definitely useful, an extremely high-ROI one. The efficiency gap between having AI write code and having a person write code — that's the efficiency gap in preparing data. The way you get training samples is automatic: as long as you have enough accounts and enough money, it scales horizontally; whereas having people write data is manual and doesn't necessarily scale horizontally. From a post-training perspective, for getting bootstrap data, the efficiency difference between the two sides is enormous. As for why ByteDance didn't produce SOTA, it isn't necessarily related to distillation — Google didn't either, and both ByteDance and Google do everything, so they may be aiming for a unified all-modality model, and early fusion makes the problem more complicated.
DSH isn't a good Harness, but it's a good kernel
Good Harnesses or Agents all make a lot of trade-offs and have strong opinions: Codex has a plugin mechanism but you can't change its own Agent loop; Pi thinks you don't need many tools; OpenCloud values access through various IMs. DSH's approach is that everything and anything can be changed, so it's definitely not a good Agent or a good traditional Harness — people who spray it saying it's badly made, I completely agree. But its path has value: why is the whole coding harness a static thing, why can't this round of conversation be like this and that round like that, shouldn't the whole agent runtime also be changing dynamically.
Wearable software: everyone needs their own kitchen
Overseas they call this Wearable Software. Two images make the point: one is a maker's studio, below it a kitchen — every chef's or electrician's workspace looks a little different, the equipment, the layout, the spatulas and seasonings they habitually use are all different. That's how everyone uses an Agent too: some use Cmax, some use the Codex client, some use Herdr. This need used to be hard to satisfy — everyone gritted their teeth and used a combination of different software, and customising in very small places was hard; today writing code to customise to my needs has become less difficult, but you shouldn't build from zero either — you take some piece of software or kernel and change it into what you need.
AI-native org: let AI drive the work loop closed
What works better isn't changing the org so you use AI better, it's using AI to perform the org's functions — AI is the org's competitor. Concretely: file a bug in Feishu, and the Agent automatically looks at online logs and the current codebase to answer, knowing who you are, your email, the actual running state of your account; every day it scans online alerts and posts them to different groups; every week it automatically scans which modules are duplicated and need refactoring, where docs haven't been updated and aligned, and AI sends you a PR with the changes. When a designer wants to change the UI, engineering hands them the codebase to change themselves, with a check gate added so they can only commit to a branch, and engineers set up the environment for them so they can simulate the online server locally.
In their own words · checked verbatim
I think OPC itself doesn't hold up, but I think small teams do hold up — I've always held this view. FDE probably isn't a lasting business; it's more like this phase's process for filling the gap in how enterprises absorb AI, and as the gap slowly gets filled, the FDE hand on the clock disappears.
我觉得OPC本身不成立 但是我觉得团队归功小是成立的 我其实一直是这个观点 FDE可能它不是一个长久的业务 它更像在这个阶段去填补企业渗透AI缝隙的过程 随着慢慢填补了 FDE这个时针就消亡了
The most important thing is finding who pays you. They'll say, for OPC, is the most important thing building this Harness or what? I say it doesn't matter — what matters is paying.
最重要是找到谁给你付钱 他们就会说 OPC我们最重要是搭这个Harness 还是干嘛干嘛 我说不重要 不重要是给钱
Where are most people in the world stuck? Recently, by chance, I've been mixing with a group of entrepreneurs on some things. They think you're an AI expert, but the questions they ask me are all like how to install Codex and how to store a credit card.
世界上绝大部分人的卡点卡在哪里呢 我最近机缘巧合反正有一群企业家 一起混一些事情 他们觉得你是AI专家嘛 但是他们问我的问题 都是类似于Codex怎么安装 以及信用卡存储
What every company has to solve isn't an AI problem, it's still a business problem. Whether I'm a retailer today, a restaurant today, or a manufacturing enterprise today, the problem I have to solve isn't how efficient my AI is — it's that my restaurant needs customers to come and spend money eating my food.
所有的公司要解决的并不是AI问题 还是业务问题 我今天是个零售商也好 我今天是一个餐馆也好 我今天是一个制造业企业也好 我要解决的问题并不是我的AI多有效率 而是我这个餐馆得有客人来 花钱吃我这个麦
Because what AI actually revolutionises is yourself. For employees, it's a little bit here to take their jobs; for management, it's a little bit here to take their power — operating on yourself is hard.
因为AI它其实革命的是自己 自己对于员工来讲 其实少少有一点事是来抢饭碗的 对于管理层多多少少 它有一点事来抢权的 让自己开刀自己这事
And the way you get training samples is automatic, and as long as you have enough accounts and enough money, it scales horizontally; but having people write this data is manual and doesn't necessarily scale horizontally.
而获取训练样本的方式是自动的 而且你只要有足够多的账号跟足够多的钱 它是横向扩展的 但是你要人去写这个数据 它是手动的 且不一定能横向扩展
But Deepseek's Agent feels like everything can be changed. I think an all-purpose thing is never equal — it definitely faces this problem: what is this thing actually for?
但是Deepseek的Agent的感觉就是所有东西都可以改 我觉得一个万能的东西万万不等的 就是他一定他面临这个问题 这东西到底该啥用呢
It's actually not changing the org so you use AI better, it's using AI to perform the org's functions — AI is the org's competitor.
其实就不是改组织 让你更好的用AI 而是用AI来完成组织的职能 就是AI是组织的竞品
Figures
| Decline in active users of third-party agent apps | may be down a third from the peak | 3:55 |
| Manus cost per task | 20 USD gone on a single task | 5:02 |
| Cloud Code subscription price | 200 USD | 5:02 |
| Traditional enterprise VP annual salary | maybe only 500k | 14:08 |
| Traditional enterprise R&D lead annual salary | at a not-small company, maybe only 800k | 14:08 |
| Kimi subscription price | 199 RMB | 23:11 |
| Grok subscription allowance | X Premium includes roughly 30 USD worth | 23:11 |
| Array microphone dev board price | roughly three to four hundred RMB | 1:02:28 |
| Camera price | maybe one hundred–odd RMB | 1:02:28 |
Glossary
- OPC / one-person company
- A company run by one person; the show argues it doesn't hold up, small teams do.
- FDE / forward deployed engineer
- The role that goes on-site to help enterprises land AI; the show argues it dies once the gap is filled.
- Harness / agent shell
- The agent runtime wrapped around a model, deciding tools, loops and interception points.
- Wearable Software
- Software changed into what you need based on some kernel, rather than built from zero.
- early fusion / late fusion
- The two decisions about when to add visual capability in multimodal training.
How to listen
Founders and engineers building AI applications, enterprise services and agent products, and managers pushing AI transformation inside their companies.
The opening news roundup and the closing chat about making videos and Jarvis can be fast-forwarded.