Don't rush to buy tools; enterprise AI starts with building a working model
Wang explains that unlike older technologies like databases, everyone has their own understanding of AI, so widespread adoption won't work through brainstorming. Enterprises should first build a closed-loop working model in one department, then use that as proof to convince the rest of the company to follow.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Brain in a vat has no limbs
Wang's first reaction to seeing GPT-4 was: the problem of human thought is essentially solved—as an AI researcher working in applications, he immediately felt his past research was no longer necessary. But this new intelligence is just a "brain in a vat": no hands, no feet, only able to give answers in a chat window. Users have to manually copy generated code back to the editor. What FDE does is give this "brain" hands and feet, memory and scene-matching capability, turning it from a chat tool into an agent that can run in enterprise production systems.
— Wang ShuyuanProblem feedback doesn't leave the company
Wang says Google Cloud FDE's advantage isn't "stronger models" but a shorter debug loop: when FDE discovers capability boundaries on the customer site, they feed that directly back to internal product or model quality teams, then verify again after updates ship—internal communication within the company. If independent or third-party FDE can't solve a problem, they're looking at another company's tools; cross-organization communication is orders of magnitude slower. Third-party FDE's advantage is staying neutral between model vendors and helping customers compare options, but that's a tradeoff, not a capability advantage.
— Wang ShuyuanFDE sells model homes
Wang uses renovation as an analogy: a single house doesn't need a model home, just hire someone to renovate. But renovating an entire neighborhood means you must build a model home first—the model home may not have a toilet or mop, but it must have running water and electricity, meaning the demo doesn't need every feature, but the data pipeline must actually work in the real scenario and form a closed loop. His provocative take: enterprises can't push AI through brainstorming because unlike databases, where only one department understood the technology, everyone now has their own understanding of AI, making consensus impossible. The right approach is to make one department's model home beautiful and functional, then use that as proof to convince other departments to follow.
— Wang ShuyuanThe boundary between R&D and PM melts first
Wang observed that before the Agent era, the line between his AI R&D team and PM team blurred first—developers started talking directly to LLMs to write project proposals and argue for algorithm value, work that used to be purely PM; PMs took R&D documents and asked LLMs for strategic direction, work that used to be purely R&D. He calls this phenomenon "FDE BP" (business partner): as vibe coding spreads, pure technical ability becomes secondary, what matters is whether you can reverse-engineer from customer business problems what kind of agent workflow is needed. This is making teams reconsider whether whiteboard coding belongs in hiring conversations anymore.
— Wang ShuyuanDon't ride the main channel, value is in the ripples
Wang describes a real case: a customer wanted to translate an entire ad image with pixel-level text from one language to another. He immediately objected because that sits squarely on the model's main channel—sure enough, two or three weeks later Google released Nano Banana and the whole custom system would have been obsolete. His conclusion: FDE should stand on the ship's sides looking for ripples, not at the bow. Needs on the main channel the model vendor will solve themselves soon. FDE's real job is finding the gaps to the sides and behind, the gaps the model still misses but customers genuinely need.
— Wang ShuyuanMechanism can't buy trust
Wang says mechanism alone can't build trust—it can only help eliminate the last tiny bit of doubt. He uses Manus as an example: when multiple cloud vendors and model companies were pitching, what won was Google's FDE colleague solving concrete problems with enough initiative that trust slowly accumulated; only then could they discuss mechanisms like early model checkpoint access. He's blunt: agents still can't replace FDE because building trust face-to-face between people isn't automated yet.
— Wang ShuyuanNext wave explodes in knowledge workers
Wang's provocative take: the first AI wave already cashed out in programming; the next will blow up in knowledge work—sure, knowledge workers use fewer tokens per person than programmers, but feed the AI every meeting someone's attended and document they've written since joining the company, and tokens add up fast. He had a second ChatGPT moment using Gemini Agent for manager work: he doesn't have to tell AI which documents to look at or which meetings matter, just describe what he wants done and AI runs for a long time pulling context together. He calls this process "self-distillation" for knowledge workers and argues many enterprise AI failures aren't about AI itself, but about companies that skipped their digital transformation.
— Wang ShuyuanWhat took two months now finishes in one pass with agents
Wang contrasts his lived experience: using the old workflow (requirements document, architecture document, progress document all kept in sync with every change tracked) to build his wife's career coaching website—database design and full-stack deployment—took about two months of spare time. After Manus, he re-ran the same project without maintaining three documents or tracking changes, end-to-end, and got a basically functional page in one go. He sees agent evolution in layers: first round judges answer quality, second judges task completion, the next round will judge whether it matches how human brains process information at their natural pace—sometimes he wants one sentence, but the agent produces a five-minute read, and that mismatch hasn't been solved.
— Wang ShuyuanIn their own words · checked verbatim
We define FDE as an accelerator helping customers move from AI concepts to actual enterprise production system deployment.
我们是会把FDE定义为帮助客户 从AI的概念走向实际生产系统落地的加速器
Wang Shuyuan1:02
Now I think we'll eventually have countless brains in vats. These brains in vats themselves are very cheap. But although they have no hands and no feet, they can't do anything else.
那我现在觉得说以后我们就会得到无数个刚中之脑 这刚中之脑本身的价格是非常便宜的 但是它虽然没有手没有脚 在当时不能做其他任何的事情
Wang Shuyuan4:04
You can use a very simple analogy: FDE is about building model homes. What's different between building model homes and traditional renovation? Think about it.
可以用一个非常简单的比喻来说明 FDE就是打造样板间的 打造样板间跟传统的装修有什么不一样 大家可以想一想
Wang Shuyuan15:07
Let me give you a provocative take. It's very simple: you must choose the former; the latter won't work. In other words, you must build the model home first.
我讲一个我的暴论 可以 非常简单 一定选前者 后者是走不通的 或者换句话说 就是一定要先打造样板间
Wang Shuyuan20:12
The PM thought ‘now that this is set up like this, I can go forward with it’. Then we discovered our R&D and PM were actually doing the same thing. Really magical.
PM就觉得说 这个东西已经这样了 我就可以去找公众团队了 然后我们发现说 我们研发和PM之间在干 同样一件事情 特别的神奇
Wang Shuyuan30:21
But just after this, maybe without two or three weeks passing, Nano Banana came out. What this means is: the need still exists and there are still many edge cases, but if we had gone ahead and built it then, that entire custom system wouldn't be needed.
但是就在这个事情过后 可能没有两三周的时间 nano banana就出来了 这个事情就是说 需求还存在 还有很多edge case 但是说如果我们当时去做的话 那整套系统都不需要了
Wang Shuyuan38:24
Mechanism can't help you build trust. Right, mechanism can only help you eliminate that last tiny bit of doubt.
机制不能帮助你建立信任 对 机制只能帮助你排除说最最后的那一丁点怀疑
Wang Shuyuan43:27
And gave me very reliable advice. I think the core insight is: a knowledge worker can self-distill.
然后给我非常靠谱的建议 嗯 我觉得他的核心就是说 一个知识工作者 他是可以自我蒸馏的
Wang Shuyuan48:32
Figures
| Interval from requirement to Nano Banana release | two or three weeks | 38:24 |
| Team size a manager can lead (traditional model) | 50 people | 45:27 |
| Personal website build time (old workflow) | approximately two months (spare time) | 52:54 |
Glossary
- FDE
- Forward Deployed Engineer—engineer posted at customer site to implement AI capabilities as production systems.
- brain in a vat
- Philosophical metaphor for intelligence without hands and feet, unable to interact with the external world.
- FDE BP
- Wang's coinage meaning FDE should work like a business partner, reverse-engineering technical solutions from customer business problems.
- Nano Banana
- Google's image generation and editing model.
- self-distillation
- Wang's term for knowledge workers feeding accumulated documents and meeting records to AI.
How to listen
Founders and technical leads aiming to get AI into enterprise production systems, plus people building FDE or delivery teams.
If you're familiar with the history of AI tool evolution, you can skip the opening section about vibe coding history.