Original · Unique Research · 2026-06-11
Editor's note: The first-person report and its judgments belong to the original Chinese author (Wu Wei, Unique Research founder). This English rendition retains the Silicon Valley reflection, the AI Coding arc, the FDE debate, the taste debate, the super-individual discussion, the closing reflection, and the complete 15-question Q&A with Nodesk.ai founder Song Jian. All named companies, products, and figures are preserved. Founder statements are source attributions, not independently verified findings.
AI Future Talks
AI Doesn't Replace Taste — It Exposes That You Never Had Any
In conversation with Nodesk.ai founder Song Jian (宋健): in the AI era, where exactly is a person's irreplaceability?
"
AI Coding has only two states — it hasn't started, and it can't stop.
I'm Wu Wei (吴畏), founder of Unique Research; I have no development experience, but a while back I tried to build a small tool with AI.
At first it really hooked me. Tell it what you want, it gives it to you. But soon problems came: endless bugs to fix, most of my time spent tugging back and forth with it. What came out was miles away from a real product, and I couldn't even say where the gap was. Most anxiety-inducing of all: I didn't know when to stop.
This week on Wu Xiaobo Channel's livestream I talked for nearly two hours with Nodesk.ai founder Song Jian. He said one line that immediately made my anxiety concrete:
"AI Coding has only two states — it hasn't started, and it can't stop."
The biggest feeling from that conversation wasn't what AI can do again, but that it is amplifying everything. It amplifies your ability, and your flaws. It amplifies your judgment, and your confusion.
Three Weeks in Silicon Valley: A Hangzhou Founder Mopping WiFi Outside OpenAI
Song Jian went to the US in May for three full weeks. SaaS industry conferences, Google I/O, early exploration of going-global business — a packed schedule.
After returning he told me about a "pyramid" model: compared with China, Silicon Valley's advantage splits into four layers.
At the very top is the information density of model frontier. He went near OpenAI, but OpenAI doesn't allow visitors in, so he checked in downstairs, connected to the WiFi, and talked a lot with Codex team members and some researchers. He said the feel of that area is a lot like early Hangzhou's startup ecosystem around Alibaba. Around OpenAI and Anthropic, a whole swarm of startups has sprung up, some already valued at tens of billions.
One layer down is the richness of talent. Big tech is scattered across the Bay Area, and talent mobility still leads.
One more layer down is the volume of venture capital. An old topic, but he said the feeling was still strong this time.
The bottom layer is the startup atmosphere.
I asked him: compared with past US trips, what's the biggest change? He gave an answer that surprised me: involution.
"Over at OpenAI, 70 hours a week is the baseline. 80 to 100 hours isn't unusual either."
We used to say Chinese founders are known for being involuted. But what Song Jian felt this trip is that AI iterates too fast — a model update may come every three weeks. Over in the US, from big-tech AI Labs to startups, everyone is pushing hard.
His words: "No longer will we say only China is involuted and America isn't. Chasing a dream means no slacking; everyone must build on effort and think about how value is created."
The AI Coding Rollercoaster: Emotion Ran Ahead of Ability
Song Jian has lived AI Coding firsthand. Nodesk.ai was founded in March 2025, wanting from day one to embed AI into the coding workflow. But the actual journey was nothing like outsiders imagine.
March to June: nobody wanted to use it. Cursor was still very early, Claude 3 wasn't out yet; the capability couldn't carry it.
June to September: pushed forward with difficulty. Product design and requirement understanding still had to rely on people; AI only slowly edged into development.
September to December: explosion. Claude upgraded to 4.0+, capability jumped, usage surged.
December to now: became normal. Colleagues don't just use it to write code; they use it for many daily things.
Too many people now swing between two extremes. Either they think AI can do everything, full self-driving; or they cling to traditional development, scrutinizing everything with a critical eye, not daring to let AI touch a single line of code.
"Neither state is right," he said. "Too conservative is definitely wrong; technological revolutions have proven again and again they aren't a joke. But you can't just leap in either, dropping all guardrails at once."
I feel this firsthand. I tried AI coding myself; what came out was miles behind others', then endless bugs waiting, not knowing when to stop. After listening, Song Jian said something that stung a little:
"It's unfair to blame the AI Coding tool. It only demos a rough tool with you; you can't really take it and use it as WeChat."
He went on: "Behind WeChat lies a whole system. Your questioning ability and thinking must at least match Zhang Xiaolong's. A WeChat that merely looks like WeChat isn't the same as an industrial-grade WeChat; how do you take responsibility for the consequences?"
I pressed: what if the model itself has already learned from the world's best product managers? Do I no longer need that ability?
He paused and said: "It doesn't have my true feelings."
FDE: A Necessary Path for AI Landing, or Outsourcing in a New Jacket?
This year a concept got very hot — FDE, Forward Deployed Engineer. Palantir coined it first; now OpenAI and Anthropic are also building similar teams. In short: send engineers on-site to truly wire AI capability into the client's business systems.
What Song Jian is building at Nodesk.ai is, in a sense, a Chinese version of FDE. What they deliver to enterprises isn't just an Agent product, but how to embed that Agent into real business loops — the client's e-commerce operations, customer service, product selection, listing and delisting.
I asked him: what's the essential difference between FDE and the software outsourcing and private deployment China has done for twenty years?
He unfolded it along a timeline from digitalization to intelligence. 2000 to 2010 was the IT era, like large complex machinery that had to be customized by IBM and Oracle. 2010 to 2020 was the cloud era; going on cloud was the foundation for SaaS. But China's problem is that it entered the next stage before fully finishing each one. It rushed to digitization before finishing informatization, to intelligence before finishing digitization. The debt owed at each layer looms larger under AI.
"Every enterprise has specific needs and pain points; one API can't solve them. FDE inevitably must exist."
I agree with this logic. The need is real.
But a real need doesn't mean a healthy business model.
The core dilemma of China's 2B market over the past twenty years isn't that no one understands clients; it's that many companies are dragged into the swamp of project-based, heavy-delivery, low-reuse work. Every client says it's special; sales promise customization to close the deal; delivery keeps patching holes to pass acceptance; the product team is torn apart by project demands. In the end the company is busy and has revenue, but with low margins, long cycles, and poor scalability.
Renaming it FDE won't automatically solve these problems.
I think FDE's real dividing line is: is it feeding project experience back into the product, or just packaging human delivery with AI?
More directly: after delivering one client, do you leave behind a pile of project documents, or a set of industry Agent capabilities portable to the next client?
If it can't be reused, FDE is a cost center. If it can, FDE is a product-evolution mechanism.
Palantir's FDE works for one key precondition: behind it sits the strong product base Foundry. FDE isn't a swarm of consultants rushing on-site to hand-build requirements; it's wiring client needs into reusable platform capability.
Frankly, many Chinese AI startups say they're learning Palantir but may actually just be putting an Agent hat on traditional software outsourcing.
Nodesk.ai serves three industries at once: e-commerce, industrial manufacturing, and medical tech. Song Jian says their edge is professionalism, stability, and cost-performance; I don't doubt that. But I care more: serving three industries at once, how high is the reuse rate?
This may be the soul-searching question every Chinese AI company doing the FDE model can't avoid.
The Taste Debate: AI Doesn't Replace Taste, But It Exposes You Never Had Any
This was the fiercest part of that night's discussion.
Song Jian's core view: a person's irreplaceability lies in taste and empathy. He put it bluntly, "it doesn't have my true feelings." AI can write code, make content, analyze data, but it can't feel that the livestream room's AC is too cold, nor understand why a certain product proposal "just isn't right."
I agree with this direction. But I think we need to dig one layer deeper.
The "taste" many people talk about doesn't survive dissection.
You say a proposal lacks refinement — is it actually chaotic information hierarchy? Out-of-control visual density? Text too full? No whitespace? Or just not hitting the user's mind? If you can only say "it feels off" but can't say where, that's not taste, that's intuition. Intuition has value, but it's fragile. Once AI's output quality reaches 85 points, most people's intuition gets drowned.
What taste is truly hard to replace? It's the ability to make integrated judgments in ambiguous scenarios. What to keep, what to cut, what's good on data but shouldn't be done, what has no short-term ROI but long-term value. This isn't just aesthetics; it's a combined decision of aesthetics, responsibility, scenario, risk, and consequences.
Song Jian gave a great example: "In a photography expert's hands, the iPhone 17 is a great camera; to me it may just be a phone."
Same tool, totally different output. The difference isn't in the iPhone; it's in the person.
But here's the key question: a photography expert beats an ordinary person not only because he "has taste," but because he knows light, composition, subject relationship, emotional expression, and post-processing boundaries. A large part of these can actually be broken down and trained.
So my judgment is a bit harsher than Song Jian's: AI won't replace real taste, but it will replace a lot of experiential routines that people mistook for taste.
If your taste is just following trends, copying paradigms, applying templates, then it really isn't irreplaceable. AI is faster, steadier, and cheaper than you.
I asked Song Jian whether taste is innate or learned. He hesitated and said it may be closer to innate.
I understand what he means, but want to add a layer: taste's ceiling may have a talent factor, but its floor can be raised through training. What ordinary people should really do isn't comfort themselves with "I have true feelings," but train their judgment to be concrete enough.
Only when you can break "a feeling" into judgeable details does taste turn from a fig leaf into a real ability.
Song Jian also mentioned a judgment I strongly agree with: facing the same Agent capability, the gap between people may not shrink but widen. This exactly confirms the "amplifier" logic. AI amplifies what you already have. If what you have is systematic judgment, you'll fly. If what you have is only experiential routines, you'll be left behind.
The Super-Individual Illusion: One Person Plus a Pile of Agents May Be a New Kind of 996
"One person plus a pile of Agents" is this year's sexiest startup narrative. No team to hire, no meetings, no management — one person orchestrating a flock of AI, doing what once needed a whole team.
Song Jian says the MVP stage can indeed work. One person strings together all the processes, calls Agents to finish work that once needed an entire team. He also mentioned Peter of OpenClaw, one person wielding dozens of Agents to write a product.
But there's a huge survivorship bias hidden here: people who can do this were never ordinary to begin with.
They usually already have engineering ability, product judgment, architectural sense, and extreme self-drive. AI is an amplifier for them, not a lifeline.
What about ordinary people? I'm a ready example. A pile of Agents to me isn't a pile of employees; it's a pile of uncontrolled sources. More bugs, messier versions, denser information, more exhausting judgment. In the past I couldn't write it; now AI writes a pile of things I don't know are usable. Capacity isn't insufficient; it's overstocked but uncontrollable in quality.
Song Jian openly admitted this too. He said when one person coordinates different Agents, the information density and overload "exceed the working cognition built up over more than a decade." Some people lack architectural ability and don't even know where the problem is, running around putting out fires.
I asked him how many Agents one person should have. He said it depends on that person's bandwidth, then listed four factors: questioning ability, taste, architectural ability, plus physical stamina.
Physical stamina.
Think about it and it really seems to be true. AI raised the task concurrency, sped up feedback, and amplified human judgment pressure. This isn't ease; it's a higher-density kind of labor.
So the story "AI makes people freer" only holds for some people. Those with clear goals, a judgment framework, architectural ability, and a sense of boundaries are indeed freer. For those without these, AI may bring not freedom, but faster anxiety and denser information.
You could say: AI doesn't turn everyone into a super individual; it fits a bigger lever onto the few who already have a systematic ability. As for everyone else? By day tuning Agents, by night fixing Agents, in the middle of the night doubting whether they're managing AI or AI is managing them.
What If the Answer Is Already There Tomorrow?
Near the end I asked Song Jian what daily habits he uses to train his thinking.
His answer surprised me.
He said he notes problems encountered by stage in a notebook, looks at them every day, tries to think about them, but doesn't force solving them today or tomorrow. Unsolved problems stay hanging, never deleted.
"Most startup problems have no solution, or no solution right now. If you drive yourself to death every single day, you won't make it to tomorrow."
Then he said:
"
What if the answer is already there tomorrow?
After recording the livestream I kept thinking. In the AI era everyone runs hard, fearing being eliminated, fearing not keeping up, fearing falling behind. But Song Jian, a founder who soaks in AI every day, leaves himself space precisely by: don't rush, let it hang, let problems walk with you for a while.
AI is indeed amplifying everything. Amplifying ability, amplifying the gap, amplifying anxiety, amplifying speed.
But maybe it's also amplifying something else: the value of patience.
When everyone is using AI to speed up, the person who can slow down and think clearly about the problem itself may go further instead.
Selected Q&A
Q1: On these three weeks in Silicon Valley, what was your biggest impression?
Song Jian (Nodesk.ai founder): I have a kind of triangular cognition. At the top is still the density of intelligence cognition; the impact and stimulation are stronger than at home. In the middle is the richness of talent — whether downtown San Francisco or the Bay Area, big tech is distributed at every point. Then the bottom has two layers: one is the volume and richness of venture capital, the other is the startup atmosphere. I still feel the whole Silicon Valley startup atmosphere is most worth looking forward to.
Q2: Compared with past US trips, what's the biggest change this time?
Song Jian: Involution.
Over at OpenAI, 70 hours a week is now the most basic work time, and 80 to 100 hours is really happening. We used to say Chinese founders are known for being involuted, but this time I could feel OpenAI, Anthropic, including these big-tech AI Labs, iterate extremely fast; a model update may come every three weeks. Startup app companies are very involuted too, extremely so.
Q3: How did Nodesk.ai's AI Coding evolve step by step?
Song Jian: We learned by doing.
In March there was an idea: can AI do design, development, testing across the whole flow? Back then Cursor wasn't mature, Claude 3 hadn't come out. March to June was very slow; nobody wanted to use it. June to September was using it with difficulty. September to December was explosive use, after Claude entered 4.0+. December to this March became normal. Now many colleagues don't just use it as a coding tool; they use it for many daily things.
Q4: What's the biggest challenge of AI Coding?
Song Jian: The biggest challenge is how to govern and scrutinize AI. Its overflow capability in Coding is fairly strong; as the one issuing instructions, how do you ask better, even sharper questions? Otherwise it's just two states: hasn't started, and can't stop. Once you start, it can write this and that, unstoppable. We've stepped in quite a few holes here.
Q5: What's your team's age structure?
Song Jian: Just over 30 people, five Gen-X (post-1980), everyone else is post-1990, among whom those born 99-05 are about 60%.
Q6: Can people without a development background use AI Coding to make products?
Song Jian: Our product-and-dev team genuinely has deep mobile-internet-era experience; even new grads are pure science backgrounds. Nobody is fully cross-domain. The product-and-dev team is actually expert-oriented, then uses AI Coding as an amplifier. But other roles — HR, finance, commercialization, operations — are all proactively embracing this change.
Q7: Is it necessary for ordinary people to hand-build Agents? Do you encourage it?
Song Jian: Hand-building is the best behavior for understanding capability evolution. But that doesn't necessarily mean owning the ability turns you into an expert. In a photography expert's hands the iPhone 17 is a great camera; to me it may just be a phone. System stability plus overall cost-performance, plus continuous industry know-how. You must understand AI and the industry; you can't rely on stitching both sides with code alone.
Q8: What is an AI-native organization?
Song Jian: I'll say it in three layers. First, in mindset, AI First — believe AI can help you solve problems better. Second, in methodology, practice beats ideas; you must have soil that lets AI trial-and-error, unlimited Token, allowing waste and experimentation. Third, in tooling, introduce differentiated AI capability into the company. None of the three can be missing. Also, AI First doesn't equal AI All-in; between them there's still the human.
Q9: What's the difference between AI First and AI All-in?
Song Jian: Too many people put AI and people in opposition. AI First is, when facing something, first thinking what AI can help with and what value it creates. AI All-in is, in everything, thinking how to replace people. The directions are completely different.
Q10: How does an AI-native org form differ from a traditional one?
Song Jian: Everything is different. Before it was treadmill collaboration — understand requirements, form a team, different roles, set OKRs; maybe 7 people form 1 squad. Now the starting point may be one person, one person stringing everything together, 1 person forming 1 squad. After validating PMF and running MVP, scaling still needs team fission to amplify. But before MVP, one person can indeed call Agents to do what a team once did.
Q11: How many Agents is reasonable for one person?
Song Jian: It depends on the person's bandwidth. Questioning ability, taste, architectural ability, plus physical stamina. None can be missing. Architecture splits into information architecture, code architecture, and business architecture. Information architecture is your understanding of the whole; code architecture prevents a mountain of spaghetti code from arriving at unprecedented speed; business architecture measures ROI.
Q12: Where does taste come from? Innate or learned?
Song Jian: I really want to say learned, but I do feel taste may be harder to develop later than questioning ability. It's the crystallization of a composite.
Q13: Why did people become more involuted in the AI era, doing more work?
Song Jian: A few reasons. First, not subtracting well; the concurrency window widens, blowing up physical stamina. Second, one person coordinating different Agents; the information density and overload exceed working cognition built over more than a decade, psychologically and physically overloaded. Some is because architectural ability is insufficient, not even knowing where the problem is, turning into running around putting out fires.
Q14: Where does the name Nodesk.ai come from?
Song Jian: Two layers of meaning. Macroscopically, not being defined by an established table, not defined by a track. Microscopically, breaking away from taking the human as the absolute starting point; it isn't necessarily just one person at one desk.
Q15: For an ordinary super individual, what do you think they should do most?
Song Jian: First, throw cold water on yourself; let yourself off, don't set too idealistic a goal. Small and vertical may be most worth doing. Second, don't mythologize AI, don't smear it either; the best way is to use it every day but not be addicted, keep it normalized. Third, don't work so hard. Lie back for two days; the unsolved problem may solve itself. Make peace with yourself, and your stamina, architectural ability, and taste will rise in a spiral.