Zhang Xiaojun | 张小珺Jùn｜商业访谈录 Summary

Podcast: 张小珺Jùn｜商业访谈录 Episode: 142 Participants: Zhang Xiaojun (host), Dai Yusen (guest, Managing Partner of Upgrade Capital / 升格基金)

1. Key Themes

The "Return Problem" Has Not Been Solved — It's Just Been Deferred

Dai Yusen's central thesis is that while Anthropic's revenue surge looks like proof that AI is generating returns, it's actually just the first step in a three-part chain: Input → Output → Result. The AI industry has made extraordinary progress on the "input" side (token spending, GPU buildout), some on "output" (code generated, reports written), but virtually none yet on "result" (incremental profit for end customers).

"Anthropic's revenue is not the final return — it is actually the investment of its customers. People buying all those tokens are not burning them for fun — they want results in the end. So this pushes the return problem one step down the chain to Anthropic's customers. Those customers buy tokens and burn tokens, and what they want is the software they build. But software alone is not enough — that software has to sell, or has to reduce costs, and generate profit. That is the result." 00:23:49

"Right now the number is too big to ignore. If you look at the entire AI hardware spending this year — $700 billion. And Anthropic, by year-end, is expected to hit $100 billion ARR, meaning each month $10 billion of tokens are being purchased. The people spending that $10 billion — how much money are they making? That question needs to be answered in a relatively short timeframe." 00:33:26

Harness / The Orchestration Layer Is Becoming the Operating System

One of the most developed themes in the episode is that the "harness" — the layer wrapping AI models — is evolving into something analogous to an operating system, with models becoming like processors. This is not obvious because the market has historically dismissed these products as "just wrappers."

"I think the harness is becoming more valuable... Users are more loyal to their harness than to the underlying model. Many friends have the experience of constantly trying to rescue their Xiaolongxia [OpenCraw] — they keep reviving it using Claude Code or other means. Why? Because what they care about is the memory stored inside that harness. And people are willing to swap out the underlying model to cut costs, but they don't want to swap the harness." 00:43:57

"The harness is becoming the OS layer, and the model is becoming the processor layer. Just like in the Windows era, you could swap between Intel and AMD CPUs — whichever was cheaper or newer — that's very similar to how users today can plug and unplug models into their harness." 00:54:44

Coding Ability Is Horizontal, Not Vertical — This Changes Everything

A non-obvious realization that crystallized over the past year: coding was previously categorized as one vertical domain among many (alongside healthcare, finance, etc.). The insight that it is actually a horizontal capability that amplifies every other domain fundamentally changes its strategic importance.

"Even when I was talking with senior researchers at leading labs — if they weren't themselves working on coding — they would often say coding is just one vertical direction. But now they've realized: coding is not a vertical. Coding is horizontal. It can strengthen office work, strengthen healthcare, strengthen the pace of research. So the implication of coding is very different." 00:17:12

"Cloud Code is very important. If Anthropic had only sold its API without Claude Code, I think it would be a very different company. Claude Code created a massive flywheel of high-quality usage data, and from what I understand, this data feedback genuinely enabled them to train better coding models." 00:18:23

2. Contrarian Perspectives

Most AI 2C Consumer Applications Are a Trap — Don't Try to Out-ByteDance ByteDance

The prevailing excitement around AI consumer applications, particularly those modeled on information feeds and entertainment, is a misallocation of energy. Competing in ByteDance's game, using ByteDance's rules, is nearly impossible.

"If you make a product that looks like an AI information feed — where content is AI-generated and users scroll through it — I think this is very hard to create major opportunity from. Because in ByteDance's game, beating ByteDance is extremely hard. For example, if you're making an AI casual entertainment product, from day one your distribution is controlled by the existing champions. Your competitor from day one is Douyin and Kuaishou and you have to be better than them immediately — and they didn't have that challenge when they started." 00:40:24 01:39:25

"Many current AI 2C applications that everyone is chasing are probably not a good investment direction." 01:41:45

DAU Is the Wrong Metric for the Agent Era

The entire mobile internet era was built around maximizing DAU × time-on-app × monetization efficiency. In the agent era, this framework becomes actively misleading. The best product may actually reduce user visits.

"In the agent era, how many users open the app daily is no longer the most critical metric. Possibly more critical is the METR metric — how long can an agent continuously complete valuable tasks. So how to build a good harness that lets the model complete tasks for long periods becomes more important. Liberating user attention so the AI can work — that's the key." 01:22:32

"The absurd scenario is: you optimize for DAU, so you need users to come back and check — but possibly a better product is one where the user hands off the task and never has to look again. If a user's visits actually decline, it might mean your product is excellent at completing tasks without requiring attention." 01:27:16

Wearable AI Hardware Will Repeat the Mistakes of the "New Consumer" Wave

The current wave of AI-powered wearables (recording devices, AI pendants, etc.) is likely to disappoint for the same structural reasons that China's "new consumer" brand wave collapsed: adding AI to an existing category does not create a fundamentally new category.

"I think many AI hardware products will repeat the mistakes of the new consumer wave. Back in 2020, every product category got remade — skincare, toothpaste, everything. But people later realized that taking a traditional category, adding some small innovation, and then finding a livestreamer to sell it doesn't make it a fundamentally new category. Similarly, many AI hardware products haven't solved fundamental problems — they've added a chatbot or some AI perception capability, but have they created a genuinely new demand category? I'm skeptical." 01:55:49

"If you just give the hardware a chatbot or some AI sensing capabilities... the incremental value relative to the hassle [of wearing it daily] is very limited. To be blunt, it's often a demand invented by VCs — the VC thinks they're very busy and needs daily to-do reminders. But for most ordinary people, recording everything you do each day — the incremental value is quite limited." 01:57:16

The Next ByteDance-Scale Company Will Not Look Like ByteDance

A counterintuitive argument: founders (especially ex-ByteDance) searching for "the next ByteDance" by replicating its playbook are structurally disadvantaged because they're entering the champion's home court.

"Every revolutionary company in each era solved user problems with a distinctly different interaction model. OpenCraw doesn't even have its own app — it has no base of its own — but it lives everywhere: in your WeChat, your Telegram, everywhere. What I'm saying is: taking the successful paradigm of the last era — information feed, paid acquisition, monetization, traffic flywheel — and transplanting it into the AI era means you're competing directly with the champion of the last era. That's very difficult." 01:39:55

3. Companies Identified

Anthropic

Leading AI lab behind the Claude model family
Mentioned for: organizational discipline, data flywheel from Claude Code, rapid iteration cadence, enterprise focus that differentiated from OpenAI's consumer-first approach

"Anthropic — every person who goes in for an interview has a values interview. This is a kind of shared directional alignment. They're not so focused on individual heroism. Dario apparently syncs his internal thinking memos with the whole company every two weeks. It's a very aligned organization." 00:15:43

Manus (曼努斯)

AI agent / harness company (portfolio company of Dai Yusen)
Mentioned for: pioneering "wide research" (parallel multi-agent task execution), building sandboxed browser environments for AI, influencing Anthropic's Claude Code design

"Manus was doing wide research — what we now call agent teams or agent swarms — spinning up 10 or dozens of subagents simultaneously to execute different tasks. They were doing this before Claude Code and before others were doing agent teams. From what I understand, when Anthropic was building Claude Code, they actually drew on a lot of Manus's experience from building wide research." 00:49:05

OpenCraw (小龙虾 / OpenCraw)

Open-source local AI agent harness built by a single developer (Peter)
Mentioned for: multiple product innovations — heartbeat mechanism for scheduled tasks, persistent memory via a single continuous conversation thread (memoryMD), integration into messaging apps (WeChat, WhatsApp, Telegram), running natively on Mac with full file/calendar access

"OpenCraw runs on your Mac, so it can access all files on your Mac, access your calendar, your messages. These experiences are ones that a product like Claude Code — a CLI — doesn't have. The heartbeat mechanism lets it complete scheduled tasks. Its memoryMD system is very interesting — while ChatGPT and Claude Code all try to separate sessions to prevent context mixing, OpenCraw puts everything in one big conversation and uses daily memory consolidation. Researchers were initially skeptical, but users feel very strongly that OpenCraw remembers them." 00:50:04

Cursor

AI coding IDE
Mentioned for: proving that "wrapper" companies can build proprietary models using feedback data; Composer model trained using Claude as base + high-quality usage feedback data

"Cursor proved that even if you only have the 'shell,' the data that shell generates, combined with post-training, can produce a genuinely good model. Cursor's Composer — based on Kimi's pre-training and Cursor's own high-quality feedback data — is actually a quite good model. This shows that 'model-based product' and 'product generates model' can simultaneously be true." 00:42:07

Kimi (Moonshot AI)

Chinese AI model company (portfolio company of Upgrade Capital)
Mentioned for: becoming what Dai believes is possibly the world's best open-source coding model; significant strategic pivot toward agentic coding

"We invested in Kimi, which is probably the domestic company most focused on agentic coding as a technical model company. In 2025 they completed a massive transformation and are now, we believe, possibly the world's best open-source coding model." 00:35:25

5G (灵巧手公司 / Dexterous Hand Company)

Chinese robotics component company focused on dexterous hands
Mentioned for: early conviction on dexterous hands as the key bottleneck for robot manipulation; benefiting from the rise of egocentric data collection trends

"He started by making a very high-torque motor, then discovered that hands were crucially important — at a time when very few people were making hands, and even fewer were making high-degree-of-freedom dexterous hands. He said: the best robot manipulation will require human-like hands, human hand data is the easiest to collect, and hand costs will rapidly decline. Now with egocentric data training becoming dominant, dexterous hands that are isomorphic with human hands are becoming critically important." 01:14:38

Slock (投资组合公司)

Human-agent collaboration platform (portfolio company of Upgrade Capital)
Mentioned for: pioneering "build Slock using Slock" — all internal operations, task management, and coding done on their own platform; example of agent-native organizational design

"We invested in a company called Slock — it's like a Slack where humans and agents collaborate together. Their entire company operation, task management, and coding is done on their own platform. So they build Slock using Slock. This means AI participates in essentially all of their company operations from day one." 01:36:14

Stripe & Coinbase

Payment infrastructure companies
Mentioned for: proactively building agent-native payment infrastructure, issuing "credit cards" to AI agents

"Everyone has noticed that payment is a problem for agents. So companies like Stripe and Coinbase are issuing human-style credit cards to AI. Cloudflare, which used to block AI, recently launched a service letting agents register and use various services with equal access to humans." 01:29:52

Aura Ring & Whoop

Health wearable companies
Mentioned favorably in contrast to AI gimmick hardware — succeeding because they are genuinely health hardware that happens to use data intelligently, not "AI hardware" as a marketing label

"Aura Ring and Whoop — you could say they have some AI, but they're not really positioning as AI hardware. They're health hardware. I think the key is to be genuinely useful, not to claim you're worth a lot because you used AI." 01:58:05

4. People Identified

Dai Yusen (戴雨森)

Managing Partner, Upgrade Capital (升格基金)
Why mentioned: The primary guest. Known for early conviction in agents (invested in Manus and Kimi pre-hype), willingness to publicly revise views, deep first-principles thinking about AI economics

"As an early-stage investor, getting your face slapped [being proven wrong] is normal. When you keep getting proven wrong, it often means the industry is changing very fast. And if you're doing early-stage investing and the industry is changing very fast, that means there are many opportunities." 00:02:45

Stanley Druckenmiller (史丹利·德鲁肯米勒)

Legendary macro trader / investor; former lead portfolio manager for George Soros
Mentioned as Dai Yusen's personal idol for secondary market investing; embodiment of "strong opinion, weakly held"

"My idol in secondary markets is Stanley Druckenmiller — the former wheel-man for Soros, now running his own fund. He's a trader, not a fundamental investor. So my own secondary market style isn't 'buy a stock and hold for five years' either — it's more trading-oriented." 00:06:50

Dario Amodei

CEO and co-founder of Anthropic
Mentioned for: building a uniquely aligned, top-down organization with biweekly internal memo syncs; choosing to focus on enterprise/productivity rather than mass consumer products

"Dario apparently syncs his internal thinking memos with the whole company every two weeks. It's a very aligned organization — almost like some of China's high-combat-effectiveness companies." 00:15:43

Peter (OpenCraw founder)

Serial entrepreneur (previously sold a company for ~$100M); creator of OpenCraw
Mentioned for: embodying the ideal "AI product manager" — power user who noticed a gap (couldn't use Claude Code while eating), built a bridge from mobile to desktop agent, kept innovating

"Peter had previously sold a company for $100M, so he's already an 'experienced operator' in our taxonomy. While using Claude Code, he thought: when I'm eating, how can I continue using my Coding Agent? So he naturally built a hook connecting his Mac's Claude Code to his IM — controlling his Coding Agent remotely through IM. Step by step he built OpenCraw. He has very deep AI understanding, he's an extreme power user, and he understood the subtle technical differences that enabled experiences nobody else had." 01:20:49

Liu Songming (刘松明)

Co-founder of a world model / robotics company (portfolio of Upgrade Capital); Tsinghua PhD dropout; author of RDT series
Mentioned for: deep technical foresight, working on large-scale robot learning data collection (5-meter data) almost simultaneously with Jenialist; strong entrepreneurial spirit combined with research depth

"Liu Songming worked on large-scale collection of 5-meter [operation] data for model training — almost simultaneously with Jenialist — as well as Agentic RL work. They also have strong entrepreneurial spirit. Technical foresight plus entrepreneurial spirit is very important right now. And the willingness to give up a lot to actually start a company." 01:13:23

Ding Ning (丁宁)

Co-founder of a world model / robotics company (portfolio of Upgrade Capital); Tsinghua professor; author of Simple VLA series
Mentioned alongside Liu Songming as a next-generation young academic entrepreneur from Tsinghua with high research impact and strong founder qualities 01:11:22

5. Operating Insights

The "Strong Opinion, Weakly Held" Discipline Applied to Public Statements

Dai Yusen makes a specific and actionable observation: publicly stated views create a psychological trap that causes people to hold positions longer than the evidence warrants. The discipline is not just internal conviction-updating — it's managing the social cost of changing your mind publicly.

"When you express a view very forcefully, you become constrained by what you've said. And if it was said publicly, there are even more constraints. That's also why I was hesitant about recording this second episode. But Wazong from Jike persuaded me — he said: if you think your view was proven wrong and you're unwilling to speak again, then you are truly bound by your own view. But you should be a continuously learning, continuously evolving individual." 00:08:12

Operational implication: When making public commitments (investor letters, podcasts, team announcements), explicitly state the conditions under which you would change your view. This makes it easier to update without social cost and prevents anchoring to past positions.

"Organizations Can Only Bear a Fixed Amount of Accountability" — The Real Bottleneck to AI ROI

A precise and overlooked framework for why AI-driven productivity gains at the organizational level are slower than expected: accountability, not capability, is the binding constraint.

"I thought of a framework: the amount of accountability an organization can carry is limited. As long as AI cannot end-to-end complete a job, it can write ten times or a hundred times more reports — but a person still has to be responsible for the results of those reports and the actions they drive. AI cannot be held responsible. You cannot fire it. When the accountability a person can bear is fixed — determined by their salary or the losses they can absorb — just because AI writes better models doesn't mean that person can suddenly manage ten times more investments or bear ten times more responsibility." 00:31:54

Operational implication: When building AI into workflows, don't just measure output volume gains — redesign accountability structures so that AI completion of verifiable, bounded tasks can genuinely transfer responsibility, otherwise productivity gains will stall at the human accountability bottleneck.

Building Agent-Native Companies from Day One Creates Compounding Structural Advantages

Companies that build their internal operations on their own agent platform from day one accumulate an AI-legible context corpus that legacy companies cannot replicate.

"A company that's been running for 10 years — its biggest problem is that its context and data are not visible to AI. Getting that context in is extremely hard. But some new companies from day one have all their context AI-visible. AI deeply understands what they're doing... They build Slock using Slock. That means AI participates in essentially all of their company operations, task management, and software development from the start." 01:35:45

6. Overlooked Insights

Agent-to-Agent Marketplaces Will Create an Entirely New Pricing Layer Based on Accumulated Personal Context

This was mentioned briefly and treated almost as speculation, but it contains a profound investment thesis: as agents accumulate unique personal/organizational context, they will generate differentiated outputs for the same task — creating the conditions for an agent-to-agent economy where the premium over base token cost reflects the value of accumulated proprietary knowledge.

"My agent hires Zhang Xiaojun's agent. Say I need to prepare interview questions — my agent pays your agent. Maybe 1,000 yuan of that is the token cost, but 9,000 yuan is because your agent has accumulated your proprietary knowledge and experience. So your agent's output is much better than mine burning 1,000 yuan of tokens on the same task... Six months ago, there was no reason to hire your agent because your agent was running Opus and mine was also Opus — we both just burn tokens equally. But in the scenario I described, your context and pipeline information might never be distilled into the model — the model will probably never have Zhang Xiaojun's specific knowledge. So your agent produces something different from everyone else's." 01:01:07 01:02:34

Why this is significant and overlooked: The existing mental models for agent value (better model = better output) miss this. The real moat will be accumulated private context, not model quality. This points to investing in: (1) platforms that help individuals/organizations build rich agent memory, (2) infrastructure for agent-to-agent transactions (micropayments, trust, verification), and (3) companies whose product creates irreplaceable personal or organizational context repositories. No one else in the conversation flagged this as a major investment theme — but it is one.

The Transition from GUI-Optimized to CLI/API-Optimized Products Is an Underappreciated Platform Shift

Mentioned very briefly in the context of agent-friendliness, this is actually a major structural change that threatens incumbent software companies and creates greenfield opportunities.

"When your users are no longer primarily humans but primarily agents, maybe the number of humans opening your app daily is no longer that important — what matters is how many agents are using it. At that point your product design philosophy changes: you need to be agent-friendly, so maybe your CLI or API needs to be better, not your GUI needing to be fancy... In the future, maybe you won't need that software at all — I'll communicate directly with your underlying database. Then your interface becomes useless." 01:25:58 01:46:34

Why this is significant and overlooked: This is a direct threat to the entire SaaS GUI industry — not from AI replacing functionality, but from AI making the human-facing interface layer unnecessary. Companies building CLI-first or API-first interfaces for AI agents, or middleware that translates agent requests directly to data layers, are positioned to benefit from the obsolescence of traditional SaaS UI. The window to build before incumbents react is now.