Episode 2 of 雨森的创投观察 (Yumori's Venture Capital Observations) on the Zhang Xiaojun Podcast, featuring Dai Yusen, partner at Zengge Fund Management. It follows Episode 1 (episode 124 of the program), where Dai named the year's keyword “The Year of R” — a year of reality and return to the original state — and was more cautious than ever, emptying all his stocks in the second-tier market before the New Year. Half a year on, some of those opinions look off the mark. After a brief inner struggle, he decided to keep recording and sharing investment thoughts, framing the project as joint progress in 2026 with AI. The episode covers rapid AI model and company churn, the emerging “Harness as operating system” analogy, startup advice in fast-changing technology, the “strong opinion, weakly held” principle, and why being “hit in the face” is normal — even enjoyable — for an early-stage investor. It also discusses the unresolved return-on-investment question for massive token expenditures, the distinction between models and “harnesses” (agentic products), coding as a horizontal enabler, organizational changes needed for AI adoption, the rise of agent-to-agent marketplaces, and the long-term trajectory toward an AI-native digital world.
Key Points
Podcast context: Episode 2 of 雨森的创投观察 on the Zhang Xiaojun Podcast; source: https://www.youtube.com/watch?v=XEhf371Aeso. Episode 1 was episode 124, where Dai called the year “The Year of R” and emptied all his stocks in the second-tier market before the New Year.
Model landscape shifts quickly: OpenAI, Google, Anthropic, and others have traded leadership in coding, revenue, and user growth multiple times in 2025.
“Strong opinion, weakly held” – adapt views when underlying reasons change; market feedback is like reinforcement learning.
OpenAI’s revenue pieces: subscription growth has slowed; advertising progress has been slower than expected; corporate coding revenue surged unexpectedly due to agentic coding advances.
Anthropic’s strength in coding: coding was not an initial intention but emerged as training data included more code. Anthropic’s top-down organization (e.g., Dario Amodei’s biweekly memos) contrasts with OpenAI’s angel-investor-like resource allocation.
Harness vs. model: a powerful model alone is insufficient; a long-term harness (like Claude Code/Cloud Code, Codex, OpenCore) collects high-quality user data, enabling a data flywheel. Harnesses can be built by non-model companies.
Harness as abstraction layer: just as Windows/DOS and iOS abstracted away hardware, AI is undergoing the same shift. Three years ago an AI developer had to build the agentic loop, deal with memory, and handle guardrails — the “Harness” pieces inside the model. With Claude Code / Harness technology, you supply a skill or a shell and the harness runs the agentic loop. The harness becomes like an operating system and the model like a processor.
Return question unresolved: massive AI hardware profits ($7T this year) and token spending (Antelope AR $1B/month) require real profits; payback may take years, and many companies have already reduced token usage.
Coding is horizontal, not vertical – it strengthens office work, medical work, and research. The bottleneck is innovation, not programming capacity.
Agent era metrics: shift from “attention is all you need” (DAU) to “attention is not all you need” – agents liberate human attention; key metric is how long an agent can complete valuable tasks.
Ecosystem evolution: three stages – make agents for humans, adapt agents to the human digital world, build a native digital world for agents.
Organizational change required: like the shift from steam engine shaft to electric wires enabling factory flow lines, AI demands organizational restructuring, not just tool adoption.
Investment focus: invest in excellent entrepreneurs (small genius, old driver, scientist, supermodel) before trends become obvious; favor first-tier companies with differentiated views.
New Labs are popular but not universally favored; China is hardware-focused (robots, world models) while US sees more research-oriented New Labs.
Horizontal vs. vertical: avoid “following the trend steadily” (稳稳地跟随); do a genuinely big innovation. In early technology waves, horizontal capability lets you adapt to each new wave and find market information sooner; vertical SaaS is safer in later stages but can trap you when technology moves fast.
Taste vs. execution: AI separated ability from execution and judgment. Execution became increasingly AI-based, but judgment — and taste — remained the open question. Taste might also be replaced by AI if done correctly, since AI can go to the lab to get rewards; but at present AI can’t move by itself.
Being “hit in the face” is normal and enjoyable for an early-stage investor. Mr. Wa of GEEK convinced Dai to continue recording: if you think your ideas are being ignored and stop saying them, you are really bound by your ideas — you should keep learning and evolving.
Deliberate practice with AI: start with simple projects to build proficiency; outsource execution but not understanding.
Apps cited: Poverty, Manus, GenSpark, OpenCloud, Hermes. Lovable, formerly “Dijon,” is also expanding in parallel.
Concepts
Harness: An agentic product that adds context, tools, memory, and agent loops around a base model. Examples: Claude Code/Cloud Code, Codex, OpenCore, Minus, Manus, GenSpark, Hermes. The harness becomes the OS, the model the processor.
Agentic loop: The control loop previously built by each AI developer; now supplied by the harness so developers can provide only a skill or shell.
Agent (Agenda): An autonomous entity that can be given long-duration tasks without requiring human attention; the main form of AI value delivery in 2025 and beyond.
Strong opinion, weakly held: Borrowed from the second-tier market; hold a strong point of view but do not be kidnapped by it.
Horizontal vs. vertical: In early technology waves, horizontal capability lets you adapt to each new wave and find market information sooner; vertical SaaS is safer in later stages but can trap you when technology moves fast.
“The Year of R”: Keyword from Episode 1: a year of reality and return to the original state, accompanied by cautious positioning and emptied stock holdings.
Reinforcement learning analogy: Expressing thoughts to get feedback, including negative signals; the market as an honest feedback environment.
Taste vs. execution: AI separated ability from execution and judgment; execution is increasingly AI-based, judgment and taste remain open questions, though taste might also be replaced by AI if done correctly.
Second-tier market: The secondary market, source of the “strong opinion, weakly held” principle and a reference point for company momentum.
Return question: The unresolved challenge of whether the massive capital expenditure on AI tokens (hardware and inference) will yield real profits for end customers, not just model providers.
In-distribution vs. out-of-distribution (OOD): AI excels at solving problems within the distribution of existing human data (e.g., coding, writing). Truly novel (OOD) creation – like original jokes or new mathematical theories – remains a human strength.
New Lab: A specially funded, free-form research organization focused on exploratory AI research, often separated from existing model company structures.
Details
Episode Context and “Being Hit in the Face”
Dai Yusen opens by acknowledging the difficulty of returning for Episode 2. Episode 1 (episode 124) had him naming the year “The Year of R” — a year of reality and return to the original state — and he was more cautious than ever, emptying all his stocks in the second-tier market before the New Year. Half a year later, many say some of those opinions look off the mark. After a brief inner struggle he decided to keep recording and continue sharing investment thoughts, framing the project as joint progress in 2026 with AI.
Mr. Wa of GEEK convinced him to continue. The argument: if you think your ideas are being ignored and stop saying them, you are really bound by your ideas — you should instead be someone who keeps learning and evolving. Dai likens this to a non-standard learning method, much like reinforcement learning: you express thoughts, get feedback, and need high-quality feedback signals, sometimes negative ones. The market is valuable because it will not flatter you — if the market says you are wrong, you are wrong, unlike daily life where few people criticize you directly. Being slapped in the face means the industry is changing very fast; fast change in early-stage investment also means many opportunities. Many people are being hit in the face by AI.
On being “hit in the face” (打脸), he told Mr. Wa he would not record next time, because that would limit his ability to make a flag (立flag). Mr. Wa immediately said he was convinced — because if you don’t record after feeling you’ve been hit, you really are beaten. Keeping to summarize his thoughts is a process that lets him think about unthinkable problems and maybe think about them more clearly. You can practice: when you express something, you find some places you clearly know that you understand (or think you understand), while other places are confused — this is very clear when you say something out loud. He sees this as a process of his own self-organization, and very valuable. As an early investor, being hit in the face is often a happy thing — “enjoy being slapped.” Will they keep doing this? Yes; next time they may be hit again. He feels this is normal, because early investment is really a habit of being hit.
Model Shifts and Revenue Dynamics
In the past six to eight months, model leadership has oscillated: OpenAI dominated in November/December (DAO of OpenAI cited at ~$800M–$1B), Google’s live model (GMI Live 3) emerged in December, Cloud’s coding ability peaked in January, Zopic’s (likely Claude Opus) revenue grew very fast and surpassed OpenAI in March (it also exceeded OpenAI in the secondary market, with the March trend described as “Zopic at 100 million”), and by May Codex had more new users than Cloud Code, though GPT‑5.5 also showed strength and Opus 4.7 came down somewhat (attributed to a series of system problems). OpenAI is seen in a rollback or “you chase me, I chase you” pattern. An OpenAI cooperation announcement moved a company’s share price by tens of points, as did announcements about who would buy whose profit (e.g., Oracle).
Dai Yusen’s earlier prediction that OpenAI’s revenue would decline was partially accurate (subscription growth slowed, advertising was slower than expected) but wrong about corporate coding revenue, which grew significantly because of agentic coding advances enabled by Cloud 4.5 and 4.6. He subsequently adjusted his portfolio, adding storage, CPU, and bottleneck hardware investments.
Anthropic vs. OpenAI Organization
Anthropic is relatively closed to China; most information is second-hand. Its organization is top-down: Dario Amodei issues a thinking memo every two weeks, and every interviewee undergoes a value‑aligned interview. This contrasts with OpenAI, where many celebrity researchers pursue individual directions and resource allocation resembles an angel investor (Sam Altman, former YC president). Anthropic’s consistent direction helps it “control sand and bury” (post‑release maintenance), whereas OpenAI products often lack post-release protection. Yusen notes that Anthropic’s choice of coding as a specialty was not intentional – coding performance improved naturally as training data included more code.
The Harness Becomes the OS
Cloud Code, Codex, Minus, OpenCore, Hermes, Manus, GenSpark, and OpenCloud are all “harnesses” that wrap a model with context, tools, agent loops, memory, and guardrails. The model is the processor; the harness is the operating system. Just as Windows/DOS and iOS abstracted away hardware — developers stopped handling CPU memory and IO interfaces and dealt only with APIs, and mobile developers never had to think about communicating with Apple’s processor or camera — AI is undergoing the same shift. Three years ago an AI developer had to build the agentic loop, deal with memory, and handle guardrails; that made app development hard. With Claude Code / Harness technology, you supply a skill or a shell and the harness runs the agentic loop for you. In the Windows era you could use an Intel CPU or an AMD CPU — plug in whichever, cheaper or more expensive, whichever you preferred — and similarly users can now plug a model into a harness and build new capabilities on top of that OS.
Rendering diagram…
This is roughly the common sense emerging around building applications on top of Claude Code, or Claude Code becoming an OS: you install Claude Code but instead of using it directly you use an application running on Claude Code. Claude Code is a CLI, terminal-based, for many people; many are not opening the terminal to call the model directly but are running a map interface on the Claude Code runtime. It is characterized as an interesting trend this year.
Brand lock-in is strong: users invest heavily in configuring tools like Cloud Code, reducing motivation to switch. Codex has engaged in a price war, offering similar capabilities at ~50% lower cost than Cloud Code’s equivalent tier, but GPT‑5.5 narrowed the gap. OpenCloud innovations include running on Mac, accessing files/calendar, a heartbeat mechanism, and memory MD (single context organized daily). It lives inside familiar IMs (WeChat, WhatsApp, Discord), a key reason for its popularity.
Open Source and Innovation
OpenCloud exemplifies how open source can attract users and later monetize via paid versions. Security concerns (tied to computer, reading files) make big companies cautious, but startups can “move fast and break things.” The harness space also allows non-model companies to innovate – e.g., OpenCloud, Manus, and others were initially dismissed as “shells” but are now considered critical.
Return Question and Token Economics
The real chain is input (tokens) → output (software) → result (profit, revenue increase, or cost reduction). Current investment assumes results will materialize, but the link is unproven. Simply multiplying engineering capacity (e.g., ten times more engineers via AI coding) does not automatically increase revenue; the bottleneck is knowing what to build. Many companies that burned large token volumes in March have already reduced usage, suggesting failed value delivery. Hyperscalers like Amazon raise capital for data centers, shortening the time demand for answers. Antelope’s AR is expected to reach $1 billion per month by year-end, meaning $10 billion spent on tokens must eventually justify returns.
Coding as a Horizontal Field
Coding is not a vertical like medical or finance – it strengthens all fields. High-quality user coding data makes models better at programming. Agentic loops that previously could not run now can, enabling longer-term, higher-value tasks. Even for non-programming office work, AI can generate more output but does not replace human responsibility for decisions (e.g., investment decisions still require a person to bear the loss).
Organizational Impacts
New companies (e.g., Slack, Cloud Code, Codex) are built from day one with AI participating in management and software development, blurring divisions between front‑end, back‑end, UI, and testing. Established companies struggle because their context and data are
