A 3-year record of AI industry evolution. The quality of Wang Kai's industry judgments has been repeatedly verified by time (the evolution of OpenAI vs Microsoft vs Google, Apple's AI path, the rise of Anthropic, the Scaling Law debate, the AI hardware wave, the SaaS crisis).
1. Core Judgments (sorted by confirmation time)
1.1 Early high-certainty judgments (most verified)
| Time | Judgment | Current Status |
|---|---|---|
| 2023-04 | MJ will definitely be banned + foreign AI image generation bans | ✅ Verified |
| 2023-05 | Multiple domestic large models will coexist, there will be no leading stock | ✅ Verified |
| 2023-09 | The wave of AI changing hardware will definitely come, and Apple will seize it | Partially verified (Tesla FSD/iPhone 16 AI) |
| 2023-10 | Chinese AI entrepreneurs face "four big mountains" | ✅ Verified |
| 2023-11 | SD cannot survive independently in the long run | ✅ Verified (acquired/atrophied) |
| 2023-11 | Search engines transition from "providing information" to "providing answers + solutions" | ✅ Verified (SGE) |
| 2024-02 | Apple is bound to be the biggest winner in the AI era | Pending (slow execution) |
| 2024-02 | The value of "brand equity" rises significantly in the AI era | ✅ Verified |
| 2024-03 | RAG is a data goldmine, vertical RAG has a future | Partially verified (vertical AI search is widespread) |
| 2024-05 | Google AI implementation > OpenAI (product orientation > technology) | ✅ Verified |
| 2024-06 | Siri is a certain ticket to the AI era, eating up all shallow-interaction apps | ⚠️ Apple executes slowly |
| 2025-02 | Domestic AI products can no longer charge fees | ✅ Verified |
| 2025-03 | Manus's viral popularity is mainly due to narrative + public sentiment | ✅ Verified |
| 2025-09 | AI replacing jobs is still in the early stage, large-scale manifestation will occur in 1-3 years | Under verification |
| 2026-02 | SaaS valuations are affected by AI replacement panic, most of the panic is excessive | Under verification |
1.2 Judgments Wang Kai admitted he got wrong
| Judgment | Reason for error | Lesson |
|---|---|---|
| In December 2022, firmly believed in AI but did not buy NVDA | Did not convert cognition into investment returns | "Cognition that is not monetized equals no cognition" |
| Search would be immediately impacted after ChatGPT launched | Underestimated the lag effect | "Change is cumulative, not achieved overnight" |
| Smartphones would definitely replace feature phones immediately | Same as above | Nokia still has a market capitalization of 40 billion US dollars today |
2. Role Positioning of Large Companies
graph TD
A[AI 时代竞争层级] --> B[硬件层]
A --> C[系统层]
A --> D[模型层]
A --> E[生态层]
A --> F[应用层]
B --> B1[苹果 - 软硬一体生态<br/>未来 AI 硬件最大赢家]
B --> B2[OPPO/小米 - 中国本地化 AI 手机]
B --> B3[Tesla / 机器人 - 物理 AI]
C --> C1[iOS - 苹果 Siri 系统级 AI]
C --> C2[Android - Google + Gemini]
C --> C3[Windows - Copilot]
D --> D1[OpenAI - 模型领先 + 应用平台化]
D --> D2[Anthropic - 纯 AI 路线 + 工具链]
D --> D3[Google - 全栈数据 + 生态]
D --> D4[国内: DS / 千问 / Kimi / MiniMax / 智谱]
E --> E1[Chrome + Google 全家桶]
E --> E2[微信 / 抖音 / 飞书]
E --> E3[Salesforce - 企业数据生态]
F --> F1[Cursor / Devin - Coding 应用]
F --> F2[Cursor 类 Agent 应用]
F --> F3[小创业 team - 极细分点]
3. Evolution of the Big Three (OpenAI / Anthropic / Google)
3.1 OpenAI's Path
- 2023: A pure technology company, Sam = the Pope of AI
- 2024: Started building an ecosystem (GPTs / Store / 4o)
- 2025: Started building a product platform (Custom GPTs / Apps in ChatGPT) + launched ChatGPT Agent (with dedicated model optimization)
- Wang Kai's judgment: Is copying Google's ecosystem route, but retains model leadership → the strongest among independent large model companies in the long run
3.2 Anthropic / Claude's Path
- 2024: Pure AI route, model quality is king
- 2025: Three major ecosystem layouts: MCP / Skills / Code, Claude Code changes development collaboration
- 2026: Claude Code is a foundational collaboration tool in the AI era
- Wang Kai's judgment: Pure AI route, most aligned with the token economy, but needs to bind with ecosystems such as Google to counter OpenAI's platformization
3.3 Google's Path
- 2023-2024: Was considered "lagging in AI", but Wang Kai repeatedly said watching Google's launch events is more useful than OpenAI's
- 2024-2025: Gemini model capabilities are rapidly catching up
- 2026: Integrated Gemini into the Google full suite + Stitch + Veo + Genie,ecosystem advantages are beginning to show
- Wang Kai's judgment: "Google is a company where product orientation > technology orientation, and it ranks first in the world in AI implementation"
3.4 Who will win? Wang Kai's "context engineering" perspective
- Context engineering becomes a consensus → Google holds the most high-quality context (Search/Gmail/Drive/YT)
- Alibaba / ByteDance / WeChat are also huge context holders
- Independent large model companies must connect to data ecosystems to survive
- In the future, users' GPT/Claude will look the same on the surface, but the connected data/tools will be different, leading to vastly different efficiency
4. Model Capability Evaluation (Wang Kai's Discipline)
| Model | Wang Kai's regular usage scenarios | Evaluation |
|---|---|---|
| Claude (Pro $250/month) | Decision-making, judgment, long-form analysis, Coding (Claude Code) | Highest quality for investment judgments, but API data is inaccurate; will upgrade to client + terminal + Skills in 2026 |
| GPT-5 / GPT-4o / o3 | Long tasks, reasoning, research reports, market analysis | High quality in thinking mode, became the main workhorse again in 2025 |
| Gemini | Deep research, NotebookLM (essential for reading papers/podcasts) | High cost-performance, "even Old Huang uses Perplexity" |
| DeepSeek | Do not use! Absolutely cannot be used for investment | Severe hallucinations with numbers when using web search |
| Qwen / Doubao / Kimi | For experimental use | Restriction to Chinese information sources is an inherent flaw |
Key disciplines:
- Use the best model
- Compare multiple models (open Cursor / Qoder / TRAE simultaneously)
- Do not save tokens, use the most expensive subscription
- Use thinking mode more often
5. The Three Scaling Law Debates (Wang Kai's Stance)
| School | Claim | Representatives |
|---|---|---|
| Brute-force Aesthetics School | More data + more computing power = higher intelligence = AGI | OpenAI / Anthropic / Google |
| Application Layer School | Large model upgrades will not eat up all opportunities, application scenarios are the core | Zhu Xiaohu / domestic VCs |
| Ceiling School | Scaling Law is already experiencing diminishing marginal returns | Kai-Fu Lee / Cohere CEO |
Wang Kai's stance: Leans toward the Brute-force Aesthetics School (firmly believes AGI will arrive before 2035), but acknowledges the application layer still has space when it can solve "details that AI is not good at".
6. Five Levels of AI (OpenAI 2024-07)
graph TD
L1[L1 Chatbot<br/>对话能力<br/>2022-2023] --> L2[L2 Reasoner<br/>专家级推理<br/>2024 o1/r1]
L2 --> L3[L3 Agent<br/>多步任务自主<br/>2025 Devin/Claude Code]
L3 --> L4[L4 Innovator<br/>原创性突破<br/>?]
L4 --> L5[L5 Organization<br/>完整组织<br/>?]
Wang Kai's current positioning:Transition period from L2 to L3, consumer-side Agent explosion still requires token prices to drop another 90%
7. AI Hardware Wave
Key judgments
- Phone manufacturers will definitely participate: AI phones are the third revolution in the mobile phone industry (OPPO advertorial tone)
- Apple is the ultimate winner: All-round advantages in hardware + system + ecosystem + brand, but slow execution
- AI hardware review/media ecosystem: Currently in a blank period (analogous to early mobile phone review bloggers)
- Physical AI: The next wave as described by Jensen Huang, robots integrating into daily life
Wang Kai's investment targets
- Apple (core) → later exited position
- ISRG (Intuitive Surgical medical robots) → sold after position surged 20% in 2 months of holding
- Tesla (FSD + robots) → ongoing adjustment
- Do not buy OPPO / Xiaomi (insufficient cognition, not monopolies)
8. SaaS Crisis Judgment (after the 2026-02 crash)
Market panic logic: AI replaces labor → layoffs → SaaS charges per seat → revenue declines → valuation collapses (CRM/NOW/ADBE plummeted)
Wang Kai's rebuttal:
- No deterministic problems within 3 years
- The real risk is not the present, but AI changing reading and writing(everyone wears AI glasses, no need to open CRM)
- But SaaS giants have proprietary data + channels + services, and will still hold advantages in the AI era
- Buying SaaS now is a "good asset, good volatility" opportunity
Comparison with Google's reaction to the AI wave:
- When ChatGPT went viral in 2023, Wang Kai believed Google's search monopoly would break → in reality Google did not plummet
- The current SaaS dip is only a bit of panic, clearly showing the market's increased trust in AI
9. Redefining "Risk" in the AI Era
Wang Kai's summary after studying Howard Marks' *The Most Important Thing*:
- "Volatility ≠ Risk"
- Real risk is fundamental changes and the disappearance of moats
- AI provides a quantitative framework for "risk", allowing retail investors to price risk like insurance actuaries
5 variables for risk quantification(built into prompts):
- Whether the company's fundamentals have actually changed
- Macroeconomic risk (Fed/tariffs/wars/black swan events)
- Industry risk (AI replacement/policy/competitive landscape)
- Liquidity risk
- Position concentration
10. Evolution of "Commercial Monopoly" in the AI Era
| Internet Era | AI Era |
|---|---|
| Platform monopolies (Taobao/WeChat/Douyin) | AI Agent platform monopolies (ChatGPT/Claude) |
| Page Rank determines information distribution | Agent Rank determines Agent invocation frequency |
| SaaS intermediary monopolies (Salesforce) | End of intermediary monopolies, AI directly connects supply and demand |
| Content sites rely on search traffic | Content site weight shifts to AI training data |
11. Investment Framework (Three Layers)
- Technology stack layer(bottom layer): chips → models → inference → applications
- Value chain layer(middle layer): data → decisions → actions
- Competitive landscape layer(top layer): monopoly / oligopoly / red ocean
Wang Kai Only places heavy bets on monopoly/oligopoly layers(Apple, missed NVDA, Google, ISRG)
12. Old Huang newtype (Comparative Perspective)
- newtype Old Huang transitioned from PR/marketing to independent development, leaning more toward the personal/small business path
- Wang Kai leans toward investment/asset allocation + accounting
- Mutual corroboration: "What you should do in the AI era is not AI itself, but the supporting infrastructure"
- Wang Kai repeatedly says: "View the present from the perspective of AGI", Uncle Huang says "AI makes 1-person companies possible"
13. Time Anchors (Biggest AI Industry Event Each Year)
| Year | Landmark Event | Wang Kai's Observation |
|---|---|---|
| 2022-11 | ChatGPT launch | "AI unlocks opportunities wave by wave" |
| 2023-03 | GPT-4 + domestic large model free-for-all | "Multiple large models will definitely emerge in China, no leader" |
| 2023-11 | OpenAI internal strife | "Sam is irreplaceable" |
| 2024-05 | GPT-4o | "Multimodality, camera input" |
| 2024-09 | o1 reasoning model | "Democratization of slow thinking" |
| 2025-01 | DeepSeek v3/R1 | "Reassessment of China-US AI competition" |
| 2025-03 | Manus | "Beginning of the Year of the Agent, sentiment > technology" |
| 2025-08 | GPT-5 | "Thinking mode as default" |
| 2025-11 | Popularization of Claude Code | "AI Coding is a foundational skill" |
| 2026-02 | SaaS crash | "The market has begun to value SaaS using the endgame of AI replacement" |
References
- Wang Kai RSI AI Investment Strategy and the "Good Asset, Good Price, Good Volatility" Framework - Investment Strategy
- Wang Kai's AI Era Entrepreneurship Methodology and Business Judgment Framework - Entrepreneurship Methodology (the other side corresponding to industry judgments)
- Wang Kai's AI Coding Practice: From Cursor to Claude Code Terminal Workflow - AI Coding Industry Evolution
- Wang Kai's Agent Engineering Practice - Agent Industry Evolution
- newtype AI Business, Industry and Investment Observations - Old Huang newtype Comparative Perspective
- newtype-ai-agent-architecture - Agent Architecture Evolution
- newtype Model Evaluation, Fine-tuning and Hardware Evolution - Model Comparison Benchmarks
2026-06 Addendum: AI Costs, Market Growth and Software Entry Points Should Be Judged Separately
The following are phased views from four of Wang Kai's posts in June 2026. The text has been read in full; financial report and stock price charts have not been verified, and predictions do not equal outcomes that have already occurred.
Do not attribute all market expansion to consuming more Tokens
On June 6, the author raised a counterexample to the narrative of "using more Tokens → generating more revenue → continuing unlimited usage". AI companies are in a new growth market, and revenue growth may mainly come from market expansion; for mature ride-hailing or advertising businesses, even if employee efficiency improves, constraints such as demand and ad inventory will not automatically disappear. For product managers, evaluating AI investment requires separately examining market increments, efficiency increments, and the ceiling of sellable resources. A rise in consumption itself cannot prove that revenue increments are brought by AI.
On June 15, he made a sensitivity estimate based on the news about Meta reining in AI spending: he tentatively took the "billions of dollars" mentioned in the report as 3 billion US dollars, and observed the order of magnitude of costs using the metric of approximately 38,000 US dollars per person per year, accounting for about 10% of salary. These were the values and calculation statements the author used at the time; financial reports or employee-related metrics were not rechecked for this piece. The method of converting total expenses into per-capita figures and salary proportion can be used for reference; however, the ratio of a single company cannot be elevated to an "absolute ceiling" for the whole of society. After technology costs and capabilities change, this assumption itself must also be recalculated.
The risk to software may lie in interaction entry points, not just in programming becoming cheaper
On June 1, continuing from Section 8 of this page: the author believed that the previous decline in software stocks included short-term sentiment, and the revenue of some enterprises had not yet been significantly affected; at the same time, long-term risks still existed. His concern was that people would no longer open CRM systems to query information, and instead AI assistants would proactively extract, process, and deliver information, so the interaction value between software and people would decline accordingly. The two layers of judgment can hold true simultaneously: the fact that short-term revenue does not collapse does not mean that long-term entry points are secure.
When examining your own product, you can ask separately: Do users still need to open it directly? Which link in data, permissions, security, or workflows still requires it to bear responsibility? This is business entry point analysis, not proving that a certain stock is cheap based on a price rebound.
Market capitalization differences and lock-up expiry dates are hypotheses to be tested
On June 22, the author proposed an explanation related to lock-up expirations and transaction structures for the market capitalization difference between Zhipu and MiniMax, and predicted that a decline might occur around July 9. This batch of materials does not include subsequent market trend verification. Such records should be retained as dated prediction samples for future comparison with results and other causes; they must not be written as proven laws in the knowledge base.