The core distillation of Wang Kai's three years of investment practice. From the failure of FastFacts SEC filing automated trading in 2023, to the launch of the RSI AI strategy in 2025-05, and the disciplined system that scaled capital to 250,000 USD by 2026-Q1. In one sentence:AI is not a fortune teller that predicts ups and downs, but a ruler that quantifies and prices fuzzy risks; what you earn in investing is not money from market rises, but money from market sentiment.
1. Core Mindsets (memorize them)
| # | Mindset | One-sentence explanation |
|---|---|---|
| 1 | Good asset, good price, good volatility | Three layers of filtering, none can be omitted. A good asset is the prerequisite, a good price is the margin of safety, and good volatility is the entry signal |
| 2 | AI is a ruler, not a fortune teller | Let AI quantify fuzzy variables that humans cannot accurately measure (risk level, intrinsic value, margin of safety), rather than predict stock prices |
| 3 | Absolute rationality + strict discipline | AI has no fast thinking, only thinking. Hand decision-making power over to rules, and hand rules over to AI |
| 4 | Earn money from risk pricing | Not money from good companies' stock price rises (that is what Buffett earns), but money from the market's mispricing of short-term risks |
| 5 | Waiting > Thinking | It is not hard to think things through clearly; it is hard to wait for the right price. Any company can be waited on to reach a fair price, and most can be waited on to reach an undervalued price |
| 6 | Understanding = being able to act on it | Intellectually understanding versus behaviorally understanding are completely different. Being able to hold firm and add positions when down 50% is true understanding |
| 7 | View the present from the perspective of AGI | Assume AGI arrives on 2035-01-01; looking back, which of today's decisions will be proven correct? Do not bet on the main AI trend; instead, bet on supporting sectors (the MCN approach) |
| 8 | Token ROI perspective | The economic value of AI helping Jack Ma make decisions versus helping you write an article is worlds apart. Proactively move toward "high-value decision links" |
2. Three-year evolution timeline
graph TD
A[2023-04 FastFacts 立项<br/>抓 SEC 公告 + AI 决策 + 自动交易] --> B[2023-10 内测 3 万美金<br/>发现量化 team 比我们快 5 秒]
B --> C[2023-11 关项目<br/>承载资金天花板 10-30 万美金]
C --> D[2024 重仓 Apple<br/>错过 NVDA 是最大错误<br/>2022-12 就坚信 AI 但没买英伟达]
D --> E[2024-08 AiPathly + Quanta Quest<br/>仍是 AI 应用层创业思维]
E --> F[2025-Q1 大转折<br/>站在 AGI 视角看当下<br/>从做 AI 应用退回 AI 投资]
F --> G[2025-02 AI 套利探索<br/>跑了量化老路]
G --> H[2025-05 RSI AI 策略 v1<br/>5/28 首笔买 PDD<br/>1 万美金起步]
H --> I[2025-06 引入 好资产好波动 理念<br/>放弃量化技术路径<br/>3 万美金]
I --> J[2025-08 RSI v2 季度收益 4.3%<br/>多邻国亏损<br/>限定千亿美金以上龙头]
J --> K[2025-09 自建金融 MCP<br/>从 FMP API 拿准确数据]
K --> L[2025-11 升级 Claude 125 美金/月<br/>10 万美金资金池]
L --> M[2025-12 KeepRule 上线<br/>RSI v3 引入 内在价值计算<br/>18 万美金]
M --> N[2026-01 升级 250 美金/月<br/>RSI v4 三层框架完整<br/>21 万美金]
N --> O[2026-02-05 SaaS 大跌防御<br/>切换微软纳指<br/>保现金 45%]
O --> P[2026-05 owner-earnings.vercel.app<br/>KeepRule 26 大师原则库<br/>25 万美金]
Key turning points
- 2023-12-29 reflection: First systematic record: "Heard a big shot say the only AI-related thing he did was invest in Nvidia, didn't take it to heart", realizing that investing is the fastest way to realize cognition
- 2024-02-19 WLD reflection: Firmly believe in cognition; you must invest both time and money. Cognition without action is failed cognition
- 2025-04-01 (highlighted, pinned post) View the present from the perspective of AGI:
- Do not bet on points along the main AI trend (mainstream logic loses money)
- Do not even bet on points directly related to AI (the competition timeline is too long)
- Assume AGI arrives on 2035-01-01; looking back, what actions taken now will be proven correct
- 2025-08-02 epiphany: The real deep significance of the AI wave is not "intelligence equality", but "rationality/slow thinking equality" — the vast majority of people do not think through the vast majority of their decisions
- 2025-11-12 Duan Yongping interview baptism: Distinguish between "logically understanding" and "behaviorally understanding"
- 2026-01-03 behavior defines understanding: Being able to act on it is understanding; if you incorporate a discipline perspective + basic IQ + some common sense, it is very hard to lose money in investing
3. RSI AI strategy: detailed explanation of the three-layer framework
3.1 First layer: Good asset (prerequisite)
Initial definition(2025-Q2): Market cap above 100 billion USD + monopoly in a niche sector + long history + past volatility proving long-term value
Pitfall corrections:
- Duolingo (700 → 320 USD) kept falling; after buying ten times in a row he realized it "was not a good asset", a negative example of the consecutive buying discipline
- SaaS (CRM/NOW/ADBE/INTU) crashed in 2026-02 → review: 3-year revenue certainty was fine, but the market was already pricing in panic about the "endgame of AI replacement"
- Motorola Solutions (MSI) had a 70 billion USD market cap after its spin-off, Nokia 40 billion USD; counterexamples remind: major transformations do not happen with a single "snap"
Final rules:
- Prioritize above 100 billion USD (including special exceptions like Moutai, SPGI, etc.)
- Must have a monopoly/absolute advantage in its niche sector
- Exclude assets that "meet scores in the short term but cannot pan out" (such as Duolingo, early IBM)
- Do not buy inverse ETFs (such as SQQQ triple-short Nasdaq), which violate the premise that "good assets go up long-term"
- Do not buy small and mid-cap stocks (completely avoid anything below 30 billion USD)
3.2 Second layer: Good price (margin of safety)
Core concept: "Being cheap enough is not the key; understanding well enough is the key" (Duan Yongping)
Calculation tools: Two open-source Vercel apps built by Wang Kai himself
- owner-earnings.vercel.app - Intrinsic value calculation framework (synthesis of multiple classic models)
- qoderqianyimeijinrsi.vercel.app - Buffett strict version (most companies are directly passed)
Calculation framework features(polished repeatedly over a 6000-word prompt):
- Four-quadrant of revenue/cash flow/moat/management
- Give ranges, not single points (to avoid precise errors)
- Calculate the "return/risk" ratio, not just look at absolute price
- Sony test: runs a diversified, non-standard company in 5 minutes, with quality exceeding that of junior/mid-level analysts
Discipline:
- Any company can definitely be waited on to reach a fair price
- Currently evolving toward "any company can definitely be waited on to reach an undervalued price"
- Do not place high-price orders for quick fills (placed multiple high-price orders on ASML above 680 → now at 1500, a hard lesson)
3.3 Third layer: Good volatility (entry signal)
What AI does here: Score "risk" (1-5 scale, 5 is most dangerous, do not buy)
Risk variables(built into the prompt, for AI to quantify):
- Whether the company's fundamentals have actually changed (versus short-term sentiment)
- Macro risks (Fed/tariffs/wars/black swan events)
- Industry risks (replacement by AI/regulation/competitive landscape)
- Liquidity risk
- Position concentration
Typical cases:
- Boeing: government launched investigation → risk rating 2 → strategic buy → risk lifted → close (5%+ return)
- Apple: Trump tariffs → short-term drop → buy → rose 10% 3 days later, close
- Alibaba: rumor of a Hangzhou businessman surnamed Ma being a spy → short-term 17% drop → buy → risk eliminated, recovered
- ISRG: long-term 2-month decline → stable risk rating (fundamentals unchanged) → kept adding positions → single-day 20% rebound, large profit
- VZ: duplicate order placed by mistake after fill → immediate stop loss (correct when discipline is violated)
- Duolingo: bought ten times in a row but risk rating never issued a rising warning → negative example, limit target size
Capital allocation rules(dynamic):
- Risk 1: full position buy (40,000 USD)
- Risk 2: 3/4 position
- Risk 3: 1/2 position
- Risk 4: reduce position
- Risk 5: clear position
3.4 Overall position management
graph LR
A[总资金 25 万] --> B[策略仓 50-55%]
A --> C[现金 45-50%]
B --> D[股票 70%]
B --> E[BTC/加密 15%]
B --> F[纳指 QQQ 15%]
C --> G[等更便宜的好价格]
C --> H[防极端风险<br/>疫情/战争/恐慌]
4. Four uses of AI (ranked by importance)
Use 1: As a ruler (most core, most valuable)
Let AI use a stable framework to quantify fuzzy variables that humans cannot quantify:
- Market risk level ("How panicked is the market now? Score 1-10")
- Corporate intrinsic value (synthesis of multiple models → give ranges)
- Single stock risk (score separately for short, medium, long term)
- Emotional state (you can also let AI score whether you are "overly panicked vs rational")
Analogy: An actuary at an insurance company; after giving AI a stable framework it becomes a ruthless actuarial machine. As AI penetrates deeper into the "decision-making" link, its essence is to reduce the cost of human irrational decisions.
Use 2: Interrogate the decision-maker
Prompt: Let AI act as an investment master, ask deep questions to interrogate trading decisions ("Why buy? What is the basis? What are the risks? Why this price? Why is the position this large?"), guiding thinking mode.
Effect: Force "fast thinking decisions" to become "slow thinking decisions". Even if you do not end up using AI to make decisions, running decisions through this extra step avoids a large amount of emotional trading.
Use 3: Data analysis assistant (Claude Code / Cursor)
Key shift: Wang Kai realized in 2025-09 that "Claude Code is not just for writing code, it is a powerful data analysis tool", and has since used AI Coding as his primary data analysis tool.
Typical scenarios:
- Pull 10 years of historical data for a strategy → validate hypotheses (the "largest market cap company" strategy underperforms the Nasdaq)
- Run Berkshire cash ratio vs S&P 500 PE history → calculate future risk cycles
- Compare 30-year annualized returns of Moutai/Apple/S&P 500/Berkshire → Moutai earned 1x more than the S&P over 30 years
- Run signal statistics: how many buy signals appear per month on average, what average holding period is optimal
Prerequisite: Must build your own financial MCP (using FMP API), to bypass the flaw that Claude/GPT web search cannot get numbers right.
Use 4: Assisted decision-making (L2 driver assistance)
Wang Kai repeatedly emphasizes: At this stage it is not L4 full autonomy, it is L2 assistance.
- I control the route myself (stock selection criteria, position/risk limits)
- AI helps me quantify, calculate, interrogate, and check
- Do not let AI directly decide buys/sells (except in ultra-high-frequency signal scenarios, which are already occupied by quant teams)
5. Negative lessons (never commit these)
5.1 On AI tools
- ❌ Do not use DeepSeek to trade stocks: Inaccurate data, severe hallucinations, no decision coherence
- ❌ Do not directly trust numbers from Claude/GPT web search: Financial data accuracy is extremely low, it may fabricate numbers outright
- ❌ Do not keep asking about trades in one single conversation: Large model context length limits → early context is lost → decisions become disordered
- ❌ Do not pile all N requirements into one prompt: Instruction following will break down
- ❌ Do not ignore potential bias in web search: Information sources related to small-cap stocks are all self-media/interested parties; conclusions after AI synthesis are untrustworthy
5.2 On investment behavior
- ❌ Do not use AI to run stock screening reports just to confirm your own ideas: This is essentially using AI for psychological comfort
- ❌ Do not let AI co-create "complex analysis frameworks" to impress yourself: Retail investors have no basis to judge framework quality; complexity does not equal correctness
- ❌ Do not compare returns to the three major indices: That is a vague goal; losing money is losing money, profit is the goal
- ❌ Do not buy just because a stock feels cheap: Good assets can keep falling; you must calculate intrinsic value first
- ❌ Do not apply "track logic" to the secondary market: That is a VC primary market playstyle; retail investors who inherit it get hurt
- ❌ Do not try to backtest the entire strategy: AI strategies are inherently unbacktestable, but data at each node can be verified
- ❌ Do not predict short-term stock prices + do not draw conclusions based on short-term stock prices: Short-term stock prices are distorted and do not provide a basis for review
- ❌ Do not modify order prices just to get a quick fill: It loses money every time; the VZ case triggered the immediate stop-loss rule
5.3 On the AI investment industry
Wang Kai has repeatedly attacked industry noise:
- Douyin AI stock learning machines (hardware for retail scamming)
- Using AI to run "700% get-rich-quick returns" (the ex-OpenAI guy's PR is wildly overblown)
- HKUST AI stock trading paper (selecting Nvidia datasets to run cannot validate strategy potential)
- Caixin cover story "AI Investment Revolution" ("AI is a millisecond-level trader" is wrong; reasoning models take at least 2 minutes)
6. Value investing learning path (Wang Kai's)
Reading order (as actually practiced by Wang Kai):
- Duan Yongping Xueqiu (3 minutes a day, subscribed via Songshu Kuaikan)
- Buffett shareholder letters
- "The Intelligent Investor" - Benjamin Graham → cures the "track logic" disease
- "One Up On Wall Street" → "Don't trust your inner feelings"
- "Trader Vic: Methods of a Wall Street Master" → inspired "waiting is the greatest discipline"
- "The Most Important Thing" - Howard Marks → defines risk as "volatility ≠ risk"
- "Poor Charlie's Almanack"
- "Thinking, Fast and Slow", "Metacognition" → read deeply in coordination with thinking mode
- Himalaya Capital COO interview → the high standard of "understanding" in value investing
Reading takeaway: Classic books must be re-read, because previously insufficient ability (like a small model) meant much information could not be received. After AI makes "thinking mode the default", reading the same books reveals completely different things.
7. Differences from typical schools
| School | Core | Capital size | Wang Kai's stance |
|---|---|---|---|
| Buffett/Duan Yongping value investing | Buy companies + hold long-term + calculate next 10 years of cash flow | Large capital, low turnover | Learns the mindset but cannot execute it (looking 10 years out is too hard) |
| Quant | Mathematical models + large computing power + high frequency | Hundreds of millions of USD and above | A losing domain, retail investors have no chance |
| AI arbitrage(early Wang Kai direction) | Catch sudden events like SEC filings + AI instant judgment + automated trading | 300,000 USD ceiling | Speed is locked down by 2 professional teams |
| RSI AI Strategy(Current) | Good assets + AI pricing risk + short-to-medium-term volatility | 250,000 USD → 10 million USD in the future | Currently doing |
| Buy following Buffett's holdings | Copying homework | Arbitrary | A single logic works but cannot earn excess returns |
The relationship between the RSI AI strategy and value investing:
- Duan Yongping's perspective is 10-year cash flow; the RSI perspective is 1–2 month volatility
- Buffett earns money from the rise of good companies themselves
- RSI earns money from the market's short-term mispricing of risk (the "mania vs. depression" market sentiment)
- But the prerequisite for both is the same: good assets, which you must understand
8. Tool Stack (as of 2026-05)
AI Models
- Claude Pro $250/month(Absolute core): judgment, decision-making, documents, Claude Code
- ChatGPT Plus: GPT-5 thinking mode, research reports, long tasks
- Gemini: Deep research, NotebookLM (essential for reading papers/podcasts)
- Do not use: DeepSeek (never use for investment)
Financial Data
- FMP API(Paid, monthly fee around $100): Self-built MCP, bypassing the flaw of AI web search being unreliable
- Xueqiu (Songshu Kuaikan) subscription to Duan Yongping: 3 minutes per day
AI Coding (already a basic skill)
- Claude Code $250/month(Terminal, main force)
- Multiple instances of Cursor / TRAE (ByteDance overseas version) / Qoder (Alibaba)
- VS Code
- Xcode to check APP effects
Self-built Tools
- owner-earnings.vercel.app Intrinsic value calculation
- keeprule.com Investment master principles library (26 masters)
- global-liquidity-dashboard Liquidity dashboard
- fmp-market-structure-radar Overall market structure radar
- buffett-wiki.vercel.app Buffett principles Demo
Browser Automation
- Chrome MCP (controls a real browser, not headless)
- AgentMail (Agent-dedicated email, YC S25)
- Parallel multi-browser, multi-account system
Psychological / Decision Support
- IFS (Internal Family Systems) — introduced mainly by Wang Kai's wife, to treat the impulse to "violate discipline and change unit prices"
9. Performance Data (as of 2026-05)
| Period | Close Return | Capital Pool | Return Rate |
|---|---|---|---|
| 2025-05 (first PDD trade) | Launch | 10,000 USD → 30,000 USD | - |
| 2025-06 | $580 | 30,000 USD | ~2% |
| 2025-07 | $440 | 50,000 USD | - |
| 2025-08 | $927 | 50,000 USD | ~2% |
| 2025-09 | $992 | 50,000 USD | - |
| 2025-Q3 cumulative | $1,700 | - | Quarterly 4.3% |
| 2025-10 | $2,500+ | 50,000 USD | - |
| 2025-11 | $1,510 | 100,000 USD | Outperformed the three major indices |
| 2025-12 | $2,200 | 180,000 USD | - |
| 2026-01 | $2,200 | 210,000 USD | Monthly ~2% |
| 2026-02 | $4,000 | 250,000 USD | - |
| 2026-04 (Iran War) | Single QQQ trade $1,400 | 250,000 USD | Significantly outperformed the index |
Comparison Benchmarks(30-year average annualized):
- S&P 500 ~10%
- Nasdaq ~12%
- Moutai ~30% (over 30 years, cumulative return is 1x more than the S&P)
- Berkshire ~20%
- Buffett's historical average ~20%
Goal: Within 2 years, run this strategy on $10 million of own + partner capital.
10. Discipline Checklist for Followers
Before executing the RSI AI strategy, you must ask yourself:
- Do you have your own investment discipline? Or are you purely copying homework?
- If Duan Yongping sells Moutai and says his view has changed, will you sell?
- Do you understand tech stocks better than Buffett? Why do you dare to invest when he doesn't?
- Is your buying logic based on "calculating the company's intrinsic value" or "feeling the stock price is cheap"?
- If a position has a 40% floating loss, can you add to it? Or will you waver?
- Can you spot errors in the reports AI gives you? Or can you only trust them?
- Does the financial data you run come from the FMP API or from Claude web search?
- If you violate discipline once, can you stop loss immediately (the test of duplicate VZ orders)?
- Can your total capital withstand a 50% loss? Is your position size reasonable?
- Can you wait several months without trading?
11. Memorable Quotes from Wang Kai (memorize 5, enough for a lifetime)
"AI is a ruler, not a fortune teller. It is a tool for taking the temperature of a manic market."
"The hard thing in investing is 'always believing.' Buying is easy; holding through tests is hard."
"I understand some things, and there is more I don't understand. That gives me great certainty."
"Buffett earns money from the development of enterprises themselves; I earn money from the market's mispricing of risk."
"Waiting is the hardest thing, harder than thinking things through clearly."
"You should consume more tokens, not save tokens. No matter how expensive it is per month, it is cheaper than missing one decision."
"The most valuable ability in the AI era is 'the efficiency and quality of exponentially consuming tokens.'"
"The process of doing things is actually not that important; there are only 3–5 key decisions." (Li Yong of Yuanfudao version)
References and Extensions
- Wang Kai's AI-era entrepreneurship methodology and business judgment framework Entrepreneurship methodology (Internet-era dividend period vs. AI-era multi-project parallelism)
- Wang Kai's AI industry judgments and observations on big-tech competitive landscape Industry judgments (ecosystem judgments on OpenAI/Anthropic/Google/Apple/Microsoft)
- Wang Kai's AI Coding practice: from Cursor to Claude Code terminal workflow AI Coding practice (Claude Code data analysis, self-built MCP)
- Wang Kai's AI-era decision-making system Decision-making system (thinking, absolute rationality, AI-assisted decision-making)
- Wang Kai's Agent engineering practice Agent engineering (Chrome MCP, automated tasks, Claude Code Loop)
- newtype AI business, industry and investment observations Comparison with Lao Huang's newtype investment logic (more long-term oriented + personal entrepreneurship)
- wiki-yitang-scraping Yitang data scraping (related to the investment data acquisition approach of this wiki)
History
- 2026-05-16: Read all 433 posts in full (bucket-investment-strategy.md, 9615 lines) → topic-driven rewriting, reorganized by "mindset / timeline / three-layer framework / AI usage / negative lessons / tool stack / data / quotes", removing the original article's chapter order