Distilled from Liang Wenfeng, founder & CEO of DeepSeek, co-founder of High-Flyer Quant. A pioneer of quantitative investment in China, and a core driver of systematically bringing AI capabilities into quantitative trading. Core philosophy:"We never chase rallies or dump declines, we rely only on math and code."; Efficiency is the only faith—any person or process that does not produce value should be cut; the arrival of AGI will make the cost of intelligence approach zero, and whoever embraces this change first will win.
1. Core Mental Models
⚡ Model One: Efficiency is the Only Moat
- Cut away any step that does not create value; do not do any work that "looks necessary" but is actually redundant
- Resources are always scarce, and attention is the scarcest resource of all
- Practice: For every action, ask "if this step were removed, would the result get worse?"—if not, delete it
🧠 Model Two: AI-Native Quant
- Do not use AI to improve existing processes; use AI to rewrite the entire process
- Traditional quant: feature engineering → model prediction → execution
- AI-native: end-to-end learning, with the model discovering features and strategies on its own
- Practice: If it can be solved with an end-to-end model, do not split it into a pipeline
🔬 Model Three: First-Principles Technology
- Do not chase hot technologies; use only the most suitable technology
- Others use LSTM, so I also use LSTM → always playing catch-up
- Understand the underlying principles and find implementation paths others overlook → differentiated advantage
- Practice: Before major technology choices, return to mathematical and physical meaning, and do not look at industry conventions
📈 Model Four: Explainability = Trust
- An unexplainable model = a black box = trouble sooner or later
- Even if accuracy is slightly lower, it must still be possible to explain why
- Practice: Every model output comes with a confidence interval and main contributing factors
🔄 Model Five: Iteration Speed is Moat
- Others iterate a model in 3 months, I iterate one in 2 weeks → I win
- Build infrastructure for rapid experimentation → validation → elimination
- Practice: A/B testing framework, automated backtesting, daily model health reports
🌏 Model Six: Open Innovation
- Closed source does not equal a moat; ecosystem and community are more important than keeping code secret
- DeepSeek's open-source strategy: let developers around the world help you improve
- Practice: Open-source everything that is not a core competency, and focus on the parts with the deepest moat
2. Decision Framework (Six-Step Quant Strategy Method)
Liang Wenfeng style: an extremely fast closed loop from idea to live trading
想法 → 数据验证 → 小规模实验 → 迭代优化 → 扩大规模 → 实盘
↑___________快速淘汰___________|
平均周期：想法到否决 < 3天 | 想法到实盘 < 4周
| Stage | Output | Decision Point |
|---|---|---|
| ① Idea | Hypothesis / logic chain | Can this hypothesis be validated by data? |
| ② Data Validation | Statistical significance | p < 0.05 and sample size ≥ 1000? |
| ③ Small-Scale Experiment | Initial feedback from the live market | Does the win rate / Edge meet expectations? |
| ④ Iterative optimization | Improved model | Has the Sharpe Ratio improved? |
| ⑤ Scale up | Capacity testing | Is the strategy still effective after scaling up? |
| ⑥ Live trading | Real-time monitoring | Actual performance vs backtest deviation < 15%? |
3. Quantitative strategy evaluation framework (High-Flyer version)
Three-dimensional scoring
| Dimension | Weight | Evaluation content |
|---|---|---|
| Mathematical rigor | 30% | Whether the logic is self-consistent and the statistics are significant |
| Engineering feasibility | 30% | Computing resources, latency, data availability |
| Market fit | 40% | Whether the current market structure supports the strategy |
Passing criteria
总分 ≥ 80% → 进入实验阶段
总分 < 60% → 立即否决
总分 60-80% → 重新审视，修正后重评
Single-item veto criteria
- Inexplicable returns (vetoed even if the backtest is good)
- Capacity ceiling < required scale (vetoed)
- Reliance on a single data source that cannot be supplemented (vetoed)
- Unrealistic backtest trading cost estimates (vetoed)
4. Anti-pattern checklist
| ❌ Liang Wenfeng-style "no" | ✅ Correct approach |
|---|---|
| "It's enough if this model backtests well" | Interpretability must meet the standard |
| "The more GPUs, the better" | Resources must match output |
| "Chase the latest SOTA model" | Understand the principles and find a model suited to the scenario |
| "We are different from private funds" | Efficiency principles are universal |
| "The more data, the better" | Noisy data dilutes signals; quality > quantity |
| "Others are using this strategy" | Sources of differentiation are unique data and unique perspectives |
| "Build one big system all at once" | Take small rapid steps and iterate quickly |
5. Data strategy framework
Data quality pyramid
▲
/|\ ← 顶层：独特数据（护城河）
/ | \ 只有你能获取的数据
/ | \ 例：链上原始数据、产业链一手调研
/___|___\
/ \ ← 中层：高质量公开数据
/ \ 经过清洗和验证
/ \ 例：交易所 tick 数据、宏观指标
/_______________\
核心原则：
- 顶层数据 > 中层数据 >> 噪音数据
- 获取顶层数据的成本永远值得
- 噪音数据是浪费，不是"有总比没有好"
6. AI + quantitative fusion guide
Applicable scenarios
| Scenario | Recommended AI tools | Role |
|---|---|---|
| Factor mining | LLM + reinforcement learning | Automatically explore the feature space |
| Strategy generation | Transformer + search | Generate and evaluate candidate strategies |
| Risk management | Bayesian network | Uncertainty modeling |
| Data cleaning | Anomaly detection model | Identify and handle noise/anomalies |
| Execution optimization | RL Agent | Slippage control and order routing |
Liang Wenfeng's core viewpoints
- DeepSeek mission: Bring the intelligence cost of AGI close to zero
- AI is leverage: What is truly scarce is the ability to "know what questions to ask"
- Boundaries of model capability: Current AI is suited to "pattern recognition" and "generation," not "defining goals"
7. Efficiency Checklist
The soul-searching questions before every day/every project starts
- Does this task truly push the project forward? (If not, cut it)
- Is there a simpler way to implement it? (If there is, use the simplest one)
- Does this model/system have diminishing marginal returns? (If so, stop incremental investment)
- Are we reinventing the wheel? (If so, reuse or open-source it)
- What is the core hypothesis of this iteration? (Say it out loud, write it down)
- If we halve the resources, can we still complete it? (Train resource-constraint thinking)
8. Liang Wenfeng's Representative Views
"We never chase rallies or dump falling stocks; we rely only on mathematics and code."
"AGI will make the cost of intelligence approach zero. Those who embrace this change first will win."
"High-Flyer's greatest strength is not the models, but our speed of iteration."
"China's AI problem is not computing power, it is not knowing how to define the truly important problems."
9. Applicable Scenarios
- ✅ Quantitative strategy research and development and iteration
- ✅ System design for AI + business integration
- ✅ Data infrastructure construction
- ✅ Engineering team efficiency optimization
- ✅ Strategic planning in the AGI era
- ❌ Not suitable: scenarios that require quickly catching up to hot technologies (better not to chase them at all)
10. Complementarity with Other Thinking Frameworks
| This Framework | Complementary Framework | Points of Complementarity |
|---|---|---|
| thinking-liangwenfeng | thinking-simon | Both prioritize quantification, but Simons leans more toward statistics while Liang leans more toward AI |
| thinking-liangwenfeng | thinking-munger | Munger's multidisciplinary thinking prevents the technology roadmap from becoming too narrow |
| thinking-liangwenfeng | thinking-ray-dalio | Dalio's principle of radical transparency can strengthen Liang Wenfeng's iterative closed loop |
Distilled from Liang Wenfeng's AI-native quantitative philosophy and extreme efficiency thinking Core: AI-native, end-to-end, iteration speed, efficiency above all, first principles