Adversarial Deep Deduction: Ten Historical Thinkers Who Influenced Liang Wenfeng An Archaeology of the Cognitive Pedigree and Technical Evolution of the DeepSeek Founder Step 1: Questioning the Problem Itself This deduction introduces a distinctive problem absent in the Zhang Yiming version: the input data is virtually non-existent. Zhang Yiming left behind at least 20+ public interviews, internal letters, and speeches. Liang Wenfeng? There are only about 2 to 3 deep, long public interviews available (such as 36Kr in July 2024, and occasional brief appearances at Wuzhen or WAIC), totaling less than 3 hours of spoken material. Attempting to reverse-engineer an intellectual evolution from this is akin to illuminating a 500-square-meter hall with the light of three matches. Therefore, what you are asking me to perform is not a traditional “reverse deduction,” but rather reverse archaeology—reconstructing the cognitive molds from the observable outputs of DeepSeek (its papers, open-source models, and organizational style). This hypothesis is not fragile, but its single point of vulnerability is fatal: Hypothesis: DeepSeek’s technical direction and product forms can be entirely attributed to the intellectual throughput of Liang Wenfeng alone. If DeepSeek’s core breakthroughs (the MLA attention architecture, the MoE sparsification strategy, and the R1 pure RL reasoning path) are actually the result of collective emergence from a young, 200-person team, where Liang Wenfeng’s role was merely to “provide money, provide time, and stay out of the way,” then this list belongs to the DeepSeek Research team, not Liang Wenfeng. Your list would then become a CEO’s reading list masquerading as a team’s collective intellect. I know this hypothesis is likely wrong—Liang Wenfeng writes code himself, DeepSeek’s organization is highly flat, and core technical decisions are clearly top-down—but if it is even 50% true, the following rankings must be discounted by 50%. DeepSeek Research • 对抗式深度推演 / Adversarial Deep Deduction Step 2: Three Independent Failure Modes Failure Mode 1: Cognitive Pollution of Quantitative Trading Liang Wenfeng spent 8 years in quantitative trading (High-Flyer Quant, 2015–2023) before founding DeepSeek. These eight years were not just a career phase; they completely reshaped his cognitive framework: Quant Thinking = Everything is a Signal: Trends? Signals. Volatility? Signals. News sentiment? Signals. There is nothing that is not a signal. Strategies Must Be Backtestable: Do not do what cannot be verified or quantified. The Market is a Zero-Sum Game: Your alpha comes directly from someone else’s loss. Scale is Not a Virtue: Capacity is fundamentally limited, characterized by diminishing marginal returns. These mindsets are highly visible in DeepSeek’s technical decisions: In DeepSeek’s model outputs, every token is a “signal”—devoid of redundancy; the essence of the MoE architecture is to make irrelevant experts shut up (sparsification is noise reduction); R1 utilizes pure RL instead of RLHF because human feedback is too slow, heavily biased, and unbacktestable. The Trap: If you mistake a quantitative trader’s mental framework for “Liang Wenfeng’s worldview shaped by reading historical figures,” you confuse cognitive habits with intellectual lineage. If Liang Wenfeng had gone straight into AI without doing quant trading, he would have developed an entirely different methodology even if he read the exact same books. Failure Mode 2: Technical Dissemination Channel Fallacy Liang Wenfeng entered Zhejiang University in 2002 as a computer science undergraduate. His undergraduate years (2002–2006) coincided with the most barren period for public academic literature in the Chinese internet landscape. At that time, Chinese universities did not use Google Scholar, and arXiv had just launched (1999) with almost zero Chinese users. His most critical formative period (ages 18–25) means that the Western ideas he encountered were likely absorbed through MIT OpenCourseWare (launched in 2002), translated Chinese textbooks (likely 3–5 years behind the English editions), BBS discussions (Smth BBS, Yitahutu BBS), and printed versions of English PDFs. This implies that the intensity of the ideas Liang Wenfeng absorbed was determined not by “how important the thinker was,” but by what books were accessible in Building 5 of the Yuquan Campus at Zhejiang University in 2005. A niche author whose PDF was easily obtainable might have exerted a greater influence than an academic giant whose books required maritime shipping. Failure Mode 3: Self-Reconstruction Narrative of a Quant CEO Turning to AI In his 2024 interviews, Liang Wenfeng described himself as a “technological idealist” focused on “basic research,” “not doing applications,” and “caring about AGI.” But remember: when he founded High-Flyer in 2015, his goal was to build a Chinese version of Two Sigma or Renaissance Technologies, earning the highest risk-adjusted returns. • • • • DeepSeek Research • 对抗式深度推演 / Adversarial Deep Deduction His narrative shift is outcome-driven: because DeepSeek succeeded, the story that “I was pursuing AGI from the very beginning” becomes a coherent and justifiable narrative. If DeepSeek had failed, his autobiography might have been titled “A Quantitative Trader’s Gamble on GPU Compute.” For intellectual genealogy deduction, this is fatal: Liang Wenfeng’s retrospective claim that his inspiration came from Shannon, Hinton, or Turing might be true, or it might be a post-success narrative unification. The true cognitive curve might look like this: 2015 reading Markowitz (Portfolio Theory) → 2018 reading Knight (Uncertainty and Risk) → 2021 reading Sutton (Reinforcement Learning) → 2023 reading LeCun (World Models). Four stages, four people, none of whom spanned the entire journey. Step 3: Ten Thought Figures Ranking Principle: Not “who is the greatest,” but “whose ideas are most deeply embedded in DeepSeek’s observable product forms.” (Deduced with maximum credibility under the condition that the three failure modes are accounted for)

1. Alan Turing Observable Evidence: DeepSeek R1’s reasoning display is the inverse proposition of the Turing Test—it actively displays its thought process, letting you know it is not human, yet its reasoning path is clearer than a human’s. Intellectual Radiation: Turing’s 1936 paper On Computable Numbers defined what can be computed. Liang Wenfeng’s entire DeepSeek wager is built on a Turing proposition: “A sufficient number of computable steps can simulate human-level intelligence.” This is not a proven proposition, but Liang Wenfeng is betting billions of dollars and tens of thousands of GPUs on its correctness. Depth of Deduction: In his 1950 Mind paper, Turing proposed the concept of a “child machine”—instead of programming intelligence directly, write a program that can learn and let it grow on its own. DeepSeek R1’s pure RL approach (not giving the model human-annotated reasoning templates, but only providing right/ wrong feedback on the final answer) is the 2024 engineering realization of Turing’s child machine concept. Without Turing’s framework, DeepSeek as a company has no philosophical reason to exist. DeepSeek Research • 对抗式深度推演 / Adversarial Deep Deduction
2. Claude Shannon Observable Evidence: DeepSeek’s Multi-head Latent Attention (MLA) architecture. Deduction Logic: Shannon’s greatest contribution was proving that information can be treated mathematically—it has bits, entropy, and channel capacity. DeepSeek’s MLA essentially does the same thing: compressing the computational overhead of attention to its theoretical lower bound. The core breakthrough of DeepSeek V2—MLA reducing KV Cache by about 75%—is a Shannon-style information-theoretic optimization: removing redundancy, preserving mutual information, and maximizing information transmission efficiency under a fixed compute budget. Adversarial Point: Yet, Liang Wenfeng might have never read a single original page of Shannon. MLA might have been proposed by a DeepSeek researcher during a seminar, and Liang Wenfeng simply said “do it.” If a Shannon diluted through a technical team is credited to Liang Wenfeng, then this list is merely Liang Wenfeng accepting an award on behalf of his team.
3. Geoffrey Hinton Observable Evidence: DeepSeek is the most extreme executor of the Hinton school of thought (connectionism, distributed representations, backpropagation). Deduction Logic: Hinton’s lifelong bet—that a sufficiently large neural network, trained on enough data, will yield general intelligence—is DeepSeek’s doctrine. Hinton’s 2012 AlexNet opened the doors to deep learning; in 2023, Liang Wenfeng pushed Hinton’s bet to its limit: using the MoE architecture to scale parameters to 671B, and pure RL to drive reasoning capabilities close to OpenAI’s o1. Crucial Evidence: DeepSeek R1’s pure RL path (learning to reason from environmental rewards rather than human-annotated examples) is the engineering implementation of Hinton’s 2017 assertion: “Human supervision is too small; we need unsupervised learning.” Adversarial Point: Liang Wenfeng has never publicly mentioned Hinton’s name. This can be explained in two ways: (1) Liang Wenfeng does not want to reinforce the narrative that Chinese AI is merely following Western academic giants; (2) Hinton’s influence is a default setting of the technical environment for him (like the air he breathes). If it is (2), Hinton’s actual impact should rank first. DeepSeek Research • 对抗式深度推演 / Adversarial Deep Deduction
4. Richard Sutton Observable Evidence: The core technical route of DeepSeek R1 = Reinforcement Learning reasoning from scratch. Deduction Logic: Sutton’s mathematical foundations of reinforcement learning are more than papers; they constitute a worldview—an agent learns optimal strategies by interacting with an environment, without needing human demonstrations. DeepSeek R1’s reasoning training workflow (SFT cold start → large-scale RL reasoning training → RL for alignment) is a textbook application of Sutton’s framework. Deeper Evidence: Sutton’s famous 2019 essay The Bitter Lesson—“70 years of AI research shows that general methods that leverage computation eventually win out, while special-purpose methods leveraging human knowledge only work in the short term”—is the philosophical bedrock of DeepSeek’s entire strategy. Liang Wenfeng follows it unreservedly—open-sourcing models, releasing weights, and building no application barriers. If ranking by influence on irreversible decisions, Sutton belongs in the top three.
5. John von Neumann Observable Evidence: DeepSeek’s Mixture of Experts (MoE) architecture design. Deduction Logic: Von Neumann defined the template of a computing architecture (memory, control unit, arithmetic logic unit, input/output). DeepSeek’s MoE externalizes this template: each expert is an independent ALU, the gating network acts as the control unit, and the KV Cache serves as memory. Deeper Isomorphism: Von Neumann was the first to combine computation theory with game theory, demonstrating that the same mathematical framework could describe both computation and strategy. The game-theoretic thinking Liang Wenfeng practiced in quantitative trading (treating the market as an adversarial game) transformed at DeepSeek into “the game of compute allocation during training”—allocating the most expensive computing resources to the problems that need them most. DeepSeek Research • 对抗式深度推演 / Adversarial Deep Deduction
6. Liu Cixin Observable Evidence: DeepSeek is not a standard profit-driven commercial enterprise—it is a research institute betting on civilization-level outcomes. Commercial logic dictated that DeepSeek should follow other Chinese AI companies into B2B applications and private deployments. Instead, DeepSeek chose the most foundational, cutting-edge, and short-term least lucrative endeavor: AGI research. Deduction Logic: The influence of Liu Cixin’s The Three-Body Problem on the worldview of China’s technological elite cannot be ignored. The core axiom in Three-Body—“Survival is the primary need of a civilization; civilization continuously grows and expands”—manifests in Liang Wenfeng’s technical choices: the Chinese-speaking world needs its own AGI explorer independent of OpenAI’s path. The concept of a “technological explosion” in Three-Body is DeepSeek’s wager. If OpenAI’s closed-source model is the Sophon, DeepSeek’s open-source path is nanomaterials. Adversarial Point: The influence of Three-Body on Chinese tech CEOs has become a narrative cliché and a form of identity performance. It is as natural as breathing, and therefore cannot uniquely explain Liang Wenfeng.
7. James Simons Observable Evidence: Liang Wenfeng is the only person in China who successfully traversed the path from quantitative trading to AI research. Deduction Logic: Simons and Renaissance Technologies were the first in the world to apply pure mathematical methods to financial markets and achieve sustained alpha. Liang Wenfeng’s High-Flyer Quant is China’s closest equivalent to Renaissance (fully automated trading, purely quantitative, zero fundamental analysts). Simons proved that a mathematician could earn billions via mathematical models and use that capital for foundational scientific research. Liang Wenfeng’s structure is strikingly similar: earning money through quant trading, buying GPUs, and conducting AI research. Adversarial Point: Yet, this may be the “influence” Liang Wenfeng is least willing to acknowledge. The Chinese financial community has a complicated relationship with quant trading, and in interviews, Liang Wenfeng describes himself as a “Zhejiang University Computer Science graduate” rather than “CEO of High- Flyer Quant.” DeepSeek Research • 对抗式深度推演 / Adversarial Deep Deduction
8. Noam Chomsky Observable Evidence: DeepSeek’s attempts at reasoning interpretability—displaying its chain of thought (CoT) rather than presenting a black-box output. Deduction Logic: A distinction must be made here—DeepSeek is not a follower of the Chomskyan school. On the contrary, DeepSeek proves Chomsky wrong—no preset grammatical rules are needed; given enough parameters and data, language emerges entirely from statistical distributions. When a scientist spends his life proving what you do is impossible, and he is the most powerful figure in that discipline, he effectively shapes your path through opposition. Chomsky’s failed premise underscores the significance of DeepSeek’s success.
9. Wang Yangming Observable Evidence: An extreme version of “the unity of knowing and doing” (知行合一) within DeepSeek’s organizational culture. Deduction Logic: Wang Yangming said, “To know and not to do is not to know.” DeepSeek is the only AI company in China where the core founder personally writes code and actively participates in discussions on model training strategies. Liang Wenfeng’s style is “cultivation through action,” managing the team while remaining deeply involved in engineering. Furthermore, Wang Yangming’s “Mind is Reason” finds its AI-era engineering expression here: relying not on external complex data or human annotations, but entirely on the model’s internal knowledge accumulation and self-evolution.
10. Friedrich Hayek Observable Evidence: The distributed representation of knowledge in DeepSeek’s MoE architecture and a spontaneous order free from central planning. Deduction Logic: One of Hayek’s most crucial ideas is “dispersed knowledge”—no central planner can possess all information; knowledge is distributed among countless individuals, and the market mechanism (price signals) is the best way to coordinate this dispersed knowledge. DeepSeek’s MoE architecture is isomorphic to this: no single central neural network processes all problems; instead, multiple “expert networks” specialize in different fields, and the gating network (acting like a price mechanism) dynamically decides which expert receives the problem. Its pure RL reasoning is a perfect engineering realization of Hayek’s “spontaneous order” inside a neural network. DeepSeek Research • 对抗式深度推演 / Adversarial Deep Deduction Step 4: Physics First, Narrative Second Zhang Yiming’s physical boundary was time (7 years of accumulation, 5,110 hours of reading). Liang Wenfeng’s physical boundary is compute. Liang Wenfeng did not form his intellectual throughput through reading—he formed it through the iterative feedback of computational resources. From 2015 to 2023, High-Flyer Quant processed petabytes of market data annually, training hundreds of models. Every model generated trading signals, every signal generated profit or loss, and those profits or losses fed back into model iteration. This loop ran for 8 years, iterating every single trading day, accumulating around 2,000 cycles. This is not wisdom gained from books—this is a cognitive filter forged by trillions of real-world feedback signals. Therefore, his “Ten Influential Figures” are essentially ten cognitive tools—not books on a shelf, but the origins of the mental models he used for quantitative trading: Shannon represents signal processing (filtering entropy from market noise); Von Neumann represents game theory (modeling counterparty behavior); Sutton represents sequential decision making (Markov Decision Processes in trade execution). He read these thinkers initially to build a more profitable trading strategy. Only later did he repurpose these cognitive tools for AGI. Step 5: Action Items for Next Week This deduction features a verifiable metric—DeepSeek has published several high-quality papers, allowing for a quantitative citation analysis: Action Point: Analyze the frequency of historical thinkers versus contemporary works in the references of DeepSeek V2, LLaMA, V3, and R1 papers. If Turing, Shannon, Hinton, and Sutton are high-frequency footnotes in DeepSeek’s papers, it indicates that the aforementioned intellectual pedigree completely holds true at the top engineering level. If the references consist primarily of contemporary peer works (Gemini, LLaMA, GPT-4) and historical figures rarely appear, it indicates that DeepSeek’s technical direction stems from an extraordinary capacity to synthesize the contemporary frontier and engineering intuition, rather than a regressive reading of historical giants. Next Week’s Action Item: Extract and analyze the frequency of cited authors and classical theories across DeepSeek’s three core papers to verify whether its underlying philosophy is a “deduction of historical thought” or the “extreme mutation of contemporary engineering.” Conclusion (The Uncomfortable Version) This question contains a trap absent in the Zhang Yiming version: Zhang Yiming’s influencers were found externally —Drucker, Shannon, Jobs—he understood what others had done and then executed it better in the Chinese market. He is a magnificent implementer. Liang Wenfeng, however, is a species forced to evolve inside a closed compute cluster. His intellectual figures are not his guides, but the tools he uses ex-post to explain his own mutational path. DeepSeek Research • 对抗式深度推演 / Adversarial Deep Deduction