Liang Wenfeng: A Trailblazer Bridging Quantitative Investment and General AI

The Dual DNA of Tech Entrepreneurship: Financial and AI Convergence

Liang Wenfeng’s career demonstrates a rare synergy between quantitative finance and AI innovation. After earning his master’s degree in Electronic Information Engineering from Zhejiang University in 2008, he pioneered mathematical modeling applications in high-frequency trading rather than pursuing conventional technical roles. This interdisciplinary approach laid groundwork for his 2013 founding of Hangzhou Jacobi Investment Management, which systematized quantitative strategy development.

Key milestones during his tenure at High-Flyer Quant include:

1. Infrastructure Revolution: Established an AI supercomputing center in 2018 with thousands of GPUs, breaking traditional quantitative firms’ computational barriers through a $150 million investment
2. Strategy Pipeline: Implemented an end-to-end workflow covering data cleansing, feature engineering, model training, and backtesting validation, enabling weekly strategy iterations
3. Risk Management Innovation: Developed dynamic hedging algorithms that limited maximum drawdown to 1/3 of industry average during 2020’s market volatility

These advancements culminated in High-Flyer becoming Asia’s first quantitative fund exceeding $15 billion AUM in 2021. Its hybrid strategy system combining traditional multi-factor models with deep learning achieved stable 15%+ annualized alpha in low-frequency trading.

General AI Breakthrough: From Specialized Models to Cognitive Revolution

Liang’s 2023 founding of DeepSeek marked a strategic pivot to general AI development. The team employed three-phase innovation to overcome large language model training challenges:

1. Mixture-of-Experts Architecture Innovation

DeepSeek-V3 surpassed Transformer’s computational limits through:

• Dynamic routing mechanism allocating 8-16 expert modules per token via attention-weighted assignment
• Sparse activation optimization reducing computation by 30% through Top-k expert selection
• Progressive training methodology expanding model scale via knowledge distillation

This architecture achieved 87.5% accuracy on MMLU benchmarks - comparable to GPT-4o’s 88.2% - at 1/5 the training cost.

2. Reinforcement Learning for Reasoning Optimization

DeepSeek-R1 introduced closed-loop reasoning systems:

# Reinforcement learning training framework
class ReasoningAgent:
    def __init__(self, model):
        self.model = model
        self.reward_model = build_reward_network()
    def generate_chain_of_thought(self, prompt):
        thoughts = []
        for _ in range(max_steps):
            output = self.model.generate(prompt + "\n".join(thoughts))
            thoughts.append(output)
            if self._is_terminal(output):
                break
        return thoughts
    def optimize_path(self, thoughts):
        rewards = [self.reward_model.score(t) for t in thoughts]
        return thoughts[np.argmax(rewards)]

The 20B-parameter reward model enabled automatic reasoning path evaluation, boosting GSM8K math accuracy from 62% to 81% through chain-of-thought optimization.

3. Open-Source Ecosystem Strategy

Contrary to proprietary approaches, DeepSeek adopted full model transparency:

• Released complete model family (7B-67B parameters)
• Open-sourced training framework including data pipelines and distributed training tools
• Established community contribution system receiving 1,200+ global developer inputs

This approach generated 500,000+ downloads and 120,000 GitHub stars within 30 days of DeepSeek-V3’s release.

China’s Unique Path to AI Innovation

Liang’s work exemplifies three key characteristics of Chinese AI development:

1. Cost-Efficiency Revolution

• Model compression techniques achieving 4x inference speed through FP32-to-INT4 quantization
• Synthetic data generation covering 90% of math training requirements
• Custom inference chips developed with cloud partners tripling single-card performance

2. Cognitive Capability Expansion

Beyond conventional NLP tasks:

• Multimodal understanding achieving 89.3% VQA accuracy through joint image-text encoding
• Tool integration enabling autonomous calculator/search engine operation
• Contextual memory extension supporting 100K-token continuous dialogues via vector databases

3. Industrial Application Development

Three primary implementation directions have emerged:

• Smart investment research: Quantitative strategy generation system reducing development cycles from months to weeks
• Scientific research assistance: Chemical molecule generation models cutting drug discovery costs by 70%
• Education: Adaptive learning systems improving efficiency by 40% through dynamic difficulty adjustment

Strategic Vision for AI Leadership

At the 2024 Tech Summit, Liang outlined three critical AI development trends:

1. Computational Democratization: Model compression enabling trillion-parameter models to run on consumer-grade GPUs
2. Capability Equalization: Open-source ecosystems eliminating technology gaps within three years
3. Vertical Application Dominance: Maximum value creation through general capability integration with industry expertise

These insights materialized in DeepSeek’s establishment of 30+ industry labs focusing on industrial defect detection and protein structure prediction. Their industrial vision model achieved 99.97% detection accuracy in automotive manufacturing, surpassing human expert performance.

Liang’s trajectory from quantitative finance to general AI demonstrates China’s innovation paradigm: overcoming resource constraints through engineering optimization, building technological moats via open collaboration, and achieving leadership through paradigm shifts. This approach not only provides new development models for the AI industry but proves that cultural choices often prove more decisive than technical selections in revolutionary technology fields. While the industry debates open-source versus proprietary models, DeepSeek’s practice confirms that true technological fortresses are built on collaborative foundations.