From Quantitative Trading to General AI: A Tech Visionary's Paradigm-Shifting Journey

The Unconventional Rise of a Tech Leader

Born in Zhanjiang, Guangdong, Liang Wenfeng began his AI journey at Zhejiang University’s AI Lab during the 2008 global financial crisis. While pursuing his master’s degree, he led a team exploring machine learning applications in quantitative trading, developing the conviction that “true technological breakthroughs emerge at interdisciplinary intersections.”

In 2015, Liang founded a quantitative hedge fund that achieved 100 billion RMB in assets under management within six years—far exceeding industry averages. The proprietary trading system, powered by machine learning algorithms, delivered 38% annualized alpha returns from 2018-2020 through automated high-frequency strategy optimization. At the peak of financial success, however, Liang made a controversial decision: redirecting core R&D resources to the nascent field of general artificial intelligence.

Strategic Infrastructure Development

Counter-Cyclical Investment Approach

During the 2019 AI winter, when most companies cut R&D budgets, Liang’s team invested 60% of annual trading profits into GPU cluster construction. This decision rested on three forecasts:

• GPU costs would decrease 40% every 18 months
• AI model parameters would grow exponentially
• Computational power would become a strategic asset

Phased Implementation Roadmap

Firefly Phase I (2019)

• Deployed 1,024 professional GPUs
• Built distributed training framework
• Solved multi-node communication bottlenecks

Firefly Phase II (2021)

• Scaled to 5,120 GPUs
• Implemented liquid cooling technology
• Reduced PUE to below 1.1

Firefly Phase III (2023)

• Achieved 10,000+ GPU cluster
• Adopted 3D packaging for density
• Supported 100B+ parameter training

This incremental approach maintained technological leadership while avoiding capital overcommitment. Industry surveys show fewer than five Chinese companies operate at this scale, with non-internet entities being particularly rare.

Technical Breakthroughs in General AI

Architectural Innovation

The team developed a Mixture-of-Experts (MoE) architecture featuring dynamic routing that assigns subtasks to specialized subnetworks. This design achieves 3-5x effective parameter utilization while maintaining total model size. Benchmark tests showed 42% efficiency gains over traditional Transformers in code generation tasks.

Data Pipeline Optimization

The multimodal data processing pipeline:

class DataPipeline:
    def __init__(self):
        self.cleaners = [
            TextNormalizer(),       # Text standardization
            ImageAugmentor(),       # Image enhancement
            AudioFilter(),          # Audio denoising
            CrossModalAligner()     # Multimodal alignment
        ]
    def process(self, raw_data):
        for cleaner in self.cleaners:
            raw_data = cleaner.transform(raw_data)
        return raw_data

This system processes 2PB of data daily while maintaining data quality through automated monitoring, keeping invalid data below 0.3%.

Training Framework Enhancements

To address communication bottlenecks in large clusters, the team developed gradient compression:

\text{CompressedGradient} = \text{Sign}(\nabla W) \odot \text{TopK}(|\nabla W|, k) \odot \text{ScaleFactor}

This algorithm reduces gradient transmission by 87% without compromising convergence speed, cutting communication overhead from 42% to 15% in 1,024-node training.

Foundational Decision-Making Principles

Long-Term Commitment

Liang views general AI development as “mountain climbing with a decade horizon,” enabling sustained investment in fundamental research despite 18 months without breakthroughs.

Cross-Disciplinary Knowledge Transfer

Financial modeling expertise informed AI development:

• Adapted backtesting frameworks for model validation
• Optimized hyperparameter search with risk control models
• Improved inference services using low-latency trading technologies

Open Ecosystem Development

The team open-sourced communication modules, attracting 300+ global contributors. This strategy accelerated development while establishing industry standards—the framework now serves as default communication component for major cloud AI platforms.

Key Strategic Insights

1. Compute Readiness: Maintain 12-18 month capacity planning with 20% redundancy for demand spikes
2. Engineering Priority: In general AI R&D, engineering optimizations often yield greater performance gains than algorithmic changes
3. Strategic Buffer: Maintain separation between core operations and exploratory research to protect long-term vision

The team now explores quantum-classical hybrid computing approaches. This relentless innovation spirit may hold the key to China’s AI leapfrog development, demonstrating that true breakthroughs emerge from non-mainstream paths requiring both strategic patience and engineering excellence.