From Quantitative Trading to AGI: The Innovation Journey of a Tech Leader

I. Technical Foundations and Early Breakthroughs

Liang Wenfeng’s technical journey began with academic-industry integration. After earning B.S. (2007) and M.S. (2010) degrees in Electronic Information Engineering from Zhejiang University, he developed core competencies in signal processing and algorithm optimization. These foundations proved critical for his subsequent breakthroughs in quantitative finance and AI.

In 2013, Liang co-founded Hangzhou Jacob Investment Management with ZJU alumnus Xu Jin, focusing on quantitative trading. This domain requires expertise in:

• High-frequency data processing
• Algorithmic trading model construction
• Real-time risk management

The team’s 2015 establishment of High-Flyer Quant marked a turning point. Through自主研发 (proprietary development) of:

1. Low-latency trading architecture: Achieved microsecond-level order processing via FPGA acceleration and kernel-level network optimization
2. Distributed computing framework: Enabled real-time analysis of PB-scale market data through a 1,024-node cluster
3. Multi-factor model optimization: Improved strategy adaptability using dynamic machine learning-based factor weighting

By 2021, High-Flyer became one of China’s first quantitative firms managing over RMB 100 billion in assets.

II. Technical Transformation in AGI Development

In July 2023, Liang founded DeepExplore AI Research, shifting focus to large language models (LLMs) and artificial general intelligence (AGI). The team addressed three key challenges:

1. Hybrid Expert Architecture Innovation

The DeepSeek-V3 model (December 2024) introduced a Mixture of Experts (MoE) architecture featuring:

• Modular specialization: 64 expert subnetworks focusing on distinct knowledge domains

• Dynamic resource allocation: Gating mechanism with softmax-based weight distribution

  # Simplified MoE gating mechanism implementation
  class MoEGating(nn.Module):
    def __init__(self, num_experts, input_dim):
        super().__init__()
        self.gate = nn.Linear(input_dim, num_experts)
    def forward(self, x):
        logits = self.gate(x)
        weights = F.softmax(logits, dim=-1)  # Normalized expert weights
        return weights

  This architecture achieved:

• 92% of GPT-4o’s performance on mathematical reasoning tasks
• 5x lower training costs compared to industry averages
• 1.6B parameter size after quantization compression

2. Reinforcement Learning-Driven Reasoning

The January 2025 release of DeepSeek-R1 implemented a three-stage reasoning optimization pipeline:

1. Supervised fine-tuning: Leveraged high-quality annotated datasets
2. Reward model training: Constructed neural network for reasoning quality evaluation

3. PPO optimization: Applied proximal policy optimization for strategy refinement

   # PPO training workflow pseudocode
   def ppo_train(model, reward_model, env):
    for epoch in range(max_epochs):
        trajectories = []
        for _ in range(batch_size):
            state = env.reset()
            while not done:
                action = model.sample_action(state)
                next_state, reward, done = env.step(action)
                trajectories.append((state, action, reward))
                state = next_state
        advantages = compute_advantages(trajectories, reward_model)
        model.update(trajectories, advantages)  # Policy gradient update

   Benchmark results:

• 94.6% accuracy on GSM8K math problems (vs. OpenAI o1’s 93.2%)
• 3x faster inference speed
• 40% reduced decision latency in robotic control applications

III. Industry Impact of Technical Breakthroughs

DeepExplore’s innovations are reshaping AI development paradigms through:

1. Computational efficiency revolution: 80% cost reduction in model training
2. Open-source ecosystem: MIT-licensed DeepSeek-V3 has spawned 120+ vertical applications
3. Reasoning standardization: PPO framework adopted by 23 research institutions worldwide

IV. Leadership Methodology Insights

Liang’s approach demonstrates three key principles:

1. Cross-domain knowledge transfer: Applied quantitative trading’s high-frequency processing techniques to AI training optimization
2. Engineering-first mindset: Maintained deployment efficiency through model quantization and architecture search
3. Open collaboration model: Established joint labs with 6 universities, resulting in 42 co-authored research papers

Current research explores multimodal AGI integration, with preliminary results showing 40% reduced latency in robotic decision-making when processing unified text-image-sensor data streams. This progress represents a critical step toward physical world interaction capabilities, offering developers new directions for technical exploration.