“We do not think the scaling law is the only path. Algorithmic innovation matters more than throwing more compute at the problem.”
“Our goal is to explore the nature of AGI. Commercial applications are secondary.”
“The chip restrictions have forced us to be more creative. Constraints breed innovation.”
“We want to make AI research accessible. Open-sourcing our models is fundamental to our mission.”
“The cost of intelligence should approach zero. That is the inevitable trajectory.”
Liang Wenfeng is among the most consequential yet least publicly visible AI leaders globally. As the founder of DeepSeek, he built a Chinese AI lab that in January 2025 released models rivaling GPT-4 class performance at dramatically lower cost, sending shockwaves through the AI industry and calling into question the American assumption that chip restrictions would keep Chinese AI permanently behind.
Liang’s public statements are rare compared to his American counterparts, but they reveal a distinctive philosophy. Where Western labs emphasize scaling — bigger models, more compute, larger datasets — Liang consistently points to algorithmic efficiency as the more important lever. DeepSeek’s mixture-of-experts architecture and training optimizations delivered frontier performance using a fraction of the compute that American labs employed.
His comment about chip restrictions breeding innovation proved prescient. Rather than being permanently hobbled by US export controls on advanced GPUs, DeepSeek found ways to achieve competitive results with available hardware, demonstrating that the relationship between compute and capability is more complex than simple scaling laws suggest.
Liang’s background in quantitative finance (he previously ran a successful hedge fund) gives him a distinct perspective on AI development economics. He approaches model training as an optimization problem where efficiency gains compound, rather than a brute-force competition where the largest budget always wins.
His commitment to open-sourcing DeepSeek’s models distinguishes the lab from other Chinese AI companies and aligns it more with Mistral’s philosophy than with the closed approaches of OpenAI or Anthropic. This openness, combined with frontier capabilities, has made DeepSeek one of the most influential AI labs in the world despite operating under significant hardware constraints.