Seek Documentary on Liang Wenfeng, R1 and What's Next", In January 2025, a Chinese AI model dropped that made the world's top tech leaders suddenly nervous. DeepSeek R1 wasn't supposed to happen—Western labs were supposedly years ahead, with language models growing ever more expensive as they got smarter. But this model could think before it spoke, cost almost nothing to run, and was open for anyone to download. Even OpenAI admitted their lead was narrowing. This is the story of how one secretive billionaire upended the AI world—and what comes next.
The Shock Heard Round the World
On January 20, 2025, DeepSeek R1 arrived like a thunderbolt. The model visibly thought before it spoke—a breakthrough that made Western labs scramble. It was unbelievably cheap, competitive with the best American models, and available to anyone to download. OpenAI's own admissions came quickly: in March, they acknowledged that DeepSeek shows our lead is not wide and is narrowing.
The timing mattered. Just weeks before, Google Gemini and ChatGPT image generation had dominated headlines. But DeepSeek was about to change the conversation entirely.
Who Is Liang Wenfeng?
The founder behind DeepSeek is a billionaire who now hides from adoring fans in his own hometown, according to a friend. He has fled his home province with family to escape further attention. That same man built Highflyer, a hedge fund that used AI to uncover patterns in financial markets—patterns no human could detect alone.
Liang graduated into a world collapsing around him during the 2008 financial crisis. After getting a master's in information engineering in 2010, he founded several companies between 2013 and 2016, culminating in Highflyer in February 2016. The hedge fund attracted $9.4 billion in assets by end of 2021, delivering returns that were 20 to 50 percentage points more than stock market benchmarks.
The fund had a troublesome quirk: its AI was too much of a risk-taker. It would double down on bets when it felt confident. After a sharp drawdown, Highflyer limited who could invest with them—but they learned their lesson and are still successful today.
Liang didn't give up on AI. He was rich now and could afford an effort dedicated to decoding not just financial systems, but the nature of general intelligence itself.
The Birth of DeepSeek
DeepSeek first formed as a research body in April 2023. By then, OpenAI had already released GPT-4, showing sparks of artificial general intelligence. Western AI labs were supposedly far ahead.
But Liang was clear about his intentions: people think there's some hidden business logic behind DeepSeek, but it's mainly driven by curiosity.
DeepSeek's recruitment strategy differed from American companies. It prioritized capability over credentials. Core technical roles went to recent graduates or those one to two years out—no KPIs, no quotas, no release schedules to chase. The goal was exploration rather than monetization.
Their first models in November 2023 weren't stunning—DeepSeek V1 drew heavily from Meta's Llama 2. But there were signs of long-termism. DeepSeek excluded multiple-choice questions from their training data so models wouldn't overperform on formal tests but underwhelm in practice. They explicitly stated: overfitting to benchmarks would not contribute to achieving true intelligence.
The Technical Breakthrough
By early 2024, the team pioneered a novel approach to getting more intelligence for less cost. Traditional models like Llama 2 used their entire set of weights—often tens or hundreds of billions—to compute a response. But DeepSeek introduced what they called the mixture of experts approach.
This involved using specialized subsets of weights depending on the user input, tapping into different expert networks within the model for each request. The innovation: certain expert subnetworks would always be activated in any response—these were the generalists. The remaining experts could specialize deeply without affecting everyday performance.
Western Labs Respond
OpenAI's leadership had argued that research secrets and money were becoming moats. In June 2023, Sam Altman replied to whether a team with $10 million could compete with OpenAI: it's totally hopeless to compete with us on training foundation models—you shouldn't try.
But Liang tried anyway. And DeepSeek R1 proved the gap between open and private models might be closing.
Now DeepSeek is preparing another shock—DeepSeek R2 expected in April or May—with Western labs watching closely.
The Stakes Are Higher Than One Company
Liang has said if artificial general intelligence is 10, 5, or even 2 years away, this story matters far more than one man, one lab, or even one nation. The question isn't just about competition—it's about what comes next.
Counterarguments
Critics might note that DeepSeek's open accessibility raises genuine concerns. OpenAI has argued DeepSeek could be compelled by the Chinese government to manipulate models to cause harm. They also worry that because DeepSeek is state-subsidized, state-controlled, and freely available, it could cost users their privacy and security.
A counterargument worth considering: the narrative around DeepSeek has become a whale caught in a net of contradicting stories. Some portray them as dangerous; others celebrate their openness. The truth likely lies somewhere between these extremes.
Bottom Line
This story is bigger than one company. What Liang Wenfeng built shows that AI development isn't just about who has the most resources—it also requires fundamental technical innovation and accessibility. DeepSeek R1 proved that a well-funded but smaller team can compete with the hyperscalers if they find better approaches. The biggest vulnerability: we still don't fully understand what DeepSeek's open weights mean for global AI safety, and that's a conversation just beginning.