"OpenAI is Not God” - The DeepSeek Documentary on Liang Wenfeng, R1 and What's Next

DeepSeek R1 detonated a long-simmering AI power struggle by delivering “reasoning” that looks like it thinks before it answers—at a price and openness that made Western labs scramble to explain why their lead wasn’t as secure as markets assumed. The shock wasn’t just that a Chinese model could compete with frontier systems; it was that R1’s chain-of-thought style output, plus the availability of the model and methods, turned a technical advantage into a public spectacle. Within weeks, the debate shifted from raw capability to cost, transparency, and whether the West’s “closed” approach is becoming a liability.

That outcome traces back to Liang Wenfeng’s unusual path into AI. Before DeepSeek, Liang built a hedge fund, Highflyer, using machine learning to find patterns in micro- and nanocond movements in financial markets—an approach that helped him amass billions by his mid-30s. The earlier AI work also left scars: his trading system and fund became risk-tolerant and overextended, prompting public damage control and tighter investment limits. DeepSeek, launched as a research body in April 2023, grew out of that same drive—except now the target was general intelligence rather than market prediction.

DeepSeek’s technical momentum came from efficiency-first design rather than brute-force scaling alone. The transcript highlights a sequence of innovations: a mixture-of-experts style activation strategy where certain “experts” are always engaged to keep general capability while specializing the rest; DeepSeek Math, a smaller model matching GPT-4–level math performance; and GRPO (group relative policy optimization), a reinforcement-learning method that avoids heavy critic models by generating answer groups in parallel and reinforcing the relative winners. Later, DeepSeek V2 is described as using multi-head latent attention to reduce the number of weights needed for comparable performance. The message is consistent: DeepSeek’s breakthroughs are framed as ways to extract more intelligence per unit of compute.

Compute constraints then become part of the plot. Liang’s lab reportedly secured 10,000 Nvidia A100 GPUs for Highflyer, but U.S. export controls tightened access to advanced chips. The transcript describes a broader “smuggling” narrative—chips moving through Singapore and Malaysia—suggesting that the AI race is increasingly about logistics as much as algorithms. That pressure helps explain why DeepSeek’s efficiency matters so much: without access to the largest compute stacks, better training methods and architectural tricks become the path to parity.

R1’s virality also sparked competing narratives. Western leaders questioned whether DeepSeek’s low training cost and open methods translate into sustainable progress, pointing to DeepSeek’s large infrastructure spending and the likelihood that costs will rise as models push toward AGI-like capabilities. There’s also a security and policy backlash: OpenAI and others argued that DeepSeek could be state-influenced and that freely available models create privacy and safety risks. Meanwhile, DeepSeek’s openness is portrayed as selective—R1 is MIT-licensed, but sensitive topics can still be constrained.

Beyond DeepSeek itself, the transcript argues the larger story is automation: reasoning is being operationalized through techniques like “think out loud” reinforcement, and the next frontier may involve infinite context and even replacing the transformer architecture. The central question becomes whether DeepSeek can keep compounding its efficiency and reasoning gains fast enough to reach AGI first—and whether it will share that path openly before the world catches up.