The Model That Shocked the World, Then Disappeared: The DeepSeek Story
The message hit the group chats of AI researchers on a Thursday evening in January 2025. Someone had uploaded the weights of a new model to Hugging Face. The model was called DeepSeek-R1. It was from a Chinese lab most people in the Western AI community had never heard of. And it was: depending on the benchmark you looked at: roughly competitive with OpenAI o1, the best reasoning model in the world at the time.
Within forty-eight hours, the response from the AI community moved through the predictable stages: initial skepticism, frantic benchmark replication, dawning recognition that the benchmarks were holding up, and then something closer to genuine alarm. Not because a Chinese lab had built a capable model: that had happened before. But because of how they had built it, and what the cost implied.
DeepSeek claimed that R1 had been trained for approximately six million dollars. OpenAI had never disclosed the training cost of o1, but informed estimates placed it in the range of one hundred million dollars or more. If DeepSeek number was accurate: and independent researchers who analyzed the training methodology found no obvious reason to doubt it: something significant had happened. The assumption that frontier AI required frontier compute budgets had been falsified.
The downstream implications cascaded quickly. Nvidia stock fell seventeen percent in a single day: one of the largest single-day market cap losses in stock market history: as investors processed what cheaper training might mean for demand for expensive AI chips. The export controls that the US government had placed on advanced semiconductors, intended to limit China AI capabilities by restricting access to Nvidia most powerful chips, looked considerably less effective if capable models could be trained without those chips. Suddenly the geopolitical calculus of the AI race required updating.
The technical paper DeepSeek released alongside the model was unusually detailed, and researchers found it credible. The efficiency gains came from several sources that were each individually known but had not previously been combined at this scale.
Mixture of Experts architecture: where a model routes each input to a subset of specialized subnetworks rather than activating the entire model for every query: allowed DeepSeek to achieve high effective capacity while keeping the compute cost of each forward pass manageable. Aggressive quantization reduced the memory footprint of model weights without significant performance degradation. And a training approach that emphasized reinforcement learning from verifiable outcomes: rather than purely supervised imitation of human-generated data: proved remarkably effective for the reasoning tasks where R1 excelled.
Please enable JavaScript to read the full article.