DeepSeek Quiz: Can You Crack the AI Disruptor?

DeepSeek was created by a technology company operating in the world's most populous country. The organisation gained global attention for producing models that rivalled leading Western alternatives while reportedly spending a fraction of the usual development budget. Its rise prompted discussions about the global distribution of AI capabilities.

DeepSeek-V2 introduced a change to how the model handles attention during text generation. Traditional attention mechanisms store a lot of data per token, which is expensive. DeepSeek's approach compresses that storage significantly, allowing the model to serve more requests at lower cost without a major drop in output quality.

DeepSeek-R1 was designed to tackle complex reasoning tasks by thinking through problems before producing an answer. This approach mirrors a specific OpenAI model that also emphasises extended internal reasoning over speed. Both models trade faster response times for improved accuracy on challenging logical and scientific problems.

DeepSeek-R1 is trained using a reinforcement learning method that differs from the more widely known approaches used by other labs. Rather than relying on a separate reward model, this technique uses a group of sampled outputs to estimate relative quality. It is considered more efficient and stable during training.

DeepSeek did not emerge from a conventional technology company. It was established and is primarily backed by a quantitative trading firm known for using mathematics and algorithms to manage investments. This unusual origin set it apart from most AI labs, which are typically funded by big technology corporations or venture capital.