What is DeepSeek?

DeepSeek is a Chinese AI model developer known for producing language models and advancing reinforcement learning techniques. The material tracks its cost efficiency, engineering innovations, and competitive performance against US models.

• Apr 2025 - Brad Gerstner said that if the US bans DeepSeek, more people would work with Huawei chips and DeepSeek to optimize it, and other countries would use it.
• May 2025 - Brad Gerstner noted that DeepSeek had only operated on US models.
• Feb 2026 - Sebastian Raschka speculated that DeepSeek might release version 4 around Chinese New Year, with a possible big change.
• Apr 2026 - Dylan Patel stated that DeepSeek’s cost on GPT-4 was 1/600 the cost.
• Apr 2026 - Reiner Pope said the DeepSeek paper reports they do a lot of expert parallelism.
• May 2026 - Kyle Corbitt explained that GRPO took off because DeepSeek did engineering work around scaling it and released an actual artifact model.
• Jun 2026 - Carina Hong reported that the best DeepSeek LM scored 103 out of 120 on an exam, while the best human scored 110.