Everybody's talking about DeepSeek

DeepSeek is a Chinese tech company that develops LLMs. Chatter about it has been fierce and frequent. I put it on the backburner because of my initial concern about the security of a Chinese-based AI—but too many respectable developers have been praising it as a game-changer to ignore the buzz.

After researching and testing DeepSeek, a few notes:

• It appears to have the intelligence of ChatGPT with the more natural language of Claude, or even better. It’s ability to reason and speak like a human is unmatched.

• It is open source, which means the code is visible to view, use, and check its security. In contrast, OpenAI (ChatGPT) and the other big models are closed source.

• DeepSeek costs software companies about 4% of what it costs to integrate and run OpenAI or Claude, which means that a lot more tech is about to have AI capabilities.

• DeepSeek learns differently from and more efficiently than ChatGPT and the other popular LLMs. Others require Supervised Fine-Tuning (SFT), which is when human trainers are needed. DeepSeek uses Reinforcement Learning (RL) and corrects against itself or outside data.

A great introduction to this gamechanger is DeepSeek R1 - o1 Performance, Completely Open-Source from Matthew Berman.