DeepSeek review

DeepSeek has a good reputation for offering powerful models and for publishing many of them openly, and it's generally considered a serious alternative to the American AIs. Negative opinions tend to focus on privacy, because it's a Chinese company and what you send to its service ends up on its servers, and on the fact that it avoids or softens some sensitive political topics.

What's confusing when reading opinions is that different things get mixed together: the chat app, the API, the open models you can run on your own machine and, now, its DeepSeek Harness. Distrusting where your data goes in the app isn't the same as assessing a downloaded model that runs on your computer offline. Another trap: many opinions come from a quick test or refer to old versions. The sensible thing is to form your own with real tasks and without putting in sensitive information until you're clear about what it does with it.

I try to base my opinion on use rather than headlines: DeepSeek is one of the AIs I work with, along with Claude, OpenAI, GLM, Kimi and MiniMax. The most detailed one is in "Ten days working with DeepSeek Harness", where I bring it into my real work, compare it with what I already use and try to give an honest verdict on what it's good for and what it isn't.