DeepSeek Adds Vision While US Stalls on Blacklist
DeepSeek shipped vision capabilities to its chat product, and the US government still hasn't added it to the entity list despite months of pressure. More than 100 firms were deemed security risks in the same review cycle, but DeepSeek survived the cut. Separately, Anthropic disclosed it found evidence of Chinese AI labs, including DeepSeek, running campaigns to extract capabilities from Claude by querying it at scale.
The pattern is a split between government rhetoric and government action. The US talks about Chinese AI as a national security threat, but DeepSeek remains accessible, Qwen runs locally on developer machines, and the actual blacklisting keeps getting deferred. Meanwhile developers in the thread were openly running these models on GrapheneOS and Linux, treating the policy uncertainty as irrelevant to their daily use.
The Anthropic disclosure is the more technically interesting thread. If Chinese labs are systematically querying frontier models to distill capabilities, that changes the calculus for any company offering API access. Your public API is also a training signal for your competitors.
So what?
If you're building on top of open-weight Chinese models like Qwen or DeepSeek, the regulatory risk is real but currently deferred. Plan for the possibility that API access or model distribution gets restricted, and make sure your stack can swap models without a rewrite. If you're running a public AI API, the Anthropic disclosure is a warning: sophisticated actors are querying you systematically, so rate limiting and abuse detection are not optional.