Overview
March opened with the consequences of two unresolved deployment questions. Anthropic's military-use limits had become a supply-continuity dispute, while DeepSeek's reported Huawei work tied model competition to accelerator independence. A clinical vision-language study supplied a narrower technical answer: formal checks can catch contradictions that aggregate accuracy conceals.
Developments
1. Anthropic's Pentagon talks collapse over two prohibited uses
Anthropic's negotiations with the Pentagon broke down over restrictions on mass domestic surveillance and fully autonomous weapons, according to reporting published March 1. The government argued that a defense supplier could not constrain lawful military activity; Anthropic refused to remove the two limits from Claude's terms.
The disagreement turns acceptable-use language into a commercial boundary with immediate distribution consequences. Anthropic preserved control over two applications it considers unacceptable, but federal agencies and contractors now face pressure to remove direct and embedded Claude dependencies before the legal status of the designation is settled.
Sources: New York Times · TechCrunch
2. DeepSeek's reported flagship couples model design to Huawei hardware
The Financial Times reported that DeepSeek was preparing a flagship language model and exploring greater use of Huawei systems to reduce reliance on Nvidia. The report supplied no final model card, weights, API terms, launch date, or measured production configuration.
The hardware pairing is the strategic claim. A model optimized around Huawei accelerators could reduce exposure to U.S. export controls and Nvidia's software ecosystem, but the commercial effect remains unknowable until token throughput, memory pressure, interconnect performance, and serving cost are measured together.
Sources: Financial Times
3. A clinical verifier converts radiology claims into logic checks
A February 27 preprint formalizes findings from generated radiology reports as propositional evidence, then uses a Z3 solver and a clinical knowledge base to classify diagnostic claims as entailed, unsupported, or omitted. The authors evaluated seven vision-language models across five chest X-ray benchmarks.
This design targets internal consistency after generation, where lexical similarity scores can reward a fluent report containing a contradictory diagnosis. Its guarantee is bounded by the extracted propositions and knowledge base: an incorrect visual finding can remain logically consistent, leaving perception error outside the solver's protection.
Sources: Clinical VLM preprint