AI digest: Pentagon partnerships and production breakthroughs
OpenAI's Pentagon deal causes drama, Google speeds up LLM retrieval by 948x, and researchers show AI can bust anonymous accounts in minutes.
This week brought military AI partnerships into sharp focus, alongside some proper technical advances that’ll actually change how we build systems.
OpenAI’s Pentagon deal sparks exodus to Anthropic
OpenAI signed a contract with the Department of Defense for “all lawful use” of its models, causing enough controversy that Anthropic’s Claude shot to number one in the App Store. Even Sam Altman admitted the deal was “definitely rushed” and “the optics don’t look good.” The backlash seems genuine - people are voting with their downloads.
Google’s STATIC framework delivers 948x speedup for LLM retrieval
Google AI released STATIC, a sparse matrix framework that makes constrained decoding for generative retrieval nearly 1000 times faster. This matters because it makes LLM-based recommendation systems actually viable for industrial use. Instead of nearest-neighbour search, you can now have models generate item IDs directly whilst enforcing business rules like content freshness.
AI deanonymisation costs just a few dollars
Researchers from ETH Zurich and Anthropic demonstrated that commercially available AI models can link pseudonymous accounts to real identities in minutes for the cost of a coffee. They used standard techniques on publicly available data - no special access needed. This isn’t theoretical anymore; it’s a production capability anyone can deploy.
FireRed-OCR-2B tackles document hallucinations
FireRedTeam released FireRed-OCR-2B using GRPO to solve structural hallucinations in tables and LaTeX parsing. The model treats document parsing as an end-to-end vision problem rather than the usual detect-extract-reconstruct pipeline. Proper OCR that doesn’t invent formulas or mess up table structures could be genuinely useful for developers dealing with technical documents.