Fetching from the wire…
Policy2026-09-26 · source-backed
Jeff Stein's Washington Sun piece (September 24) says the money comes from classified parts of the national security budget with compute as the largest cost, far above public proposals like the roughly $20M a year in the AI Security and Innovation Act. The Pentagon declined to discuss "resource allocation for its AI tools." It reopens the question of whether labs should pay for independent audits rather than taxpayers.
Each link below shares sources, entities, or timing with this story.
AA26-251A accuses DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI of extracting billions of tokens across millions of requests from Claude, GPT, Gemini and Grok since 2024, listing which US model each firm targeted. It separates legitimate distillation research from...
Moonshot AI released Kimi K3, a sparse mixture-of-experts activating 16 of 896 experts per token. That's about 1.8% of the pool live at any moment, with a 1M-token context window and native vision. Two new architectural pieces show up: Kimi Delta Attention and Attention Residu...
The Standard, citing The Information, reports the Cyberspace Administration interviewed executives at both labs over live customer conversations, some containing sensitive information, being routed through Anthropic's models to generate training data, potentially violating Chi...
Harness-of-Harness organizes existing harness executions into iterative loops, scoping work into small verifiable increments and separating implementation-time testing from independent evaluation (arXiv 2609.01481). Across GameCraft-Bench, FrontierSWE and ProgramBench with Cod...
25,000 fake accounts. 28.8 million Claude conversations. Six weeks. And the thing they were harvesting wasn't trivia, it was software engineering and agentic reasoning. In a June 24 letter to US senators and the White House, Anthropic alleged that operators tied to Alibaba's Q...
GLM-5.1 from Zhipu AI scored 58.4% on SWE-bench Pro. GPT-5.4 scored 57.7%. Claude Opus 4.6 scored 57.3%. That's the first time an open-weight model has ever topped a major coding benchmark against the best proprietary models. The specs matter. GLM-5.1 is a 754B-parameter mixtu...
MindPattern daily
One email a day at 7 AM. Sources and a take on every story. Unsubscribe anytime.