Researchers from UMD, Google, Meta, and other institutions use AutoTTS to let a coding agent independently discover control algorithms for AI reasoning. The algorithm it found cuts compute by about 70 percent compared to standard self-consistency while matching its accuracy. The whole search cost $40 and took 160 minutes.

ByteDance Seed shows that a 7B model can answer questions on long, image-heavy documents more reliably than much larger models, even when documents are four times longer than anything it saw during training. Instead of transcribing pages, the model learns by answering questions and finding the right passages on its own.

Anthropic will likely keep supplying AI models to the NSA despite being labeled a "supply chain risk." Intelligence agencies lack Nvidia's latest Grace Blackwell chips, and Anthropic's "Mythos" model reportedly runs on older hardware too. The controversial "any lawful use" clause that derailed earlier talks is not part of the deal.

When analyzing data, Microsoft Copilot invents country differences where none exist. Mathematician Adam Kucharski fed the tool identical datasets with different country labels, and Copilot delivered detailed stereotypes instead of accurate results. Thinking models catch the trick, but only if users know when to reach for them.

Deepseek is making the 75 percent discount on its top model V4-Pro permanent. At $0.435 per million input tokens, it's at least 11.5 times cheaper than GPT-5.5 and over 34 times cheaper on output. For token-hungry agentic systems, this kind of pricing could squeeze Western providers hard.