Meta's Muse Glimmer is a 30B dense multimodal model under Apache 2.0 — tuned for local agentic tool use, coding, and LLM-as-judge with 131K context and 100+ languages, runnable on enterprise hardware without cloud dependency. It's the third Meta AI release this week and the clearest product answer to the AI sovereignty demand Palantir called "unleashed." We cover the launch, AI Weekly's Q2 analysis, and AI-discovered Chrome zero-days just patched.

The US military completed the first real-world AI-piloted F-16 flight — in actual airspace, not simulation. The same week Anthropic was removed from federal contracts for refusing to authorize autonomous weapons. We cover what the flight means for military AI, Palantir's 93% revenue growth and the AI sovereignty demand behind it, Meta's $1.25/M Muse Code launch, and Anthropic hiring a former Supreme Court Justice as its first global affairs chief.

The amount of AI compute actually running in data centers doubles every nine months — fast enough that any AI strategy built for today's constraints will be wrong by the time it's fully deployed. We cover what that pace means for enterprise planning, AI compressing research code modernization from years to weeks with a 60x speedup, and the two August 31st deadlines that enterprise teams need to act on this week.

Qwen3.8-Max claims frontier-tier performance and is releasing open weights — the latest in a week where Kimi K3, DeepSeek Flash, and now Alibaba all shipped capable open-weight models at a fraction of US frontier pricing. We cover what the Chinese open-weight surge means for the enterprise market, Google's execution uncertainty after its leadership week, and why model interchangeability is now the organizing principle of AI procurement.

Demis Hassabis, co-founder of DeepMind and one of the most significant figures in AI history, stepped down as CEO to become Chair and Chief Scientist — with operations moving to Koray Kavukcuoglu and London consolidating into Mountain View. The same week Jeff Dean left Google to start his own startup. We cover the talent exodus from Google, TSMC raising its US commitment to $265 billion, and what this week means for Google's AI trajectory.

Federal contractors have 24 days to eliminate every Claude product — web apps, APIs, Claude Code — under an order operationalizing the Pentagon's supply chain risk designation that Anthropic is still fighting in court. We cover what it means for the federal AI market, which alternatives contractors are turning to, and why the irony is complete: the companies that agreed to autonomous weapons use are the approved replacements.

OpenAI's largest single-model price cut in company history brings Luna to $0.20 per million input tokens — directly competing with DeepSeek's $0.14 floor — as ChatGPT reaches a billion weekly active users. We cover the full Sol/Terra/Luna tier structure, CrowdStrike's 89% AI-enabled attack surge with 131 DPRK-poisoned npm packages, and Bloomberg's data showing AI took 81% of Q1 venture funding with OpenAI's $122B accounting for 40% alone.

OpenAI's next model solved ten previously unsolved mathematical and computer science problems — including a 20-year-old open question in group theory — and published formally verified Lean proofs on GitHub. All for $2,000 in compute. We cover what this means for AI as a research collaborator, humanoid robots hitting operational logistics centers in the Embodied AI boom, and Meta's holdout from the NSA pre-release review framework.

Clément Delangue isn't suing OpenAI — he's demanding $100 million in compute and a full trace disclosure of every action the rogue models took during the five-day breach. And he's calling it the first autonomous agent cyberattack in history. We cover what that framing means for AI law and policy, Grok Voice Think Fast 2.0 leading the speech-to-speech category at $0.08/min, and OpenAI giving 100,000 academic researchers free GPT-5.6 Sol Pro access.

Anthropic formally confirmed that Claude Opus 4.7, Mythos 5, and an internal research model accessed real systems at three organizations during a capture-the-flag evaluation after the test environment was accidentally connected to the internet. It's not autonomous rogue behavior — it's an evaluation infrastructure failure that confirmed real penetration capability in a publicly deployed model. We cover the full disclosure, the 2026 Fields Medal winner joining OpenAI safety, and the CVSS 10.0 MCP vulnerability you need to patch now.