Google Earth's new generative AI feature allows anyone to conjure fake satellite imagery with a single sentence — refugees at the border, nuclear plants in Iran, crash scenes in Amsterdam. The implications for geopolitical disinformation are immediate and severe. This is the "photos are evidence" assumption dying in real time.
MCP 2.0 (formally the 2026-07-28 spec) dropped this week and Willison calls it the most significant change to the Model Context Protocol since launch. Stateless architecture means agents can swap underlying harnesses — Claude, Codex, Goose — without losing context. The plumbing of the agentic web just got standardized.
DeepSeek dropped another open-weight model with "substantially enhanced agentic capabilities" — 304 billion parameters, available on Hugging Face at 167GB. Artificial Analysis benchmarks suggest it outperforms its size class by a significant margin. The open-weight arms race is not slowing down.
ISBNdb quietly marketed a service to sell scanned books for AI training data, then pulled all webpages and claimed the whole thing was a "test of market interest" after 404 Media published. The denial is barely plausible — but the reporting worked.
Mirendil cofounders — ex-Google and Anthropic researchers — are building systems where AI meaningfully contributes to its own development. The framing shifts AI from productivity tool to self-accelerating loop. Early, but the pedigree is serious.
Ryan Carson built Treehouse to 110 employees and a million learners — now he runs an AI divorce-law agent solo, tracking toward 4x revenue growth this month. The new leverage isn't headcount, it's parallel cloud agents operated by one person with good judgment.
Demis Hassabis announced Gemini Robotics 2, claiming robots can now "reason through every movement" to handle tasks previously out of reach. Physical AI is the next frontier everyone agrees on but few have shipped — Google DeepMind is making its move.
Miles Deutscher flagged that Anthropic released new context engineering rules specifically for Claude 5, arguing that most users are running the model wrong. Whether or not the framing is hyperbolic, the underlying point lands: prompting a frontier model without understanding its architecture is leaving significant capability on the table.
a16z posted "The intelligence explosion is a manufacturing explosion" alongside their Charts of the Week. The frame is deliberate — compute scarcity, not model capability, is now the binding constraint. Their other line, "during a gold rush, sell air conditioning," tells you exactly where they're placing bets.
DAIR.AI amplified a blunt observation: the race to make intelligence "too cheap to meter" is on, with DeepSeek-V4-Flash already proving capable on agentic tasks. This isn't abstract — it's the commoditization of reasoning, playing out in weeks, not years.