Google Earth's new generative AI feature allows anyone to prompt fake satellite imagery into existence — refugees at borders, nuclear plants in Iran, staged crashes. The tool's photorealism makes synthetic geospatial disinformation trivially easy to produce and nearly impossible to debunk at a glance.
Internal documents reveal Flock Safety coaching law enforcement agencies to lobby elected officials on the company's behalf — weaponizing civic trust to sell more surveillance cameras. A former sales rep also quit over the deception, saying simply: "They lied."
Niklas Gruhn coins a sharp new term: a meat proxy is someone who blindly pipes AI output to colleagues without reading, validating, or owning it. The concept cuts to the core of where human judgment still has to live in AI workflows.
Mirendil cofounders — ex-Google and Anthropic researchers — are building AI systems that meaningfully contribute to their own development rather than just serving as productivity tools. The self-accelerating AI thesis is moving from research paper to funded startup.
A new paper organizes 41 distinct agent failure modes by whether the fault lies in the model or the harness around it. Essential reading for anyone shipping agents to production — most failures aren't what you think.
Isenberg breaks down a vocabulary shift gaining traction: prompt engineering asks better questions, context engineering supplies better information, and graph engineering redesigns the work itself as a managed AI workflow. The framing is becoming a real hiring signal.
Mollick flagged something editors everywhere should print and tape to their monitors: AI-written articles are now citing deeply outdated AI research as if it's current, and most editors can't catch it. The gap between AI's output confidence and its knowledge currency is an editorial crisis hiding in plain sight.
Google's Gemini API lead posted a blunt reminder: "You are probably not model-pilled enough. However much you are betting on the models, it's likely under the slope of the exponential right now." Coming from someone with visibility into roadmaps, that's less a hot take and more a warning.
A new paper tests whether agent memory systems actually need an LLM at all — and finds that the expensive model calls used for summarizing interactions are often unnecessary overhead. Production teams burning tokens on memory scaffolding should read this before next month's bill arrives.
Two stories from today don't look related but are telling the same story. Google Earth now lets anyone generate photorealistic fake satellite imagery with a single sentence. Flock Safety coaches police departments to lobby politicians on the company's behalf. In both cases, a tech product is being deployed not just as a tool but as an instrument of manufactured reality — one for images, one for political narrative.
On the agent side, today's research signals are unusually concrete. Forty-one mapped failure modes. A memory paper that may save real money in production. A new vocabulary — graph engineering — that's crystallizing how serious builders think about AI workflow design. The gap between people who understand these distinctions and those who don't is widening every week, and that gap is where the next wave of value gets created.