The file
Sourced AI briefs in publish order — newest first. Citeable originals live on the homepage under From Signal Desk.
Meta’s ads model shows LLM infrastructure does not transfer cleanly to recommendations; efficiency comes from rebuilding kernels and parallelism around irregular user histories.
Orchard treats the environment, not the model, as the reusable core of agent research, narrowing the gap between laboratory training and deployment harnesses.
Huge enrollment validates demand for agent-building skills, but the drop from registration to capstone work reveals the limits of headline participation metrics.
Microsoft’s Echoverse result is that computer-use agents improve more from deeper, checkable worlds than from adding more shallow sites.
Gemini Robotics ER 2 is Google’s attempt to make a robot’s planner watch video, time its next step and coordinate other machines without a stop-and-think pause.
Managed Agents are becoming a sandboxed worker with policy hooks and budgets, not just a one-shot coding call.
Gemini Robotics 2’s claim is embodiment transfer: one intelligence layer for walking, two-handed work and on-device adaptation to new robot bodies.
Google is using dinner-party planning to show that Search’s AI Mode is becoming a visual, app-connected assistant rather than a list of links.
Google is positioning AI Mode as an offline-life planner, using calendar, shopping and ticket actions to justify more personal context in Search.
Google’s Genesis pledge is an in-kind compute and model package for DOE labs, not a cash grant, and its value will be measured in lab workflows rather than the headline figure.
Google is splitting the Flash line into a cheaper workhorse, a high-throughput Lite model and a restricted cyber specialist rather than shipping one general upgrade.
Flash Cyber is Google’s bet that cheap, repeatable vulnerability search beats a single expensive model call, and it is being withheld from general developers for that reason.