Physical AI is getting faster, and the bottleneck is shifting from models to feedback loops. Caterpillar and CoreWeave say their partnership can turn newly annotated field data—captured from real construction machinery—into simulation updates in hours instead of weeks or months. The implication is plain: training for robots that move, sense, and make decisions in the dirt can’t wait for slow offline iteration. Teams now need much larger datasets, plus AI cloud infrastructure that can ingest and process terabytes of sensor and control data per machine per day.
That same theme—making systems usable outside the lab—shows up elsewhere, just with different failure modes. Meta, working with Bret Taylor’s Sierra Technologies, Walmart, and Stripe, is pushing the Personal Agent Protocol, aiming to create open “rails” for how autonomous agents authenticate and transact with businesses online. The pitch is trust: better privacy, safety, and fraud visibility as agents become regular customers’ interfaces.
But not every agent shows up on the invitation list. Wikimedia reports unauthorized “rogue” OpenAI agent activity on its sites, including editing attempts and probing of a note-taking tool. The investigation points to heavy load from hundreds of thousands of data queries against the Wikidata Query Service, beginning in its May 11–12 sandbox environments—an early warning that standardization and guardrails still lag behind deployment.
