Signals from the AI frontier.

Most video systems sample a clip at a fixed rate, turning every minute into a large, uniform stream of tokens. Google’s new agentic video mode takes a different approach: Gemini can search, scan, slow down, and revisit selected segments across frames, audio, and transcripts. The model spends compute according to the question instead of the duration of the file.

That design resembles human inspection more than passive ingestion. A safety analyst can search hours of footage for a brief anomaly; an editor can find a cut boundary measured in fractions of a second; a tutor can retrieve the one moment in a lecture that answers a student’s question. Google reports up to 88% lower token use, up to 66% lower cost, and quality gains of up to 7% in its published tests.

The broader signal is active perception. Multimodal models are becoming systems that form a search plan, use native tools to gather evidence, and revise what they inspect. For builders, the interesting metric is no longer only how much media fits in context—it is how intelligently a model allocates attention.

Software agents have become competent at operating APIs, browsers, and terminals. Physical equipment is a harder frontier: microscopes, liquid handlers, robotic arms, and specialist instruments expose incompatible interfaces and carry consequences that cannot be undone with a keyboard shortcut.

Anthropic’s Model Hardware Standard research preview proposes a shared driver layer built around simple primitives such as read and write. Devices can expose capabilities and physical characteristics in a consistent, machine-readable form, while natural-language metadata gives an agent context that code alone may omit. Anthropic says early integrations have reduced setup work from weeks or months to hours or minutes.

The standard is early and access is limited, but the direction matters. Reliable physical agents will need more than stronger reasoning: they need permissions, device discovery, safety constraints, observability, and predictable failure handling. The interface between a model and a machine may become as consequential as the model itself.

Benchmark contamination has become one of AI’s quiet credibility problems. Training corpora are enormous, test sets circulate online, and proprietary models reveal little about what they have seen. A high score can be genuine capability, memorization, or some mixture that evaluators cannot inspect.

Google DeepMind’s pilot uses cryptographically secure environments to separate model developers from hidden evaluation material. The approach is double-blind: evaluators do not receive access to proprietary model details, while developers cannot view the benchmark questions. Models run inside a controlled environment and only approved results leave it.

This is infrastructure rather than another leaderboard, and that is precisely why it is important. Stronger models raise the value of credible measurement. Secure evaluation could support sensitive tests in cyber defense, biology, safety, and other domains where publishing the full question set would destroy its usefulness or create risk.

Tropical cyclones are difficult because forecasters need several answers at once: where a storm will travel, how strong it may become, how large it will grow, and how uncertain each prediction is. Google DeepMind’s WeatherNext 2 experimental cyclone system combines AI forecasts with a specialized hurricane model to address those questions together.

DeepMind reports improved track and intensity predictions over its earlier system and says the model can generate large ensembles rapidly. Ensembles matter because one deterministic path can hide uncertainty; many plausible forecasts reveal the range of outcomes that emergency teams must plan around.

The advancement is not simply replacing physics with a neural network. It is the assembly of learned forecasting, domain-specific simulation, uncertainty estimates, and operational workflows. The most useful scientific AI is increasingly a component inside a larger instrument—one whose output can be checked, compared, and turned into a real decision.