Do LLMs use confidence to guide behavior?

DeepMind's new paper applies causal evidence to test whether large models merely report their confidence — or actually use it to steer what they do next.

When a model says "I am seventy percent sure," is that an after-the-fact verbal summary, or the very signal it relies on to choose its next move? Google DeepMind's latest paper, published in Nature Machine Intelligence, stops leaning on correlation and applies causal intervention instead. Across several tasks, the team finds that a language model's confidence genuinely shapes its subsequent behavior — it is not simply reporting, it is steering. The result hands interpretability and alignment research a sharper instrument: calibration becomes part of the control loop, rather than a box to be ticked at the end of training.

GPT-6 Astra labels 13,000 video frames on its own

Higgsfield set GPT-6 Astra loose on parcel-sorting CCTV and watched it produce 13,038 annotations across just 81 frames — no hand labels, no review queue. The team says the use cases are endless: track package handoffs, monitor worker movement against cart traffic, and surface the bottlenecks that slow a line down.

NVIDIA open-sources MiniMax-H3 video generation

Faster than playback, fully open-sourced, and ready to try before you commit GPU. NVIDIA's Sol team has shipped a faster-than-real-time acceleration stack for MiniMax-H3 video generation, exposed through the Reactor API so teams can benchmark it on their own workloads first.

GPT-6 Astra builds a complete Vampire Survivors-style game

With Blender and Godot Engine, GPT-6 Astra put together a complete survivor-like game — an enlarged map, blacksmith weapon upgrades, and a boss. The finish is still rough, but the era of hand-built 3D content is visibly ending.