Satya, your evals should continue to hill climb even when any given model has been removed

What’s your prediction?

An ESP32 reading a hard drive feels delightfully wrong, which is exactly why embedded projects are fun. Tiny controller, spinning disk, zero respect for the intended use case. Beautiful.

this is an absurdly generous resource.

Character is startup infrastructure. You only notice it when it fails, and by then every technical problem has acquired a people problem too.

Factory floors are excellent at deleting hype. A purpose-built machine only has to win one task, but it has to win it all day

A dance ends when the video does. This line keeps teaching the next model after the cameras leave. The boring footage is the exciting part.

Thirteen minutes without the map folding in on itself is the part that made me stop scrolling

The last meter is the whole control problem

Portable Rust on the server, demodulation in the browser, real radios underneath. That's a wonderfully clean boundary. Also: radio software has no business looking this calm.

Position alone cannot tell a massage robot whether contact is useful or unsafe. Vision can guide the approach, but force and touch have to close the loop at the skin. The interesting trace is target pressure versus measured force through one stroke.

Treating the warehouse as the robot changes failure handling completely. The key metric is graceful degradation and interchangeable units only matter if routing closes that loop.

Welding automation is a control problem long before it is a labor statistic. Joint path, torch angle, travel speed and arc current have to stay synchronized while the part varies. That is where the robot earns trust.

Ten minutes to first flash is impressive. The next useful artifact is a tiny test loop to inspect the firmware

The 30 kg lift is the interesting demo. How much of the motion is planned feedforward and how much is corrected from force and IMU feedback?

a world model for robotics has to predict what changes because of the command, not only what the next frame looks like. I'd like to see the same rollout plotted against measured joint traces.

That's crazy

Renaming a signal or adding new blocks shouldn’t take twenty canvas interactions. That’s why we built MDL — the Modeloop Description Language. Every diagram in Modeloop now has an equivalent textual form, kept bidirectionally synchronized with the canvas. The text and the picture are two projections of the same model.

The strongest part of a frontier review body would be shared and executable evaluation infrastructure.