Import AI Tracks Google Orbital TPUs And Zhipu’s Self-Optimizing Model Work

Import AI’s latest edition turns two AI infrastructure experiments into one operational question: whether the next constraint for model development is where computing runs and how quickly labs can automate the work that makes it cheaper.

The newsletter first follows Google’s Project Suncatcher, an effort to put AI computing hardware in orbit.

Google is preparing with Planet to send chips on a SpaceX rideshare mission, moving the idea from moonshot framing toward a physical test of whether machine-learning hardware can tolerate the launch and orbital environment.

The space-computing section focuses less on a finished product than on the engineering evidence Google has disclosed so far.

Trillium TPUs have been put through stress testing for launch forces, and radiation testing showed Trillium TPUs withstanding a dose above the level expected across a five-year mission.

That does not make orbital training routine, but it narrows one of the most obvious hardware questions before a mission carries the chips.

Cooling remains the unresolved operating problem.

Data-center chips turn electricity into heat, and vacuum makes that heat harder to remove because ordinary air movement is unavailable.

The source frames that thermal question as the main practical hurdle after launch and radiation tolerance, especially if orbital systems are expected to do more than small demonstrations.

The reason Google is pursuing the route is tied to the scale of AI computing.

Space offers wide physical area and direct solar power, two resources that become attractive when AI training and inference demand more energy and more room.

The confirmed next step is the space chip flight, not a working orbital data centre, so the article’s strongest evidence is about environmental qualification rather than commercial deployment.

A second infrastructure thread shifts from hardware location to software labor.

Zhipu AI used GLM-5.3, its own open-weight model family, to help build and optimize infrastructure for GLM-5.3 Flash, a faster and cheaper version of the model.

The company described a loop in which engineers set objectives and boundaries, an Infra Agent analyzed problems and changed code, and an experimental environment supplied feedback.

That workflow produced the edition’s clearest operating result.

The work took less than two weeks between early adaptation and production use for GLM-5.3 Flash, and end-to-end throughput reached three times the starting baseline.

The newsletter treats the example as a sign that model developers are beginning to use their own systems not just for product features, but for the internal infrastructure work that determines serving cost and speed.

Zhipu’s advice centers on feedback design rather than broad autonomy claims.

Feedback has to be local enough to point the agent toward a launch parameter, kernel, code path or input condition.

It also has to be progressive, moving from fast syntax and type checks to unit tests, then to integration tests, microbenchmarks and production metrics.

In that setup, the agent is useful because the surrounding software environment makes failure legible.

The same edition also records a research argument from Michael Levin about minds, bodies and mathematical patterns, and a robotics post-training proposal from Perry Dong.

Those sections widen the scope from immediate infrastructure work to the research assumptions behind intelligence and embodied systems.

The practical through-line is still tooling: researchers are looking for ways to make biological, robotic and computational systems more testable.

Taken together, the issue does not show one completed AI platform.

It documents a set of experiments around where compute can sit, how models can improve their own infrastructure, and how research communities are trying to standardize methods around hard-to-measure systems.

The next confirmed milestone is Google’s space hardware test, while Zhipu’s reported production result shows that self-optimization is already being tested inside model operations on Earth.