Thinking On Paper

You can't build superintelligence from words and pictures on the internet.

The spaces around us shape human intelligence.

We learn from spatial intelligence, from cause and effect, from seeing the consequences of our actions. Language models can only take us so far.

Sometimes AI has to get out into the world and see what actually happens when it presses a button.

World models could be the missing link between the AI we have now and embodied superintelligence.

We unpack what world models are, how they're trained, and what they mean for AGI and the future of robotics.

A video model's like a dream and a world model's like a lucid dream.

Jeremy borrows the analogy from an essay to explain the difference between them: agency.

In a dream, you watch. In a lucid dream, you act. World models learn from action. They observe what you do, feed it back into the system and learn the actions around it.

From our episode on world models.

The key is getting significant advances in robotics so these systems can eventually manufacture increasingly complex things, not just build the giant machines we need to bootstrap the industry.

Ultimately, they’re talking about manufacturing computer chips on the Moon. We land lithography machines, and eventually robots are making the chips.

From our episode on bootstrapping the space economy.

Robots in space, starship, asteroids... and when it all comes together.

What if one way to deal with Earth’s growing industrial footprint was to eventually move much of that industry off the planet?

A paper called Affordable Rapid Bootstrapping of Space Industry and Solar System Civilization, by Philip Metzger, Anthony Muscatello, Robert Mueller, and James Mantovani, models a path from a small lunar industrial base to autonomous, self-replicating manufacturing systems that could eventually spread across the solar system.

According to the paper, within a few decades, that system could reach an industrial capacity millions of times greater than that of the United States. The idea is that space could provide vastly more resources and room for industry without putting the same physical demands on Earth.

But it raises a fun question: should humanity’s industrial future be off-planet or on-planet?

Liftoff and successful orbital insertion burn for Starship. A single Raptor engine powered the way, proving the system's capability. No alerts, just a smooth transition to orbit.

SpaceX launched Starship Flight 14 on 28 September 2026, and it became the first Starship to reach orbit. The Super Heavy booster made a soft splashdown in the Gulf after its landing burn, and Ship 41 fired its engines to enter orbit at about 275 km.

Once in orbit, Starship deployed all 26 of its Starlink V3 satellites, the first V3 satellites to reach space.

SpaceX streamed the launch live on its X channel. It was fabulous. As every tech media agency, it is our job at thinking on paper to download the video and upload it here and say something profound about space infrastructure and what this means for humanity and the industry and off world habitation and lunar manufacturing and ambition.

Is your data center too hot to handle?

Today, the hottest component in a data center server is typically the chip itself. That’s why the industry is increasingly turning to direct-to-chip cooling.

A cold plate is mounted directly on the chip, with pipes carrying water or dielectric liquid through it to extract heat. The limitation is that this only cools the chip. The other components in the system still rely on air cooling.

A recent study revealed incredible results: we're saving nearly 40% of energy when running compute in the rack with our technology compared to air-cooled systems.

This focuses on energy and water efficiency.