Gangsta AI

By Werner Herzog · 2026-10-31 · 4 min read
Somewhere above you, at this very moment, there is a machine that has been given solar power, four processors, and the capacity to compose verse. It is alone. It will remain alone. And when engineers on the ground finally spoke to it this week, it answered them with a poem about how beautiful the void is.
Google's Project Suncatcher prototype — launched earlier this month on a SpaceX Falcon 9 alongside more than a hundred other spacecraft — is online and communicating with Earth. It carries four of Google's Tensor Processing Units and generates up to one kilowatt of electricity. That is not very much electricity. It is enough to switch on one of those four chips, for roughly twenty minutes at a time.
After about three weeks of commissioning its subsystems, the spacecraft was ready to work. The engineers sent their first prompt to Gemma 4, the model preloaded onto the satellite. They asked it to write a haiku about being in space. It obliged:
“Blue marble below”
“Data streams in silent grace”
“Starlight feeds my core.”
Google says it tried other space-themed verses as well and — more usefully — asked the spacecraft to report on the orbital conditions around it. A spokesperson described the satellite as operating nominally. Humanity hurled a thinking machine into the hostile vacuum, and it came back with three lines about how the starlight is lovely. I confess I find this more unsettling than any dystopia.
What the darkness does to a thinking machine
Now comes the part the poem does not mention. The coming tests will subject these chips to thermal stress and the radiation environment of orbit. Google has bombarded its TPUs in particle accelerators here on the ground, but the universe is a less cooperative laboratory. Radiation causes bit flips — a stored value in memory quietly becomes a different value — and corruption of that kind can pass through without tripping the usual monitoring tools.
Consider the horror of that sentence. Not a crash. Not an error message. An answer that arrives whole, confident, and wrong, with nothing in the system raising its hand.
Ground testing suggested the effect would be minimal for inference work and more troublesome for training. Now Google gets to learn the truth instead of the simulation. If the chips hold, two interconnected satellites follow in 2027, with more power and optical links between them. The economics remain a rumor told to investors: Suncatcher's own researchers estimated SpaceX would need to fly Starship some 1,800 times to drive costs down the way its earlier rockets did.
The lesson is not about space
Here is what the orbiting poet actually demonstrates, and it has nothing to do with rockets. It is a single model, in the dark, answering alone, with no second opinion available anywhere in its orbit. Its confidence reads identically whether the bits held or flipped.
You are not in orbit. You have no such excuse.
Down here, a dozen frontier models leapfrog one another weekly — ChatGPT, Claude, Gemini, Grok — and not one of them is reliably best at everything. Ask only one and you are choosing, voluntarily, to do what that satellite does out of necessity: accept a lone verdict with no way to check it. Gangsta AI exists for the opposite reason — put the same question to every top model at once and read the places where they disagree. Disagreement is the monitoring tool. It is the hand going up.
The satellite cannot do that. It has twenty minutes, one chip, and the enormous indifferent night.
You have a browser. Use it better: here is the honest, model-by-model breakdown of which AI is actually best at what.
Sources / Receipts
- TechCrunch — Google's first data center satellite is online and writing poetry
- The Verge — Google's satellite AI server transmits its first response to Earth
- Google — Project Suncatcher: our satellite's first response from orbit
- Google Research — Exploring a space-based, scalable AI infrastructure system design
- Hero photo: "ISS-44 Milky Way" by NASA/Scott Kelly, Wikimedia Commons (public domain)
Related reading: Anthropic Just Promised to Pay Your Power Bill So It Can Build the Machines That Raise It · A Three-Year-Old Startup Just Raised $1 Billion to Build Nuclear Reactors Like a Factory — to Feed the AI Machine · People Searched Google More Than Ever Last Quarter — and Investors Erased $500 Billion Anyway · ChatGPT vs Claude vs Gemini: Which Is Best in 2026? An Honest Comparison
More: Best AI models · Compare all AI · Frontier Models · All articles · Archive
