Picture it: the world has ended in the usual inconvenient ways. Maybe it’s the towering robotic dinosaurs stomping through what used to be your neighborhood. Maybe it’s those mysteriously still-working radios that refuse to play anything but syrupy 1950s hits, on an endless loop, like the universe is punking you with a doo-wop soundtrack. Maybe it’s the nagging fear that you’ll wake up in a basement that smells like canned peaches and bad decisions, realizing—too late—that you should’ve read The Road with a little more attention.
But let’s be honest about the truly unsettling part: no internet. And with it, no always-on, politely confident AI to tell you what to cook, where to go, how to fix your cough, or whether your neighbor’s “totally safe” generator setup is going to explode. What’s a person supposed to do without a chatty digital parrot whispering life advice into their ear all day?
Apparently, someone has been losing sleep over exactly this scenario—and, in a way that feels both brilliant and deeply unserious, they built a solution: CrankGPT.
A chatbot for the end of the network
If you judge it by the name alone, you might assume it’s just a chaotic version of ChatGPT—like the same assistant, but fueled by conspiracy theories and bad vibes. It’s not that. The premise is simpler, and honestly more charming: CrankGPT is an LLM setup designed for a future where the network is gone. No cloud. No server farm. No “we’re experiencing unusually high demand.” Just a box you can power yourself.
And by “power yourself,” I mean literally. This thing runs without a battery and uses an old-school hand crank to generate electricity. It’s the kind of detail that makes you laugh until you realize it’s also the most realistic design choice imaginable for a world where outlets are decorative.
Inside the box is a Raspberry Pi 5 with 8GB of RAM, an audio input/output card, and a 20W hand generator. That’s the core of it: small computer, audio interface, and human-powered electricity—your biceps as the new grid.
What’s actually inside the box
CrankGPT comes from a two-person outfit called SqueezLabs. Their stated mission is to make AI “smaller, cheaper, and faster so you can run it anywhere.” On paper, that’s a product philosophy. In practice, it’s a mild rebuke to the current era of AI, where the default answer to any problem seems to be: “Add more compute until the lights flicker.”
According to the device description, the generator feeds into a capacitor board the duo designed themselves. The point of that board is practical and important: it keeps the voltage delivered to the Raspberry Pi stable. Anyone who has ever tried to run sensitive electronics on improvised power knows that “almost stable” is just another way of saying “about to crash.”
My favorite detail, though, is how physical the whole experience becomes. The designers note that you can feel the power load through the crank: when the system is running LLM inference and text-to-speech at the same time, the crank gets noticeably harder to turn. In other words, the device doesn’t just compute—it negotiates with you. You want a smarter, longer answer? Fine. Pay in forearm fatigue.
Small models, spoken input, and a voice that talks back
The language model itself isn’t one giant, everything-and-the-kitchen-sink behemoth. It can run a selection of smaller models. SqueezLabs suggests small variants of Liquid AI’s LFM2 (for example, 350M or 1.2B), as well as Gemma 3 in a 1B form.
The workflow is designed to feel natural, even when the world isn’t. The machine takes voice input through the audio card, converts speech to text using a custom voice agent written by the creators, then generates a response with the LLM and speaks it aloud using Piper for text-to-speech.
So yes: you can talk to it. It can talk back. And all of it happens locally. No data center. No connection. No “this may violate our content policy.” Just you, a box, and a crank—like a strange little campfire conversation with a machine that needs snacks made of watts.
Why this matters more than it should
It’s tempting to treat CrankGPT as a novelty—an art project for people who collect solar panels and iodine tablets. But there’s a serious point hiding in the joke. This is a proof of concept that shows you can run demanding software on a tight energy budget, and it’s also a reminder that LLMs don’t inherently need absurd amounts of power to be useful.
The creators frame it as a cultural irritation as much as a technical one. They say it offended their “European, small-practical-car sensibilities” to watch people throw kilowatts and thousands of tokens at tasks that small models could handle just as well—at a fraction of the cost and energy. It’s a wonderfully specific complaint, and also… not wrong.
Because here’s the question I keep coming back to: do we actually need a planet-sized engine to answer everyday questions? Or have we just gotten addicted to “bigger,” the way people get addicted to ordering the largest drink because it’s only a little more money, even if they didn’t want that much soda in the first place?
- Portability: running locally means it can work in remote or disrupted environments.
- Energy discipline: when you feel every watt in your wrist, you stop wasting them.
- Right-sized intelligence: small models can be “good enough” more often than we admit.
The supersize future… and the crank in the corner
Of course, the industry momentum doesn’t exactly favor restraint. With Silicon Valley driving, the “supersize-me” philosophy of AI development isn’t going away soon. Bigger models, bigger hardware, bigger bills—until the next breakthrough, the next bubble, or the next reality check.
Still, I can’t shake the thought that in a true crunch—whether that’s a disaster scenario, a field expedition, an outage, or simply a place where power is scarce—these “small practical car” ideas will look less like quirky minimalism and more like common sense.
And if the bombs ever do fall (cheery thought), maybe the last laugh belongs to the people who planned for a world where compute is precious and electricity is a workout. At least until the crew with necks wider than their heads shows up with their loud toys and louder opinions. In that moment, I suppose we’ll all have the same question:
“ChatGPT, what’s for dinner?”
And if the answer is “you,” well… you’ll wish you’d cranked faster.
Comments