Why we are going backwards
A running note on what we are building and why.
Every few months something bigger ships. More parameters, more racks, more power per answer. It works, and it has produced real things.
The hardware is not the problem. Performance per watt has improved for decades and is still improving. What we put on top of it ate that headroom almost as fast as it arrived. The chip got efficient. The task did not.
So we are going backwards on purpose. Fewer layers, less memory, less power, on hardware that is still yours to change. Almost nobody is working on that axis, and it is where most machines actually live.
The only number we care about
Joules per task, and bytes per task. How much energy and memory it took to finish a job and go quiet again.
A browser with a few dozen tabs sits on several gigabytes while you read one of them. Opening a news article pulls megabytes of JavaScript to put a few kilobytes of text on the screen. The chip underneath got extraordinarily good over the same period, and the gains went into the layers rather than into the work.
Embedded already solved this, and gave up everything else to do it: firmware runs for months on a small battery, but it does one job, and you cannot decide a year later that you want it running a different model. We want both. Almost everyone is heading right. We are heading up.
Everything ends up local
A machine in a field cannot wait on a round trip and cannot stop working because a link dropped. For most machines worth building that is not an edge case, it is the normal condition.
So the cloud stops being the default and becomes an optional extra. Model, data and control all have to fit on the box, inside a power budget somebody actually pays for. Hard limits produce honest engineering.
Where this actually matters
- Farming
- Most of the useful things on a farm are not the tractor. They are the pumps, the gates, the soil sensors and the cameras on posts, running off a small solar panel and expected to last a season without anyone driving out to them.
- Mining
- There is no signal underground, so a drill or a loader either takes its intelligence down with it or works blind, and whatever goes down has to last a full shift on battery through the dust and the heat.
- Space
- Every watt on a spacecraft is argued over years before it launches and every watt you spend is heat you then have to get rid of, which rules out anything that needs to check with Earth before it can decide something.
- Robotics
- Most robots work for an hour or two and then go back to a dock, and a serious share of that budget is the computer rather than the motors. We want robots that go days between charges, which means the thinking has to get much cheaper.
What we are building
A small computer, an operating system with far less in it than you would expect, and drivers written by people who know what each one costs. It does the work, then it stops.
That is the whole idea. Computers used to be like this. We would like them to be like this again.