/ 3 min read
73%. It looks like a measurement. It isn't.
There is no sensor inside a battery measuring 73% - a December 2009 lesson in state of charge, drift, and models that get more confident than reality justifies.
When a battery display shows 73%, there is no sensor inside the battery measuring “73%”.
What a system can actually measure is voltage, current and temperature. And with lithium batteries - especially LiFePO4 - voltage alone tells you surprisingly little about state of charge. For a large part of the discharge cycle, the voltage curve is almost flat.
I learned this the practical way in December 2009.
We were driving our electric car through Copenhagen during COP15, on a route that would draw “CO2” on the map. The car had battery monitoring I was quite proud of at the time. We measured voltage and used it as our indication of how much battery we had left.
The problem is that voltage moves. It sags under load, changes with temperature, recovers when the battery rests. It gave us a picture - just not one we could reliably translate into remaining range.
So there we were, on a cold December day, driving a route we had planned down to the last kilometre - without actually knowing whether the battery would last.
We didn’t know. But we had to reach the finish. So we continued.
And we made it.
In the battery systems I worked on afterwards, we relied much more on counting. How much current went in, how much came out, for how long. From that, software builds an estimate of state of charge.
Better. Not solved.
A small error in current measurement. Self-discharge that was never counted. A load the system didn’t quite see. A slightly wrong assumption about the battery’s actual capacity. None of it looks dramatic on its own - but integrate small errors over enough hours and cycles, and the number on the display starts to drift.
The software says 62%. The battery has no idea it is supposed to be at 62%.
So the real problem was never just calculating state of charge. It was keeping the digital model synchronised with the physical battery - finding reality again through known full states, voltage behaviour at the ends of the curve, temperature, corrections.
The number on the screen looks beautifully precise. But behind it sits measurements, assumptions and accumulated history - software interpreting a physical system.
That lesson has followed me far beyond batteries. And it feels more relevant than ever as AI moves further into the physical world. The models get more powerful. The underlying problem doesn’t move.
The physical world has a state. Sensors give observations. Software builds a model of what those observations mean.
And wherever software meets the physical world, one of the hardest problems is still the same: making sure the model doesn’t become more confident than reality justifies.
