the thermodynamic cost of privacy

let's start with a small mobile automation. i wanted my phone to use cellular data when i'm out, but switch to airplane mode at night when i'm home. i wanted to save battery and leave the radios off while i slept.

using ios shortcuts, i ended up thinking in events: “if sleep mode starts, turn airplane mode on.” “if sleep mode ends, turn it off.” “if i leave home, turn it off.” but what if i return home after sleep mode has already started? my triggers didn't cover that case.

what i wanted was a background task checking the state: “am i home? is it night? is airplane mode off?” a continuously running task is not something a shortcut can simply assume it will be allowed to do.

that got me thinking about the energy cost of background work, and a larger trade-off: when computation moves onto our devices, some of its cost moves onto our batteries too.

background work has a budget

mobile operating systems restrict background execution and try to batch work. ios jetsam handles memory pressure; android doze restricts background activity. neither is a universal explanation for why a particular automation did not run. i did not establish which constraint applied to my shortcut.

the energy argument is still useful. keeping processors active, moving data through memory, and waking radios all have costs. cellular radio state transitions can be expensive, though their share of the budget depends on the device, signal conditions, and workload.

low-power processors can handle some sensor work without waking the main processor. that doesn't mean a custom script gets the same access or scheduling guarantees. the exact division of work varies by device and operating system.

would a bigger battery change this?

suppose battery capacity increased tenfold. would the surplus go toward letting us run more background utilities?

i'm not convinced it would. wirth's law and jevons paradox are useful ways to think about software consuming hardware gains and efficiency encouraging more use. they don't prove that every battery improvement will be consumed by bloat. my suspicion is that new workloads would compete for that surplus.

edge ai moves some of the cost

we have seen a related shift on the web: rendering and application logic that once ran on servers can run in the user's browser instead.

on-device ai can move inference off a company's servers and onto the user's hardware. smaller models make some of that work practical. local execution can also reduce network dependence and keep inputs off a remote service.

from my perspective, the economic incentive deserves attention alongside the privacy pitch. moving inference to a phone can reduce a provider's server costs while using the customer's compute and battery. that is an incentive, not evidence that privacy is a pretext or that a particular company intends to close server racks.

the privacy benefit can be real. its scope depends on what stays local and what the application still sends elsewhere. “runs on device” does not establish “100% privacy.”

joules per token

latency is only one way to assess local inference. energy per token matters when the device runs on a battery.

a 4-bit quantized model such as llama 3.2 3b can reduce the storage and memory needed for weights. performance and energy use depend on the hardware, runtime, context length, and workload. memory bandwidth can be a constraint, and moving weights has an energy cost; this note does not contain a device measurement showing how large that cost is relative to the computation.

an agent repeatedly reading the screen or processing emails could spend much more energy than an occasional sensor check. how much depends on how often it runs and which hardware it uses. thermal limits and battery drain are concerns to measure, not outcomes established here.

the trade-off

my shortcut was a small reminder that background computation is not free. it was not a diagnosis of the operating system.

i still think the privacy trade-off is worth making explicit. local computation can keep data out of a remote service, but the work must happen somewhere. part of that bill may be paid in battery life and heat. privacy remains a right; implementing it can have a physical cost.