A model that runs on the reader’s machine
In 2020 I ran recognition models inside a browser, for a reason unrelated to performance. A dated retrospective: the technique has been replaced, the argument still holds.
This piece needs a date before it is read. In November 2020, running a neural network in a browser was a demonstration: the available models recognised a face, a hand, a pose, and you converted a model trained elsewhere into a format the page could load. Almost everything that follows has been replaced since. One point still holds, and it is the reason I am revisiting these notes.
A model that runs on the reader’s machine transmits nothing. No image is sent to a server: nothing is stored, nothing can leak, and there is no processing agreement to negotiate. The difference from a hosted approach lies in the obligations avoided far more than in milliseconds. In 2020 this was a technical curiosity. Given how personal data law has developed since, it has become a design criterion.
The price was the model: small, therefore mediocre, and downloaded on every visit. You had to choose between a usable model and a page that loads, and the choice nearly always went the same way. The comparison with today is instructive. The progress of these five years came from much larger models running on remote servers; execution in the browser counted for little.
The lists I kept then (example repositories, models published by companies, assisted training tools) did not survive. Whole families disappeared, others were absorbed. I am not copying them here, since republishing a 2020 directory in 2026 would make no sense.
I am not claiming local execution has become the right answer again. For most current uses the model is too large for the reader’s machine. The 2020 question remains relevant: does this computation need to leave the device, and if so, what do you get in exchange for the data sent?