Artificial Intelligence

Pixels of a microscopy image, a patient's vitals, the stray magnetic flux around a motor, the words in a sentence. The model has no idea what any of it means. It only knows there is structure in there worth finding.

A neural network is a graph: simple units in layers, joined by numbers called weights. Training is nothing more than searching for weights that turn inputs into useful answers.

Push the signal in. Each layer multiplies, sums and squashes. Some paths carry almost everything, most carry almost nothing. The network is composing features out of raw measurements.

At the output the model commits, and says how strongly. That output has been a diagnosis, a severity tier, a fault in a winding, a forecast a government acted on. Scale the same machinery up and feed it language, and the thing it commits to is simply the next word. Nobody expected that to be enough. It keeps turning out to be, which is the most interesting problem I have to work on right now.