Causation
How today's AI learns.
Large language models are trained to predict the next word of a text. To do that well, they learn from a large part of the public internet, and store what they learn in billions of numerical weights.
The results are impressive. These models write fluently, summarise, translate and answer questions across almost any subject. The method behind them has four consequences that matter when the setting is a secure network, an aircraft or a remote site.
It needs a data centre
Training and running a large model takes racks of graphics processors. Most organisations reach one over the internet, so their data travels to somebody else's computer to be processed.
It stops learning
Once training ends, the weights are fixed. Teaching the model something new means another round of training, done by whoever owns it.
It needs the connection
A model that lives in a data centre is unavailable when the link is down, jammed or too slow, which is when a field system most needs it.
It guesses fluently
The model produces the most likely continuation, whether or not it is true. A wrong answer reads exactly like a right one.
None of this is a flaw in the engineering. It follows from learning associations at scale. A different way of learning has different consequences, and that is what the next page describes.