Trained on All Code Ever plus every CS book and manual ever written and has gradient-descent distilled this into a vast many-dimensional matrix into which you can easily feed new data.
With the side effect that it can now be manipulated (or spontaneously update weights and "mutate") in ways that may be simply wrong, *subtly* wrong, biased, or actively hostile to you.
Or may have any of those undesirable things -- again, intentionally or not -- in the initially-trained weights.
