4 ms·
What they should do is give an exception for algorithms whose outputs can be completely explained. (e.g. a dumb summarizer that takes the first sentence of ever
by habitue 3y ago
What they should do is give an exception for algorithms whose outputs can be completely explained. (e.g. a dumb summarizer that takes the first sentence of every paragraph and concatenates them would not be considered "AI" for the purposes of the law)
Then we could sit back and watch companies trip over themselves to solve transformer interpretability.
- habinero 3y agoAny algorithm can be "completely explained". It's code, not magic.
- habitue 3y agoRight, so the algorithm that trains the neural net can be explained (autodifferentiation, backprop etc) but the neural net's functioning itself is billions of parameters which we absolutely cannot fully explain how it makes decisions