The scientist, a pioneer in neural networks, warns that ‘the capabilities of artificial intelligence continue to advance at a rate which seems faster than risk management practices’
These are just statistical models trained on the natural language output of real humans. If real humans are statistically likely to ignore instructions in a particular case (or do other undesirable things like misunderstand, lie, confabulate, etc), then the statistical model trained to simulate human output will do the same.
It would only be surprising if this weren’t the case.
Exactly.
These are just statistical models trained on the natural language output of real humans. If real humans are statistically likely to ignore instructions in a particular case (or do other undesirable things like misunderstand, lie, confabulate, etc), then the statistical model trained to simulate human output will do the same.
It would only be surprising if this weren’t the case.