Geoffrey Hinton Explains AI Dangers to Neil DeGrassi Tyson [video]

90 minute video worth every minute. When Tyson gets wowed, I pay attention.

Hinton's self promotion is lumbering towards an AI Heaven's Gate or Jonestown. His explanations are overly slanted towards common sense analogies where this tech cannot be genuinely understood by common sense, any more so than particle physics and holography can be understood with common sense.

In order to test one's thinking about anthropomorphic effects of AI agents, start with a question: Is the apparent humanity of agent behavior a function of the training, or of deep structures of the models? The training data are clearly human; the bulk of the human canon has been fed to the large models. That's while point of the training process! So, what deep structures of the models are humanoid? How are these defined and designed, and how do they work?

If the answer to the latter point is "It's a mysterious effect!" then we must assume that the agent is retrieving and organizing artifacts of the training data.

There's another electronic device that works in this way and which design predates digital systems designs: Television!

When was the last time you heard anyone wonder if the TV is hiding its true powers on the channels other than the one that's currently tuned?

Tyson is committing a sin of science journalism here by not just abiding Hinton's anthropomorphisms for model effects but by encouraging viewers to drink Hinton's expository koolaid.

As to why Hinton is becoming ever more slanted towards a cultish evangelism: I suspect Hinton became enamored with this tech beyond his own ability to grasp and has reified it in response to the excitement of his newfound celebrity and his inability to rationalize effects; an obvious hazard for engineers whose work leads to fame.