Dangers of AI

[JohnRoberts]

In the video "The AI Dilemma", did you get to the part where Tristan explained how scientist's used an LLM to "translate" the blood flow in a person's occipital lobe so to reproduce what that person was looking at? (It was a giraffe)
Ai made an interpretable language out of what an eye sees by patterning the blood in an observer's brain.

Extrapolate:
With enough Ai training, any electrical, behavioral, or biological system dataset can be potentially 'unlocked' or bridged via a language model. It's Edward Bernays' wet dream.

Now militarize it.

Sounds like any sufficiently advanced technology that can be used for good or bad.

JR

 

[boji]

Sounds like any sufficiently advanced technology that can be used for good or bad.

True. When development curves are manageable. But what happens when black boxes are trained on perverse goals, and recursively optimized?
Suppose it creates a demand for protective Ai. Having watched too many dystopic sci-fi movies and such, hope I'm overestimating the danger.

 

[JohnRoberts]

True. When development curves are manageable. But what happens when black boxes are trained on perverse goals, and recursively optimized?
Suppose it creates a demand for protective Ai. Having watched too many dystopic sci-fi movies and such, hope I'm overestimating the danger.

We are often informed by forward thinking science fiction writers... I seem to recall some misbehaving AI in "2001 a space odyssey" (1968). Arthur Clarke was a co-author of the screen play that was inspired by his short story "The Sentinel" from 1951. There were/are a handful of science fiction authors who are simultaneously cutting edge scientists well ahead of the mainstream.

It is interesting how new technologies often get adopted by bad actors (like drones). Good luck to us all.

JR

 

[boji]

Seems reasonable if language models can determine political affiliation based on reading faces, it'll do just fine with names (and before long, much more). If social constructs are physically self fulfilling (and gameable), how much more so our personalities?