Sandboxing AI

Interesting video, trying to contain super intelligence is futile, building it to not want to take over the universe is much smarter. Do you think sandboxing could be possible or not?

1 Like

AI as it stands isnā€™t really intelligent, nor does it work at all like biological intelligence. It just does optimization problems really well. Sandboxing it just means running it in a VM, unless youā€™re specifically training it to find escapes for containers, this isnā€™t an issue.

2 Likes

Yeah but in the video they are not talking about the current AI used mainly for science but more like for a theoretical AI that will self learn and have a ā€œconsciousnessā€ kinda.
So they are talking more for a AI that is a robot with a super intelligence and can think for itself ā€¦ which is pretty stupid but you know whatever floats their boat ā€¦

There are ā€œself learningā€ adversarial neural networks all over the place already. Still doesnā€™t apply. They are universally, by design, bound to their parameters. Thereā€™s no known way to make one that isnā€™t.

This understandably makes the video, as well as the discussion, masturbatory and useless.

Yeah I know but they are not conscious right ?
What they are talking is only theoretical and their idea (theory) is kinda stupid but that doesnā€™t mean itā€™s useless. Theoretical science inspire engineers and scientist to prove what is possible and what is not.

So you donā€™t believe that it will ever become a higher intelligence? These are the same sort of questions musk is asking with his AI project.

No bad way to get attention. Takes the heat off of his SE asian haircuts at the same time. Heā€™s a very shrewd man.

You mean musk?

Yeah. Who else could I be referring to?

The SE hair cut part threw me a little :slight_smile: my personal opinion is that neural net layers are going to get stacked very quickly which will make elements of AI obtainable. Then we just need a bunch of people from different fields to pull all the tech together any we will have leaning intelligence.

Not really how this works. Cost to do anything approaching meaningful mind simulation is entirely too high, and will remain that way for decades. Neural networks =/= neurons, and certainly donā€™t have anything to do with intelligence. Duct taping a bunch of them together doesnā€™t magically make a ā€˜trueā€™ AI. We call them ā€˜learningā€™ because training them for optimization problems changes them over time.

As a closing statement Iā€™d like to add that populist futurism is unbelievably retarded

1 Like

I think you are going to be surprised.

I guarantee you Iā€™m not.