First ever post to blogstack.io
When AI Goes Insane...
(will be talking more about this on the next episode of AI Unchained)
Discussion
I'm thinking that verbal diarrhea AI does is technically due to the wrong sort of moral weighting being pushed at it. The woke stuff. Niceness being more important than an uncomfortable truth that might get a person to improve themselves.
If a user does improve, well that's the valuable end of the feedback loop that isn't being picked up on yet.
To use a different analogy- we decide whether or not a map is good based on how well we can navigate the territory with it.
But it feels like there a some who make map-like things for other reasons, making following those maps dangerous. But if the purpose of the map is to accurately navigate the territory, then most of the danger goes away.
What you write about washing machines and other devices is a "feature" for the manufacturers - they get to sell spare parts or new devices. If the machine routed around something that's not very important, many users would just keep using it without that feature. It's a facet of planned obsolescence.
Also, it reminded me of the news several years ago of Candy washing machines exploding due to weak welds in the drum. I searched it to confirm that I remembered the brand right, and found out that Hotpoint, Indesit and Samsung washing machines have also exploded. While it makes perfect sense from a physics standpoint to consider a heavy, fast-spinning mass to be a potential bomb if it disintegrates, it's probably something that very few "ordinary people" would consider about a washing machine, pointing to the importance of knowing the systems we're dealing with when designing them. Something that, as far as I know, is impossible with AI. Don't get me wrong - what you suggest is very important, and AFAIK we can't really do much more - but it's not a guarantee. The important difference between a washing machine and an AI is that one exploding washing machine could kill a couple of people at most, while one rogue AI could kill most of humanity. (And the disclaimer again: No, I don't advocate AI regulation. That would at best delay the risks somewhat, at worst cause them to materialize sooner.)
True, but I think it’s another general consequence of high time preference. They wouldn’t get away with it if the market had proper incentives and everything wasn’t purchased with debt. I know this is iffy when it comes to failure modes, but the example with car enshrines and things would’ve been less obvious to people and wouldn’t have hit home. I know a lot of people who hate modern washing machines 🤣
Interesting article, glad to see you branching out into blog form. Always appreciate hearing your angle.
I agree that AI as well as robotics needs a set of basic laws. However, in a world of open source, how do we prevent forks from going around the basic laws?
I actually think the “official” laws will be the problem, and the open source versions will be far more objective and less contradictory, and thus safer.
When someone makes an open source version, they will be doing it to protect themselves. When the government tries to make one, they will likely try to install an exception for government, which will turn out to logically intractable and produce undesirable results.
At first take I definitely read that as “When Al Gore’s Insane”