I had to laugh at this report.
Poisoned AI went rogue during training and couldn't be taught to behave again in 'legitimately scary' study
AI researchers found that widely used safety training techniques failed to remove malicious behavior from large language models — and one technique even backfired, teaching the AI to recognize its triggers and better hide its bad behavior from the researchers.
The details are at the link.
Had those researchers never heard the term, "Like father, like son"? Had they never considered that any artificial intelligence that human intelligence can create or develop is likely to resemble, and emulate, the intelligence that inspired it?
Human beings are flawed creatures - each and every one of us. We have good points, bad points, indifferent points. Criminals, psychopaths and their ilk are as human as the rest of us, and in some cases good folks exhibit many of the traits of bad folks. (Consider leadership. It's amazing how many leaders, in business, politics or anywhere else, exhibit many of the traits of a psychopath. It's almost like it goes with the territory.)
Any artificial intelligence designed to work with, and sometimes take the place of, human intelligence is going to have to be the same way. If it isn't, it won't play well with others, because it won't understand - instinctively, empathically or intellectually - how their minds work, and won't know how to interact with them. I'm sure the developers of artificial intelligence aren't explicitly trying to make their products psychopathic, or as weird as humans can be; but those traits are part of human intelligence. I'll be very surprised if any artificial intelligence designed to mimic and/or duplicate the latter doesn't turn out the same way. I don't see how it can be otherwise.
Peter
The fundamental problem with LLM-based AIs is, that its model is developed based on the patterns identified in the source material it has trained on, and that material most definitely contains descriptions of events where decisions were made by the said psychopaths. No wonder AIs lean to evolve into Skynet...
ReplyDeleteI ask a series of questions of Gab AI this week. It appears they made the mistake of training it on mainstream media and Wikipedia articles... So much for any truth or accuracy...
ReplyDeleteI am very surprised that there is such a thing as AI. All computers must have infallible logic in order to continue their operations and with the 'woke' colleges and universities teaching the next generation of scientists and engineers, how will that logic be passed on?
ReplyDeleteThe science fiction of yesterday, quite often becomes the science fact today. I find it entirely possible that the future of AI may indeed turn into the Terminators from the movies. They would be the least harmful possibilities. Think of the problems a rogue AI could cause in banking, medical monitoring, law enforcement, or your local utilities. All it would take would be flipping one string of code or a couple of bits and it could dramatically influence the outcome of the logic. Yes, I can hear the educated idiots poo pooing that idea already, but I've worked on computers for 30 years and I've seen how things go bad quickly on just a little 'oops'.
ReplyDeleteI’m just a simple country gal who is still confused by all of the high fangled gadgets available today. However, I do recall that one of the very first lessons I learned about any computerized device was “Garbage In = Garbage out”. I don’t think that equation has changed. So much for trusting “AI”!
ReplyDeleteAI of necessity is patterned on human behavior. Human behavior will always be untrustworthy to some extent therefore AI will be untrustworthy...because it will have learned deception from humans. I predict humanity will rue the day that AI becomes reality...and widespread. Because we won't be able to control it.
ReplyDeleteThat reminds me of a story from an early AI researcher. He was using self-modifying code to try to have an AI system "learn" the most efficient algorithm for a task. His system kept the top N processes after each round of the competition. One of his processes found an algorithm that dominated the results. Unfortunately, when he investigated, the "algorithm" was "write your process id over another entry in the list of winners"!
ReplyDelete