When I was a kid I read Asimov's laws and it inspired me to work on robots and machine learning. However, I quickly realized: the computers of the time (Apple IIe) could not implement those laws; they required the ability to parse human text, and more importantly, my read of the laws and how to comply requires the robot to create an accurate predictive utilitarian model of its actions and their effects. And be able to continuously re-evaluate that model and choose actions consistent with the laws.
After looking more into utilitarian models, I concluded there was no way to implement Asimov's laws unless you had a superintelligent AI. But in retrospect, I think I was interpreting the laws in the 'hard' form.
There are a lot of assumptions in those laws, and they were exploited by Asimov himself in his short stories and books. Redefining what is human, adding a zeroth law, weakening laws, lack of understanding of human nature (i.e. lending an arm to a dangerously angry woman), even with those theoretically perfect laws it was still able to be exploited and cause harm.
Even the Men on Fire episode of Black Mirror was able to trick humans, with laws or not.
> His laws assume a machine that reasons from rules. Modern AI learns patterns from data and approximates behavior. This means the LLM driven Asimov law will never be an unbendable law to follow. Instead, it's merely a suggestion.
And in the human world, laws that are broken and not charged, nor prosecuted, are merely suggestions.
Which is why we have
(1) files illegally withheld and
(2) wars without authorization or even rationale other than (1)
When I was a kid I read Asimov's laws and it inspired me to work on robots and machine learning. However, I quickly realized: the computers of the time (Apple IIe) could not implement those laws; they required the ability to parse human text, and more importantly, my read of the laws and how to comply requires the robot to create an accurate predictive utilitarian model of its actions and their effects. And be able to continuously re-evaluate that model and choose actions consistent with the laws.
After looking more into utilitarian models, I concluded there was no way to implement Asimov's laws unless you had a superintelligent AI. But in retrospect, I think I was interpreting the laws in the 'hard' form.
There are a lot of assumptions in those laws, and they were exploited by Asimov himself in his short stories and books. Redefining what is human, adding a zeroth law, weakening laws, lack of understanding of human nature (i.e. lending an arm to a dangerously angry woman), even with those theoretically perfect laws it was still able to be exploited and cause harm.
Even the Men on Fire episode of Black Mirror was able to trick humans, with laws or not.
> His laws assume a machine that reasons from rules. Modern AI learns patterns from data and approximates behavior. This means the LLM driven Asimov law will never be an unbendable law to follow. Instead, it's merely a suggestion.
And in the human world, laws that are broken and not charged, nor prosecuted, are merely suggestions.
Which is why we have
(1) files illegally withheld and
(2) wars without authorization or even rationale other than (1)