I always thought that Asimov's Laws of Robotics ("A robot may not injure a human being" etc) were an interesting prop for science fiction, but wildly disconnected from the way computing & robotics actually work.
Turns out he was just writing LLM prompts way ahead of his time.
Not only wildly disconnected, but purposefully created to show ambiguity of rules when interpreted by beings without empathy. All of Asimov's books that include the laws also include them being unintentionally broken through some edge-case.
It was weird to actually read I, Robot and discover that the entire book is a collection of short stories about those laws going wrong. Far as I know, Asimov never actually told a story where those laws were a good thing.
They aren't generally potrayed as bad, either, just as things which are not as simple as they first appear. Even in the story where the AIs basically run the economy and some humans figure out that they are surreptitiously suppressing opposition to this arrangement (with the hypothesized emergent zeroth law of not allowing humanity to come to harm), Asimov doesn't really seem to believe that this is entirely a bad thing.
Just sit right back and you'll hear a tale, A tale of a fateful trip
That started from this tropic port. Aboard this tiny ship. They got lost but they called for help, and now they're totally fine. And now they're totally fine.
You are suggesting today's state of affairs to remain the same after people do the pause the world literature survey. This is exactly opposite to my point of view. At least they will be well informed as to why their views are different.
Well it's quite difficult to come up with much better rules than Asimov's.
HPMOR offers a solution called 'coherent extrapolated volition' – ordering the super intelligent machine to not obey the stated rules to the letter, but to act in the spirit of the rules instead. Figure out what the authors of the rules would have wished for, even though they failed to put it in writing.
> Figure out what the authors of the rules would have wished for
What if the original author was from long ago and doesn't share modern sensibilities? Of course you can compensate when formulating them to some extent, but I imagine there will always be potential issues.
Exactly! That was kind of the point IMO, that human morality was deeply complex and ‘the right thing’ couldn’t be expressed with some trite high level directives.
All of fiction is a distortion of sorts. Consider Wall-E movie fat people. The AI advancements shown in the movie should transitively imply that biotech, biomedical progress would be so high that we would have solved perfect health by then.
Which is because we don't have intelligence on tap. If we did, we would actually put intelligence to use in all the subjects. I would rather put intelligence to use on biology rather than philosphy.
More just that the rules are actually a summary of a very complex set of behaviours, and that those behaviours can interact with each other and unusual situations in unexpected ways.
It's funny because Isaac Asimov would have come up with some convoluted logical puzzle to justify why the robot went on a murderous rampage - because in sci-fi robots and AI are all hyperrational and perfectly rational - when in real life you'd just have to explain that your dying grandmother's last wish was to kill all the humans, because a real AI is essentially a dementia-riddled child created from the Lovecraftian pool of chaos and madness that is the internet.
I recall that story of the guy who tried to use AI to recreate his dead friend as a microwave and it tried to kill him[0].
You couldn't sell a sci-fi story where AIs just randomly go insane sometimes and everyone just accepts it as a cost of doing business, and because "humans are worse," but that's essentially reality. At least not as anything but a dark satire that people would accuse of being a bit much.
Worth noting is that the article is from April 2022 and used gpt-3. The "friend" was an imaginary friend, not a dead friend, and so probably more prone to taking actions which would appear in a fictional context. From my research it looks like the base gpt-3 model was just a text predictor without any RLHF or training to be helpful/harmless.
Certainly AI safety isn't perfect, but if you're going to criticize it at least criticize the AIs people actually use today. It's like arguing cars are unsafe and pointing to an old model without seatbelts.
It's not surprising at all that people are willing to use AIs even if they give dangerous answers sometimes, because they are useful. Surely they're less dangerous than cars or power tools or guns, and all of those have legitimate uses which make them worth the risk (depending on your risk tolerance.)
Fair enough but it's still wild that the template for AI and robots in science fiction has (usually) been to portray them as hyper-rational, competent and logical even to a fault, but the most accurate prediction of what AI and robots turn out to be is probably Star Wars, not anything by Asimov.
I use CNC machines and know how powerful stepper and servo motors are. You can ask yourself what will happen if your motor driver is controlled by an AI hallucination...
Turns out he was just writing LLM prompts way ahead of his time.