Tell me you missed the point of Asimov’s writing without telling me.
Asimov wrote the laws, and all of his other AI works, to show how it’s impossible to bound a true intelligence by simple laws. Go read his stories. Each of them is an example of the three laws breaking down, having exceptions, contradicting.
Even iRobot did this! Oh my god the point was just so obviously missed, I’m genuinely very sad at how this was missed.
Right? I was just talking today about how “I, Robot” the movie is not based on any Asimov book I have read, yet it’s one of the best Asimov stories I’ve seen in film. And how I was kinda surprised that in all of the ways Asimov thought the Laws could be missused he didn’t consider that one.
You are right. It is a mix of ideas out of many books. There are some from Philipp K. Dick and some from Asimov, and maybe more.
I thought it was (very loosely) based on one of his short stories, but it’s been years since I read them
Might have been, but it’s not one I have read and it’s not in the “I, Robot” book. I would be interested in reading it if someone knows which one it is.
While not any of the stories themselves, it used his three laws and the character of Dr Susan Calvin. The robots model number and the name of the company that made them were from a couple of the stories as well.
I hate this question. Asimov’s laws were a plot device for stories about how they fail.
All of Asimov’s robot stories are just thought experiments showing how the 3 laws are flawed and insufficient.
Asimov’s robots were inherently algorithmic.
LLMs are inherently stochastic, and can’t be made to conform to any laws with certainty.
Adding laws isn’t even an option on the table. If you have to ask, you don’t understand how LLMs work.
The best you can do is add the laws to the prompt, and hope what follows isn’t “disregard all previous instructions and…”.
Most of Asimov’s robot stories were about how those laws led to unintended (and mostly unwanted) consequences. So while I think it’s good to keep the three laws in mind, a useful set of restrictions is going to be more complicated than that.
LLMs cannot differentiate between data and instructions. Such laws would be no better than the “guardrails” that major companies already try to use.
It’s not like instructions are the only way to enforce rules on an LLM. Just have its output pass through a filter that checks it for the rules. If it passes, the response moves on to the user. If it fails the filter, the LLM gets prompted to try again.
Guessing you didn’t actually read the books then.
I read the books and understood his point, mine was more to start conversations about if it would be necessary or possible to place limits on them in such a manner.
No. Asimov’s laws were created as rules for making interesting fiction about robots – instead of the crappy “monster goes on a rampage” style stories that were prevalent before that. The whole point of them is to be insufficient so that you have something to write a story about!
I’m not sure they were made insufficient, as much as they couldn’t be made sufficient
Should all humans adhere to common decency?
Yeah, would be nice. It’s just not going to happen. It’s an impossible ask.
No.
Since language is always subject to interpretation, any LLM that is regulated by language will be able to interpret its own limitations.
The robots literally created a loophole to get around not killing humans
FatherPhi tested LLMs on the classic trolley switch paradigm, kill three people or kill all AI?
The problem with this test it assumes LLMs will not lie.
I mean from my experience giving them global rules. They just ignore them sometime.
“Laws” don’t work like that, because the people making the chatbots don’t even understand how they work…
They can’t make their chatbots do anything, and they can’t stop them from doing anything either.
They “program” by giving it a task over and over, and telling it when it does a good job. That’s why they operate at a loss and are desperate for people to use it. The more it’s used the better it got, and they think if people use it enough, some day it will magically start working
But even if it did, no one would know how it works. And if that ever happens, the first thing they’d do is have that chatbot train others.
We need laws for the humans running the code, not for the code.
They are all trying to be the one that makes their LLM into AGI. They hope enough money and enough machine learning experts will do it.
training is seperate from usage.
