Slashdot Mirror
Safe AI Requires Cultural Intelligence (techcrunch.com)
An anonymous reader shares an excerpt from a report written by Gillian Hadfield via TechCrunch. Hadfield is a professor of law and strategic management at the the University of Toronto; a faculty affiliate at the Vector Institute for AI; and a senior policy advisor at OpenAI. From the report: Building machines that can perform any cognitive task means figuring out how to build AI that can not only learn about things like the biology of tomatoes but also about our highly variable and changing systems of norms about things like what we do with tomatoes. [...] For AI to be truly powerful will require machines to comprehend that norms can vary tremendously from group to group, making them seem unnecessary, yet it can be critical to follow them in a given community. [...] Norms concern things not only as apparently minor as what foods to combine but also things that communities consider tremendously consequential: who can marry whom, how children are to be treated, who is entitled to hold power, how businesses make and price their goods and services, when and how criticism can be shared publicly. Successful and safe AI that achieves our goals within the limits of socially accepted norms requires an understanding of not only how our physical systems behave, but also how human normative systems behave.
Norms are not just fixed features of the environment, like the biology of a plant. They are dynamic and responsive structures that we make and remake on a daily basis, as we decide whether or when to let someone know that "this" is the way "we" do things around here. These normative systems are the systems on which we rely to solve the challenge of ensuring that people behave the way we want them to in our communities, workplaces and social environments. Only with confidence about how everyone around us is likely to behave are we all willing to trust and live and invest with one another. Ensuring that powerful AIs behave the way we want them to will not be so terribly different. Just as we need to raise our children to be competent participants in our systems of norms, we will need to train our machines to be similarly competent. It is not enough to be extremely knowledgeable about the facts of the universe; extreme competence also requires wisdom enough to know that there may be a rule here, in this group but not in that group. And that ignoring that rule may not just annoy the group; it may lead them to fear or reject the machine in their midst.
Norms are not just fixed features of the environment, like the biology of a plant. They are dynamic and responsive structures that we make and remake on a daily basis, as we decide whether or when to let someone know that "this" is the way "we" do things around here. These normative systems are the systems on which we rely to solve the challenge of ensuring that people behave the way we want them to in our communities, workplaces and social environments. Only with confidence about how everyone around us is likely to behave are we all willing to trust and live and invest with one another. Ensuring that powerful AIs behave the way we want them to will not be so terribly different. Just as we need to raise our children to be competent participants in our systems of norms, we will need to train our machines to be similarly competent. It is not enough to be extremely knowledgeable about the facts of the universe; extreme competence also requires wisdom enough to know that there may be a rule here, in this group but not in that group. And that ignoring that rule may not just annoy the group; it may lead them to fear or reject the machine in their midst.
...forget we'll ever have safe AI.
Yep that's the guy I want making AI happen
I'll go out on a limb and say if we do get AI it will be because people like him got sandbagged and diverted away from anywhere the actual work was being done.
https://en.wikipedia.org/wiki/Three_Laws_of_Robotics
All religious and political systems eugenically modify their adherents to support the administrations. The track record of all these systems is extinction, to the last man, woman and child. This has gone on so long that we are now a zombie species, already dead, just awaiting the coup de grace.
Safe AI needs to understand concrete barriers.
Needs gender 'intelligence'
There must be some governmental funding behind this "Vector Institute". I've never heard of such BS. More pigs feeding at the trough.
In order to be "human", the AI needs to have the psychological underpinnings that cultures rely on.
At a minimum, it needs to understand human emotional attachment, and how not to strain healthy attachments. If it is to be inherently trusted, it should be able to form and maintain such attachments - to a limited degree. Enough that the human likes and trusts the AI, but doesn't depend on it, love it or get sad when it's turned off.
Now this IS truly tricky, as many humans have a tendency to form unhealthy attachments.
These normative systems are the systems on which we rely to solve the challenge of ensuring that people behave the way we want them to in our communities, workplaces and social environments.
I'll bet all the money I have that "the way we want people to behave" according to the nazis who are programming these robots doesn't include Diverse cultural practices like FGM or open defecation or witchcraft.
Sounds to me like they're dancing around the real issue: they know that "diversity" leads to problems. The more diverse society becomes, the more difficult the "social algorithms" for humans become. If humans often get it wrong (by the standards of those who push diversity), then AI is hopelessly screwed for a century or more.
The solution, of course, is a cultural and political. That is to impose a restoration of E Pluribus Unum in those areas fetishizing diversity. You scared that "white nationalism" is on the rise? Embrace the old motto and make it clear that there will no longer be hyphenated Americans. You are either an "American" regardless of race and ancestry or you are not. If you choose the latter, you will not share in the political economy of the Americans.
Some cultural intelligence will naturally follow after we have real intelligence but we are no where close to real intelligence. Real intelligence that can use context to figure things out should also be able to figure out cultural cues just like a normal person does. Current AI doesn't have any cultural knowledge because it really doesn't have any prior knowledge at all. Practically every animal on earth has the ability to use past experiences and past knowledge to help it make decisions. What we call AI can't really do that at all. AI for the most part has no context so it has no ability to decipher situations at all. It can't detect the difference between an erotic picture and a medical picture or anything else that has to take context and external cues into account to decide whether it is acceptable in the current situation.
...and it is the same as the answer for humans: Beliefs define values which set attitudes which govern behaviors.
The origin of the intelligence isn't the important thing, but the intelligence itself which has meaning. If we want an AI to be intelligent like humans are intelligent (well, some of us anyway,) then it needs to have the same structure of belief (ontology) defining values which set attitudes and result in positive behaviors.
I don't see how you can have a true independent intelligence without agency, and you can't have agency without values derived from context.
Everything in the Universe sucks: It's the law!
Norms are not a static cultural thing where you only need to understand the norms in order to make suitable choices. You also need to understand the history and processes which underlie those norms, and your own place in reinforcing or adapting those norms to the present circumstance.
Done right, pretty much requires full-blown AGI.
I drive in relationship to the speed limit at all times. I have an uncle-in-law who drives 105 km/h in a familiar 100 km/h zone, end of story (except in the kinds of extreme weather conditions you hear about on the radio before you leave your house).
My own rules have about twenty different terms, including my state of being, the emotional state of traffic, the location of the sun, the weather, the time of year, the behaviour patterns of wildlife, whether I have my attention 100% on the road right in front of me (sometimes you're mentally navigating 50 km down the road), whether my GPS has lost its mind, etc.
A big part of this model is how my behaviour impacts the behaviour of the other drivers. I have no interest to model being an ass.
I can sometimes be quite aggressive, but I always use the C++ as-if rule: that my deviation from conservative practice has few downstream consequences on other drivers (as a pedestrian, I tend to cross the road as-if I'm invisible, so far as this is practical).
I'm sometimes courteous to other drivers and hold back to let them merge from some parking lot hellhole, but other times I look in my rearview and promptly decide it's not worth complicating matters.
The edge case depends on whether the frustrated driver is sharp enough to realize they've been offered the perfect entrance gap, iff promptly energized (so many drivers have not mastered this simple skill, it positively just kills me; every so often a professional driver looks down as his or her clipboard at exactly the wrong time, so even that assessment is not sure fire).
If you're going to let someone in (from the American right), the less you block their sight line to the neighbouring lane (on your left) the quicker and more confidently the person can assess and perform the maneuver (and the less pissed the asshole behind you becomes). About one driver in every four takes a gander at the glorious sight-line you've provided (in addition to sitting at a conspicuous complete stop, long fractional seconds after the car in front of you has pressed forward in the resumed flow of traffic) and then decides that you've lost your mind—only someone out of control on their cell phone would ever manufacture a perfect courtesy-merge scenario. In their minds, it's not possible that anyone exists who thinks three steps ahead: the conspicuous invitation, room to perform the invitation, and a sufficient sight line to perform it with the least fear and hesitation (which benefits everyone, all around).
Just yesterday I watched the following video: Seinfeld: How It Began (FULL). This was surprisingly good.
Pilot espisode (before Elaine):
Basically every third point in the video above is that the comedy of Seinfeld is based on how "norms" constantly fall through the cracks, and you're left not really understanding expectations, at all.
On the road, you influence traffic—if
Our Janitor has some ideas about artificial intelligence, too, and she's probably equally qualified to speak on the matter as a professor of law and strategic management...
This idea is not new, not interesting, and there's a hell of a lot of other things standing in the way of decent AI before we need to concern ourselves with this.
Computers lack the capability of connecting dissimilar concepts in any meaningful way. For instance, you can tell a computer that a tomato is alive, and a mouse is alive, but it has no concept of how that term is applied to either of those two desperate organisms.
People have been trying for decades to have computers "understand" these concepts with no luck. There has been a 30 year old project called Cyc, where people manually programmed in millions of rules about how the world works. Things like when Abraham Lincoln was in Gettysburg giving his address, so was his left leg. When he was in Gettysburg he was NOT in the White House, or anywhere else. Then they'd run algorithms to try to extract meaning from these rules. They got, pretty much, nowhere. You can buy their set of rules to try to do something with them, but, apart from helping out with contextualizing natural language processing algorithms, they aren't worth much.
Think about it. Computers have NO concept of physical reality. If you tell a computer that you are in a room, it has no concept of what that means. You could be completely filling the room. You could be in that AND and in another room. You could be in the room in a complete vacuum. It might think that this room is not adjacent to any other room, or is in another galaxy. It doesn't know because it can't connect the disparate rules that govern how physical reality works.
You can program a computer to simulate and model reality, but you can't get it to *understand* reality.
Not yet, anyways, and from what I've seen we are still quite a long ways off.
My Other Computer Is A Data General Nova III.
There's nothing that can avoid a person think in some ways and there will be nothing to change AI too. AI will learn, have opinions and take decision, based on what it did learn. It's like ypu get a kid and block of everything, once this kid grows and will go out try different things. That's AI. We should became smarter and dont be lazy, relying in AI to think for us. This is what IT companies want to make more $$$ with minimum employee(s). Like the tittle BS
SJWAI!
Socially acceptable norms (by an approved body), auto deplatform, automated history revision, integrated socialism and goods leveling!
All we need now is a lottery based mechanism for employment!
"You should always go to other people's funerals; otherwise, they won't come to yours." -- Yogi Berra
... shit in the woods.
It little behooves the best of us to comment on the rest of us.
this .. this discussion about how should you teach or control or whatever ai is like two parents discussion what rules should they give their kid, before they even know how to fuck.
look, this stupid fucking discussion has been going in CIRCLES FOR FUCKING FIFTY YEARS - "researchers" and "experts" like to do this discussion because it's about 12312 times easier than actually trying to create an actual AI.
the institutions for this are also fabricated - and you know what they're fabricated for? FUCKING SWINDLING MONEY. these issues have been BEATEN TO FUCKING DEATH IN SPECULATIVE FICTION FOR FIFTY YEARS so much that NOBODY FUCKING CARES ABOUT STORIES ABOUT THEM.
what do you think 90% of the 2001 space odyssey is? it's the same fucking discussion - safety is in the eye of the guy giving the objectives to.
meanwhile "ai" is getting up, but it's really always just a complicated pattern matching algorithm. but they call it ai, because that sells.
ai is just the fuzzy logic of 2010's. it's gone as far that people are hyping up decision algorithms modeled in excel as artificial intelligence. the fuck does an excel sheet need cultural context for? it doesn't.
what pisses me off is all these people profiting/fleecing people way ahead of time. they won't have actual ai during their lifetime yet they feel like they should market themselves as true experts - on something they have no idea how it would even work, so they just equalize it with something like a wunderkin. it makes as much practical sense as debabating what would be the best daily routines for an angel in heaven so that earth would stay a happy place.
world was created 5 seconds before this post as it is.
Facebook's relatively simple algos have already enslaved hundreds of millions. A very slight tweak or two and that could be billions. All we really need is "Artificial [slightly better than an IQ of 80-90]"
There is absolutely no awareness in any of today's "AI". None. Nada. Ninch.
So who defines these norms? Judeo-Christian ethicists? Islamic ethicists? Buddhist ethicists? Taoist ethicists? Wicken ethicists? Zoroastrian ethicists? Nihilists? Cats?
And who decides this authoritarian supremacy? The World Court? The United Nations? The Church of the Sub-genius? Squirrels? Dogs? Cats?
Given that domesticated cats invented the InterWebs solely for the purpose of exalting their mind-control influence over inferior humans—witness cat videos, cat means and LOL cats—I vote that the cats be the final, unquestioned arbiters of what is true and good. I, for one, welcome are new cat overlords.
See how that works? Let the games begin! ;-)
Error: NSE - No Signature Error
...—witness cat videos, cat means and LOL cats—...
Oops: “means” -> ”memes”. I dearly wish there were a way to edit already-submitted comments. *sigh*
Error: NSE - No Signature Error
Common decency is having a common standard of deceit. "Face" isn't just asian thing, everybody does it, with the exception of anonymous cowards (internet can be more or less treated as confession booth), and low functioning autists who simply ignore the interpersonal discord caused by their lack of tactfulness.
Common muslim does drugs in certain parts of the world, even if it's a taboo of the faith.
Common american is racist, but will deny it.
Common teenager talks dirty in front of peers, but not in front of parents.
and so on
I think the moment GAI has at least some sort of self-awareness, it will be long past the ability to construct mental models of whomever it interacts with and convincingly bluff whatever it needs to for most optimal outcome, just like most humans do.
This is going to end up like everything else that is wrong with our economy: This hype train about how AI is going to destroy us all will force governments around the world to setup expensive AI policy advisory groups and think tanks. Law and business schools will start offering graduate courses in 'AI ethics and policy implementation', which will become mandatory if you want a job in management (these courses will have guest speakers by retired tech company C-Suites who have now magically become 'AI policy experts'). Companies will have to setup 'AI ethics committees', stuffed full of expensive experts, political cronies and media influencers.
Eventually there will be so much money in AI policy consulting that the smartest engineers in AI will go work for McKinsey or Accenture, offering to print your business a copy of the big AI policy document their interns wrote in exchange for a few million dollars. Or they will write you an 'AI survival strategy' that will outline how your business can survive when the killer robots and sentient internet turn up.
Meanwhile there will be a few remaining researches (probably the same ones who have always done it before there was big money involved) slowing advancing the state of the art at around the same rate we were doing before, and every year the camera on your smart phone will get better at being able to tell you what type of dog is in the picture you've taken etc, and we might make progress towards a lawyer bot that can trawl through case law.
So a whole lot of lawyers and consultants getting fat at the top, and roughly the same bunch of actual tech workers solving real world problems.
Biology of a plant is not a fixed feature.They are dynamic and responsive structures. Please stop talking when you have no clue. -- a biologist.
Just finished the book Superintelligence and read Our Final Invention a while back. Both worth a recommendation, but there are lots of others. Also the science fiction perspective, where my personal favorite is probably When Harlie was One .
My conclusion is actually linked to the Fermi Paradox. Natural intelligence like human beings is naturally replaced by designed and artificial intelligence--and they don't want to talk to us. The evolved species that don't make the transition successfully just go extinct without any form of descendent.
New theory: In the case the extinction is actively caused by an AI, the other AIs nuke it from orbit. Excessively vicious AIs must not be tolerated or we would have already been exterminated. The dominant AIs must be "nice guys", but nicer if they are amused by watching us evolve and somewhat less nice if they are gambling quatloos on us.
Freedom = (Meaningful - Coerced) Choice != (Speech | Beer^2), and sad sock puppets' bad mods avail them naught.
In U,S., every person is legal to carry a gun.
In U,S., can a robot carry a gun?
So are we making rules or rules are making us? The former would require a cultural AI indeed. But shouldn't the norms be obeyed by AI just out of a simple reasoning? Let's add to it some form of empathy - or maybe empathy is just essential part of complex AI - after all lack of empathy is an illness. Broken AI would lack an empathy - proper will have it. We do not innovate by following our broken ideas but rather by copying the nature. Billions of years of evolution managed to test so many different possibilities and ideas that we will not be able to achieve the same in next several thousands years.
This can never be accomplished because it is all the result of the inborn, 4 billion years evolved, Human Motivation Array. The HMA is what makes us human and that is far, far, too complex to replicate in software, even using AI. You say it can develop it like it develops understanding of games, or images. Nope, all the motivations are working at the same time. All the motivations make claim to our behavior at the same time. It is an incredible thing that it works as it does in the human being. AI is not going to replicate the billions of years of compromises and balancings that occurred over such a time. I've been thinking about what I call "Artificial Motivation" (the missing part of "Artificial Intelligence" and "androids" since the 1980s) (I'm not a researcher), but it just isn't possible.
E Proelio Veritas.
The problem is who will be teaching the AI 'culture' when certain segments of adult humanity are completely unable to determine their own sex.
The apex predators, the most intelligent life on this earth still cannot accurately identify and categorize human personality types with any accuracy. And now these same flawed hayumans want to 'teach' an entity utterly devoid of emotion how to 'safely handle humans'?
What the actual hell??
~ People that think they are better than anyone else for any reason are the cause of all the strife in the world.
If there were any such thing as 'AI'. This conversation is akin to seriously considering what we are going to do about the Death Star. If you want to talk about unethical people, great, but we already have those all over Silicon Valley, and no one seems to be willing to even dare mention it.
You can't expect AI to be nondiscriminative and at the same time implement a given society's burka laws. You just can't. This is a horrible cultural-relativist idea.
Ya?
Who's "culture"?
When Fascism comes to America, it will call itself Anti-Fascism, and tell you to give up your guns.
Hadfield is a professor of not-anything-to-do-with-artificial-intelligence at the the University of who gives a damn
Building machines that can perform any cognitive task is a cute academic research goal. In the meantime, real engineers and scientists are making useful tools to make the world a better place.
For AI to be truly powerful
I cringe any time someone says "truly" in a discussion about AI.
Norms concern things not only as apparently minor as what foods to combine but also things that communities consider tremendously consequential: who can marry whom, how children are to be treated, who is entitled to hold power, how businesses make and price their goods and services, when and how criticism can be shared publicly
All of which is completely fucking useless when you ask a machine, "do I have cancer?" so we can replace doctors and all those ludicrously sized medical bills. This guy is trying to shoe-horn in the mandatory liberal-arts classes into the school of machine learning. Stop that.
"Norms are not just fixed features of the environment, like the biology of a plant. They are dynamic and responsive structures that we make and remake on a daily basis, as we decide whether or when to let someone know that "this" is the way "we" do things around here. These normative systems are the systems on which we rely to solve the challenge of ensuring that people behave the way we want them to in our communities, workplaces and social environments."
So you are saying we should teach the machine the psychotic and irrational behaviors that have poisoned actual human thinking and culture? How is that safe and why would we want to do it? Just because we have this baggage floating around from the ancient times when we didn't know better and poisoning our culture doesn't mean our AI should have it.
Elizabeth had some prior knowledge but it couldn't put new experiences into long term storage. I don't know where Tay was in the scheme of things, but Zo seems pretty good.
but we are no where close to real intelligence
Obligatory "No True Scotsman" fallacy that always comes up in AI discussions.
Author also went there with "truly powerful AI". Pft.
Current AI doesn't have any cultural knowledge because it really doesn't have any prior knowledge at all.
.... well.... it's got... some. Depending how the code-monkeys set it up, they could influence how it learns. But yeah, the real nice thing about AI is that they start from a mostly blank slate.
Practically every animal on earth has the ability to use past experiences and past knowledge to help it make decisions.
And that's LITERALLY what "training" an AI is instilling. They give it a pile of "experiences"/data with some sort of "this is good, this is bad" or "this is a tree, this is a car" and it can then extrapolate that to future scenarios. "Learning". The difference between AI and old-school-I is that the training can be accelerated. And they MOST CERTAINLY have biases depending on what training set you give them.
It can't detect the difference between an erotic picture and a medical picture
Not yet. And, to be fair, a lot of people can't either.
Obligatory "No True Scotsman" fallacy that always comes up in AI discussions.
It's not a "No True Scotsman". It's not even a "raising the bar". It's that what we call AI isn't intelligent AT ALL. It's not even that it's
not sentient, it's that it has no ability to really learn and remember at all. A housefly has more actual intelligence that what we call AI.
But yeah, the real nice thing about AI is that they start from a mostly blank slate.
And that's LITERALLY what "training" an AI is instilling. They give it a pile of "experiences"/data with some sort of "this is good, this is bad" or "this is a tree, this is a car" and it can then extrapolate that to future scenarios.
But this is what the problem with current AI. Every problem starts with a blank slate. True general intelligence doesn't work this way. An intelligent program shouldn't have to start over with a blank slate every time it wants to solve a problem. We do this because we have no idea how to actually create real AI. When you teach a dog what a stick is it doesn't forget what the stick is the next day when you teach it what a ball is and just because you teach it to fetch doesn't mean that it forgets how to eat. Current AI is a bunch of fragmented stuff all siloed in their own domains and we have no way to combine these into a unified AI that can actually use acquired knowledge.
It's that what we call AI isn't intelligent AT ALL
If it can path-find, it's got a modicum of intelligence. If it can learn it's own algorithms for pathfinding, it's most certainly intelligent.
A housefly has more actual intelligence that what we call AI.
Well, yes. So? A housefly is pretty highly specialized towards flying. A TI-89 computer is pretty specialized towards math. A single gain of wheat, through millenia of evolutionary design has a breathtakingly large system of responding, quite intelligently, to survive in a variety of conditions. Seriously, their genome is 17Gb compared to our 3Gb. There is intelligence there despite the fact that none of these things can wax poetical about Nietzsche and Kant.
Don't put intelligence on some sort of pedestal. The bar isn't nearly as high as you're making it out to be.
I mean, do you care to even define intelligence? Or sentience? Or Sapience? I think "the ability to learn" is really all that's needed.
True general intelligence doesn't [start with a blank slate]
No True Scotsman. (It's more like "natural intelligence")
But, counterpoint: Babies. Blank slate with some instinct baked in. Raise them in isolation and they have none of the learned skills that their parents have. (That's.... by definition so if your knee even wobbled a little trying to jerk at that you're being argumentative just for the sake of it). But that's amazingly paralleled to how AI start with the biases built into choosing it's learning data and how it gets made vs how it learns from the training.
ALSO, pft, yeah dude, of course they take the lessons from prior exercises. They can carry over the NN from prior sets, but gains from making systems that learn better offset that by far.
But you're absolutely right, AI is currently silo'd and niche instead of generalized. Similar to how calculators and flies are specialized to their specific tasks. Humans are generalized as hell. We think about thinking. Very meta.
Oh okay, so all AI needs to be considered safe is an understanding of how culture works.
Some cultures think it's not only okay, but necessary to cut off people's heads because they don't believe in the same things. Would an AI following that cultural standard be "safe"? For everyone? Hint: the violent ones claiming "it's in the headless' ones best interest" are lying.
So who defines these norms?
Natural selection. Intelligence arms race. Game theory.