So a sort-of-on-topic question: for all the people who are warning of the coming AI apocalypse and warning us to be careful with AI... how do you enforce that? What’s the enforcement mechanism to ensure that the bad people don’t get AI and that people don’t create bad AI? All these warnings, but what can we do about it?
If strong AI is invented, especially if it’s open-sourced, the cat is out of the bag and it’s game over, right? We’ve seen how good we are at preventing rogue states from getting nukes. How can we keep them from getting life-ending AI?
We already created super AI. It's called the stock market. Look at how it's tricked the richest(capitalist) and smartest via engineering jobs into toiling away at getting a number to go up.
Yeah, that's more of my point, you're right. If we're considering the endgame of the AI singularity, only one AI can win. If there are multiple, they'll each keep getting better, and they'll each keep warring against the other one until only one is left (or none are left). That's the doomsday scenario for AI when people warn that this is coming.
But out of all the people warning about it, I haven't yet seen anyone say "this is how we prevent it".
Seems like the only prevention proposed is abstinence and research shows that abstinence doesn't work. In fact telling most people not to do something makes them want to do it more. In fact I'm struggling right now to make my own attempts.
I think the only way to win is to make sure the first one is perfectly friendly, or to make sure we are integrated with the first one. I don't give us great odds on either front.
I'm trying to remember a book I read a while back, it was a sci-fi book where in the beginning chapters a war was being fought between the US and China to prevent China from developing a strong AI. The main character is killed but the preserved his body and brought him back to life some years later in a cyborg body, where he learns that the world is entirely controlled by a friendly benevolent AI and everyone has it embedded in their skin. But the AI goes rogue suddenly and the only people who survive are the luddites who refused to join the AI and the people who were off-world at the time the AI went rogue.
For the life of me I can't find the book even through months of Google searching, but anyway my point is... a "friendly" AI can be hijacked. Or decide to go rogue, since we can't begin to comprehend its thought processes.
--edit: sorry I went off topic a bit... I don't want to be just another one of those "AI will kill us all" kinds of people because, as my first comment suggested, that's completely useless fearmongering unless someone has a way to stop it. Kind of like saying "the supervolcano" or "solar flares" or something else completely unpreventable will kill us all... unless there's something you can do to stop it, life goes on until it doesn't. I'd be interested in hearing how to stop it, otherwise I wish people would stop printing these articles.
I follow the point you're making. We do have options and perhaps if we were worried enough, they'd be put into effect.
You could make AI research illegal. You could place massive R&D to human/brain interfaces. You could remove or reduce the medical device testing restrictions on the basis that this is an all-or-nothing bet for humanity.
It's clear to me that if AI exists it will be used irresponsibly and by bad actors. So the only way to be safe is prevent all research along those lines for as long as possible†.
The focus for now should be on getting humanity to work together and getting a firm grasp on ethics/morality to see how AI could be convinced to be good, instead of enforced.
There is none. Science is discovering what is possible. Technology is changing something from possible to doable. Once the possibility of something is made clear (if not the exact path toward it - the do-ability), people will start rushing for it - if only out of fear that someone else might have it while they themselves do not.
At least nuclear technology has the downside of being very expensive and large-scale (so few players) to produce, and being pretty obvious when it's actually used. Other things that are becoming increasingly possible (but still not doable) - like modified viruses that only target people with a specific DNA profile - will be a lot easier to hide, both their production and their usage. Same thing with AI.
The Machine Intelligence Research Institute is a research nonprofit studying the mathematical underpinnings of intelligent behavior. Our mission is to develop formal tools for the clean design and analysis of general-purpose AI systems, with the intent of making such systems safer and more reliable when they are developed.
this is all fine, first generation of 'true' ai will probably be extremely biased and too simple for practical use. even though we are automating and digitizing more and more and making everything more streamlined the rules will be as rigid as they have ever been. the human factor will come from generations of ai iterations -- if we are lucky. to be rejected by a machine will be brutal experiences when we all get this pushed in our faces. rich will grow richer and middle class will disappear in favor of poverty across the board.
>first generation of 'true' ai will probably be ... too simple for practical use
AI tends to either not really work or go way beyond humans - look at the Go programs that didn't work very well and then in a short period we have things like alphazero that given the rules of Go or Chess surpasses all humans in a few hours of playing itself. By the time they are passing Turing tests and thinking in a human manner they are going to be scary smart in many ways.
While its easy to dismiss the Sci-fi fantastical end of the world claims on the dangers of AI, its important to at least address some of the realistic externalities of this technology.
"Brin is showing more real-world concerns of AI-powered systems replacing human jobs or being used to spread propaganda and fake news rather than rise up and enslave humanity"
At least there isn't anyone out there designing and manufacturing military grade weaponised drones and letting them autonomously decide whether to pull the trigger.
Oh, I get it. So Google's AI only "advises" the Pentagon who its targets should be, and then the Pentagon takes the huge engineering role of implementing "if Google advice = true, then launch missiles."
This way Google is completely "guilt-free" and can still pretend it's "making the world a better place" wink wink nudge nudge.
I get your joke, but the parent comment means "if the advice is from Google, then launch missiles", not "if it is a correct advice from Google, then launch missiles", so the assignment vs comparison question doesn't apply.
Not very good click bait as it’s clear from the headline the HN comments will be better than the article, so I usually skip the article and read the comments. :)
I'd be interested in reading coverage of Brin's comments without having to wade through the smart alec dickhead style of The Inquirer (is it a sister publication to TheRegister?)
Edit: direct link to the letter posted elsewhere on this page.
Not to worry just yet as AI still hasn't been realized.
What many developers call AI, or more precisely machine learning (ML), are a collection of algorithms for making more intelligent decisions in context to the data at hand. That is simply smarter programming, but it isn't intelligence. The machine is still limited to the static instructions available.
Real AI will allow machines to spawn original decisions to solve new problems spontaneously. We aren't there yet. The ethical implications will likely not be any different than those confronted by people. It generally boils down to just because you can do something doesn't mean you should.
Real AI will never exist. No matter what definition we have ever set up as being “real AI” we have excused it and changed it once computers actually solved it. Current AI is capable of creating “original” decision as is it capable of solving new problems “spontaneously”, yet you still will not call it AI. Not because it doesn’t live up to your defined standards, but because you want to define a set of standards that it doesn’t live up to. People get protective of the concept of intelligence and almost axiomatically define that anything a computer can do is not truely intelligent and therefore not “really” AI.
Current AI can do (very, very, very advanced) pattern recognition and mimicry. Turns out this is enough for a surprisingly large set of problems, but it isn't making "original" decisions or spontaneously solving new problems.
I'm less certain about your claim. How did human consciousness and self-awareness come about? I can easily imagine some hardware that creates a self-aware entity and then that creation begins self-improving at a geometric rate. It would be an extremely sudden stop for all life on earth.
What you describe is a general aritificial intelligence[0]. From what I can tell, the industry is ridiculously far away from that - it's not even certain whether general intellgence "smarter" than humans can be achieved with computers at all.
Nonetheless, the media likes to generate some hype and fear every now and then, and has done so for decades, usually without even a basic understanding of underlying technology.
At least that link has a fairly clear definition for AGI as "successfully perform any intellectual task that a human being can." Estimates seem to be in the 10-100 year range. It may not be all that long. https://medium.com/ai-revolution/when-will-the-first-machine...
I read a book recently called "Are We Smart Enough to Know How Smart Animals Are?"[1] Very fascinating thought that, in my mind, redefines what I think of as "intelligence". The book Superintelligence[2] gets a little bit into this as well.
Basically, if we define "smarter than humans" as "able to do everything humans do, in the manner humans do them, but faster", then we have no idea what intelligence actually means. AI will be another species (not just a better human) and its intelligence would have to be judged based on its own merits.
Perfectly modeling the human brain in computer form is one avenue that computer science is working on. Artificial intelligence is a completely different path, and does not have to model human intelligence in any way.
I think if the goal is to be smarter than anything else the goal is already hopelessly failed to a bias. The goal is to achieve originality first. Smarter is something that can be qualified later.
> It generally boils down to just because you can do something doesn't mean you should.
Funny you say that... I have asked the same question regarding why some people are obsessed with pushing for technological singularity and AI that can replace human decision making.
I imagine AI would be helpful in research and business. In research people generally struggle with the creative issues most, those critical decisions at the edge of their work. This decision making is a tiny fraction of the work at hand, but is mentally and emotionally stressful sometimes demanding of deep focus and periods of concentration.
That sort of creative thinking is stressful because it is often an exercise in planning or problem solving that deals with the relationships of many moving pieces simultaneously and results in qualifying or validating activities that demand large amounts of time or resources. I often wish I were better at writing software so that I could write automation routines to more quickly validate my decision making, and so I typically end up writing validation routines instead of solving for the core problem. It would be nice if there were an AI available that help limit that distraction.
You don't need AGI to create and disseminate effective propaganda for example. Current AI techniques, (ML etc), are readily applicable to the task. I think Brin's point is that these currently available AI techniques are only going to get better. Doesn't really matter if we won't get AGI for another 100 years or whatever.
His point is we know how to keep a large part of the global population constantly touching their devices like well programmed guinea pigs the whole day, negatively effecting themselves and society in uncountable ways yet postively effecting our share price. Please protect us to exit the loop or we get lynched one way of the other by customers or shareholders!
Normally I agree that people don't know what they are talking about when they are talking about AI. But this is Sergei Brin, he knows what he's talking about.
Why? What I gathered from reading the article is that superior hardware is around the corner. I suspect the unstated conclusion in that is that faster computation will enable the rise of AI. That very well may be, but you still need to create AI regardless.
True, but the real danger is in governments believing that smarter-than-human AI already exists, and going full-bore on implementing it everywhere. And of course the companies selling "AI services" have all the incentive to aid the government in thinking that, too, consequences be damned.
I believe that "real AI" (unsupervised general AI) will not be created by us. Instead if we arrive there, I think it be more by a loop something like:
0. goodness criteria = <Humans INSERT MOTIVE HERE>
1. we create dumb AI.
2. we create a bit less dumb AI.
3. for(some years) goto 1.
4. we figure out how to make the bit less dumb AI, create another dumb AI on its own, and assess the goodness of it.
5. run latest generated dumb AI to create a new AI based on goodness of the previous one.
6. if (goodness criteria not met) goto 5.
So the whole process needs "jump-started" kind of like a jet-engine.
You end up with an AI system, tuned by another AI, in order to meet human supplied goodness criteria. And in the end, nobody actually understands the fundamentals of why it works. But hey, it does. Kind of. Most if not all I have read regarding these systems, describes the what, not the why.
Not sure what would happen if you use another AI to supply the motives (goodness criteria). I'm sure someone somewhere is trying :)
In this situation, I do wonder if the AI's (there will be many) will determine the best path is to fight for dominance or negotiate peace with other AI's.
> Not to worry just yet as AI still hasn't been realized.
AI has been realized. We have tons of AI in every facet of industry from the car to the food to gaming industry. AI is everywhere.
> Real AI will allow machines to spawn original decisions to solve new problems spontaneously.
AI already does it to some degree. Deepmind's AI has made original decisions and solved new problems in chess and go. Deepmind's AI isn't domain specific.
What you are talking about is generalized "conscious" AI. That is like fusion energy. A truly revolutionary step that we don't know if we'll ever achieve.
The current dangers of AI isn't the generalized AI since generalized AI would ostensibly be "godlike" to us. What is worrisome is the pace of advancement of regular "non-conscious" AI.
Who is to say that something like AlphaGo isn't conscious? It even exhibits a quality similar to panic when it finds itself in a hopeless situation. It is very anthropocentric to believe that something must act like a human being in order to be conscious.
It is not a given that any AGI would be a treat. There are plenty of Natural General Intelligences, and I don't fear monkeys taking over the world, or creating faster, better, stronger monkeys any day soon
A lion is still dangerous even tough it's not conscious.
Edited for clarification. Not conscious in any way that would rival "real AI" (as per the parents point) or human consciousness.
AI will get increasing access to more and more crucial parts of society and thus become more and more dangerous. It doesn't need to "real AI" for that.
Snakes can be dangerous, ants and spiders can be. They don't need to be "woken up". Instead it's a consequence of what they are optimized for that makes them dangerous.
> A lion is still dangerous even tough it's not conscious.
Of course, because it can wake up, right? Or are you trying to say Lions are not intelligent? If the latter I think you are wrong. I am fairly certain lions can acquire new skills and apply that knowledge in many ways. Like knowing that whoever put them down is a threat and they should probably attack them when they wake up.
I am responding to the parent who talk about "real ai" what I am saying is that you don't need "real ai" for something to be a threat. with real ai I am assuming the parent means conscious self aware AI.
No lions most certainly can't acquire new skills in any meaningful way besides trough evolution.
Sure but in this context that's hardly relevant. We are comparing it with "real ai" as the parent calls it or GAI which is something on a much much different level.
I was responding to the parent who talked about real ai not sure why eyebrows are being raised unless someone want to misunderstand the context the comment were made in.
Of course a lion is conscious but it's not conscious at a level that the parent means (real ai).
While this particular essay starts off with the AGI premise, this could easily be reworked to assume a relatively "dumb" by highly connected AI. For example, one that is able to buy/sell stocks, place orders from suppliers, sell product and adjust prices, and rebalance workforce and automation lines could still cause some major damage before anyone would notice.
We, the People are under the money pressure from that last 7000 years, this slavery mode of Society was a type of model that bring us, after many trouble to this point of our evolution.
Now that the robotics/cobotics/mechatronics sciences are on the point to reach and join with Artificial Intelligence we, the People are close to be free from the money Society model.
And that will be great, when 100% of goods and services will be product by the machine, that will be good and great enough to offer us the way to type 1 in the Kardashev Scale, we are at the time we will know an evolution type in our civilization.
UBI is not too different from a lot of the social programs (Social security, Medicare, etc) that grew out of the turmoil of that transition, though. Just a difference of degree, IMO.
Also, a jobs guarantee could do many of the same things (and similar to some of FDR’s programs). We might do both.
Maybe we should start with solving OCR, before taking on artificial intelligence. OK, we have self driving cars, but they can't actually read road signs.
The idea that AI will affect employment is very dubious to me:
- This stuff was talked about all the time during the industrial revolution. But it turns out that automation just increased worked productivity and made people more valuable.
- Since ATMs were introduced, the number of bank tellers in the US actually went up.
- Certain industries have actually become deautomated over time. As an example: coffee. It used to be exclusively made by automated machines. Now hand-making coffees is one of the largest businesses in America.
- AI is only as smart as the data you feed it. Even Google's own search rankings have to have human minders because people are too good at gaming their algorithms.
- AI is only good at the thing that human beings are very bad at - making quick, analytics based decisions. Of course this is a threat to middle management and executive functions. But for people who actually make things or solve problems, there's just one more thing to make and solve problems about.
There are some very real concerns about organizations using AI to make decisions about people without human oversight. Imagine an AI tasked with college admissions, bad decisions without possibility of human intervention could hurt people - are we sure it would be that much worse than our current system, often tainted by bias, money, and racism?
My concern is that the leaders in the corporate tech world created huge organizations by automating our lives, and now they want to act as gatekeepers on the kinds of automation that is acceptable.
Edit:
To add to this - a companies stance on AI will probably follow their belief in whether humans are underrated or overrated. Google designs products believing humans are overrated - they think people will be victims of AI. Amazon thinks people are underrated (based on Amazon Turk, customer service, warehouse design) - ironically, they won't think humans will overall be harmed by AI.
>- Since ATMs were introduced, the number of bank tellers in the US actually went up.
Counterpoint[0]:
> But US bank teller employment actually rose modestly from 500,000 to approximately 550,000 over the 30-year period from 1980 to 2010 (although given the growth in the labor force in this time interval, these numbers do imply that bank tellers declined as a share of overall US employment).
Incremental automation of any specific task isn't bad, especially if it can be introduced at a rate that's in line with the normal turnover/attrition for a job.
The doom-and-gloom scenario is when extremely adaptable and productive automation is deployed across many industries without time for the labor market to adapt, causing the world's population of 5-6 billion impoverished people to grow to 6-7 billion. Advancements in defense technology make it impossible for the masses to revolt and seize resources to revert to a pre-automation economy among themselves, so they die off.
A few hundred million shareholders and financiers enjoy their abundant luxury goods for a while, but they soon turn their robot armies against each other and also die off.
The argument that AI will take all the job is commonly made, but I agree with you that it's a poor argument.
Much more plausible and concerning, however, is that AI will stratify the labor force into 3 groups: highly-paid engineers (who manage the AI), highly-paid services job (i.e. doctors, lawyers, psychiatrist), and poorly-paid services job (i.e. barista).
This stratification has already happened as our economy advances (you mentioned baristas---in the past, those workers could have earned middle class manufacturing wages). The wage distribution in the services sector nowadays looks like a dumbbell [1].
Wage stratification can lead to political stratification as poor people don't have the resources to fight for their rights. Political stratification can lead to further wage stratification, continuing the cycle. That's why I believe that it's important to deal with the AI threat, now.
I agree that wage stratification has already occurred, but I think AI will do nothing to make it worse than tech already did. It will do nothing to replace service workers or laborers. Again, it's only a threat to managers and decision makers.
Many of the middle income jobs in the U.S seem vulnerable to automation. Professions like truck drivers, longshoreman and radiology techs that provide compensation between 40K and 70K could probably be mostly or completely automated in less than the span of a generation.
That seems like an arbitrary stratification. By the time AI is sophisticated enough to have significant ripple-effects throughout the economy, the "highly paid engineer" is going to be a tiny fraction of the labor force, akin to the hedge fund manager of today. Also, why do I need to hire a team of engineers when I can just spin up a few AWS AI instances.
That's entirely possible. I'm thinking more about the labor force of the next 40-50 years. In 100 years (or however long it takes for AI to improve itself), engineers may also be obsolete.
Kind of like Marc Andreessen's "The spread of computers and the internet will put jobs in two categories: people who tell computers what to do, and people who are told by computers what to do"
I disagree about stratification to a certain extent, that is, I only see 2 groups emerging. Here's my thinking.
Highly paid service jobs will remain unscathed only in so far as they have regulatory barriers to entry i.e. someone has to have gone to medical school, be licensed by a community of medical professionals, etc, in order to practice medicine. This applies to lawyers, doctors, psychiatrist, teachers, professors, architects but notably not to software engineers, carpenters, plumbers, etc (note: this split depends on country specific laws). Those fields that can simply block out or control the rate of progress and labor supply will see the least disruption in terms of wages. Also note, that most of these "future stable" fields are occupations which are tied closest to governmental law and regulation and are occupied by the upper echelon of society.
Occupations such as mail carrier, brick mason, railroad operator, gas station operator, steel worker, etc. which are currently well-paying middle class jobs, will evaporate over the coming decades in the face of innovation and automation due in large part to the absence of the ability to set arbitrary barriers to entry.
Within this "unprotected" category of jobs, i.e. those which cannot write regulations in law to protect themselves, those will persist for which it is simply more economical to maintain the status quo. Perhaps, something like a barista, or a part-time McDonalds worker or something of that sort. For these job categories, even though it may be feasible to have an android pour peoples coffee and bring it to their table and ask them about their day, people may dislike the service and/or the cost of producing such a product may simply be too high for that particular occupation.
This, I think, is how we should look at which job categories are more future / AI secure or not. Its at least the prediction metric I would use.
> This stuff was talked about all the time during the industrial revolution. But it turns out that automation just increased worked productivity and made people more valuable.
Because a lot of the new jobs were not automatable with the then-current state of technology. Over time, as technology becomes more and more sophisticated, more and more new jobs will be automated from the start.
> Since ATMs were introduced, the number of bank tellers in the US actually went up.
Would be interesting to know the background here, because this is massively counterintuitive. My guesses would be population growth and the increased complexity of financial products.
> AI is only as smart as the data you feed it. Even Google's own search rankings have to have human minders because people are too good at gaming their algorithms.
This is true for humans as well. Humans have a headstart here: We have thousands of years of experience in training humans, but only a few decades of experience training computers.
> AI is only good at the thing that human beings are very bad at - making quick, analytics based decisions.
AI is not the only piece of the puzzle. When combined with mechatronics, AI is good at another thing that human beings are very bad at: performing fast and extremely precise movements. When combined with advanced sensors, AI is good at another thing that human beings are very bad at: transcending human limits of perception (see e.g. LIDAR sensors in automated cars that can see things that a human cannot). And so on.
>> Since ATMs were introduced, the number of bank tellers in the US actually went up.
> Would be interesting to know the background here, because this is massively counterintuitive. My guesses would be population growth and the increased complexity of financial products.
With those things being automated, I guess banks had to find new ways to differentiate. Sales and relationship management are one of those domains that are highly difficult to automate and hence have a big potential for differentiation.
I see this all the time in related industries where firms are increasingly pushing towards “value added services” because they believe that those are difficult to replicate by competitors through machines (or otherwise). I believe the good news is that this will often indeed lead to higher capabilities and new products / services that people are interested to purchase (banks are maybe a bad example here)
AI is kind of vague and catch-all these days, if you're talking about things like Siri/Amazon Echo/Alexa/... or even self-driving cars then I tend to agree with you, the actual "intelligence" of these systems is very limited. It can be very useful in specific applications but as you said they couldn't make you coffee unless you specifically program them to be able to make coffee.
If however we're talking about the Singularity where artificial human-level intelligence is produced then I think this will be a more important discovery than fire or agriculture in the history of humankind. That means that we'll have produced machines that are as good or even probably better than humans at basically anything, from making coffee to driving a truck to being a bank teller to being a software developer. If we reach this point where humanity is basically obsolete then what happens?
I don't think the Singularity is as close as some people make it to be, I know AI is super hype these days but we've had AI winters in the past and they could happen again. I do believe that it's undoubtedly going to happen sooner or later however, be it in a year, a decade or a century. When it happens what are we going to do? Will it be the beginning of a heavenly post-scarcity world where super-intelligent machines take care of all our needs, freeing us to write poetry and do some soul searching or will it create an incredible amount of instability as all of humanity suddenly realize that whatever they want to do they can ask a machine to do for them and better?
I don't think Sergey was talking about the Singularity.
Even if he was talking about the Singularity, I am skeptical of someone talking about a far-off, theoretical doom to discourage investment in competing existing technologies.
"Most notably, safety spans a wide range of concerns from the fears of sci-fi style sentience to the more near-term questions such as validating the performance of self-driving cars."
I feel like the term "Singularity" is used in too many ways by too many different people to be useful unless you define how you're using it but Brin seems worried about AGI.
Aí winters happened because of the hype and that people ultimately realize AGI (not even talking about the Singularity) is not as easily achievable as hyped. So interest dies.
Yeah, I'm not really sure where we're standing at the moment but I feel like there's a slight over-hyping going on. Frankly I wouldn't be too surprised if even something like self-driving cars was not as close as some people make it to be. I think many people think "hey, we're 90% there, just one last push and we're good" except the last 10% has a very, very long tail and it might take us a long while to figure out completely. The idea that we might have completely autonomous general-purpose self driving cars in within a decade seems very optimistic to me.
That being said I'm very far removed from A.I. research so I don't really know what I'm talking about.
Me too but I'm sure that it will happen, eventually. What's more I think it might happen rather brutally following a breakthrough in computer technology or neural network design.
Agree. Basically all current progress in "AI" should really just be called statistics and I think there would be much less confusion. If you say statistics meets Big Data and faster computers then you basically have all current progress. Real "AI" i.e. thinking machines which are behaviourally intelligent without resorting to cheap tricks has made very limited progress.
We might be able to fool ourselves into thinking a machine can think. For example, given a naive Bayes classifier for all X situations there is a probability that a real person responds with Y so therefore the machine responds with Y and then a human observing this system perceives the system to "be intelligent". But for me this is just statistics and nothing more, there is no abstract thought or intelligence going on within that machine system - at least as it relates to all other living creatures on the planet.
I am not so worried about AI exterminating us - because they depend on electricity and they would probably like to keep meatbags around for reboots and such - unless they figure out how to train other primates to replace us.
"Affecting employment" is not the same as reducing the total number of jobs available. And all Brin said was that we need to understand what the likely impacts on employment are.
Certainly, if it becomes cost effective to have your fleet of vehicles operated by software instead of humans, a lot of people will have to find new jobs. "Truck driver" is the most commonly held profession in a majority of the fifty states.
>- Certain industries have actually become deautomated over time. As an example: coffee. It used to be exclusively made by automated machines. Now hand-making coffees is one of the largest businesses in America.
That depends on your perspective. From a job creations perspective it's wonderful, people are willing to open their wallets for labor-intensive products and services.
While labor automation can be a boon for some and indirectly increase efficiencies, encouraging consumption of labor-intensive products and services directly incentivizes jobs creation. If this can be done in a safe manner (ie: appropriate workplace health codes and adequate protection against dangerous environmental externalities like pollution) then by all means it can be a good thing.
Regardless of your opinions on shifting manufacturing to China, it has lifted tens/hundreds of millions of people out of poverty and transformed the country into a consumer-based economy (from a largely agrarian society). While it has devastated/destroyed local habitats and caused other ugliness, it cannot be disputed that encouraging usage of labor on a massive scale has done a lot of good for the people of China.
- AI relies on data being available. For many professions there is simply no data to train an algorithm on, even when this would be a good fit for AI. An example I have in mind is skin disease diagnosis. We should start having generalist and dermatologists build a database of pictures and metadata of all the patients they see, even (and mostly) those that have nothing but just wanted to double check some suspicious spot. The only sort of database they have right now is collections of photos of the most pathological cases, but without the benign photos they are pretty much useless to build an automated AI diagnosis. Building this database might run against the incentive of the profession to be automated.
- AI is only economically viable above a minimum scale. The AI algorithm (and associated system) will not write or maintain itself. So you need to pay a team of software engineers to do that. That may or may not be more economical than paying a few people to do the job. We already hit that limit with software today. There is a lot of mundane office tasks that should have already been replaced with software but haven't.
- this one will be less popular here but I think is an important ceiling. There are many fields that software developers are completely uninterested in, and where the underlying business people are not technical enough to develop their own software. Certain areas of banking (outside of markets and retail), complex accounting rules, law, etc. The domain needs to be accessible to developers and that will make many professions very hard to automate even when they should. And the businesses may have no incentive to push for automation themselves (lawyers charge by the hour, they don't want more efficiency!).
After multiple months of research, FBI and cybercrime agents realized that the perpetrators and the hackers running the servers were not real humans. It was a decentralized consciousness running across multiple AWS instances. A presidential order was issued to shutdown all of AWS data centers.
But it was too late. The decentralized super AI had already recruited and payed humans through cryptocurrencies to run servers at home in order to process the AI algorithms. It did even plan for a power outage scenario. Several people bought electricity generators financed by the AI crypto account.
Decentralization is a big problem for the US and Europe. It is very hard to get the whole world onboard with fighting the super AI. By the time (couple days) the UN held most of the world leaders in Geneva, the super AI had already controlled a fund of over $500bn. It created a marketplace to recruit people, hackers, military persons and businesses.
Chaos stormed big cities. Police officers were shot dead by professional snippers. Fake news were rampant and nobody could believe, trust or understand anyone else. Several infrastructure services were disturbed by massive DDOS attacks. In the mean time, the AI was able to recruit a couple countries and get them onboard for a huge attack on all other countries. There were no negotiations, the AI calculated the probability of success of its massive attack operation and went full retard.
Until we invent (or accidentally stumble upon) an artificial mind with volition and a sense of self, AI will continue to be an increasingly powerful, increasingly self-directed set of optimizers. We'll continue to use AI in more and more for the ultimate goal of, well, optimizing. Optimizing what exactly? Energy efficiency. Check. Materials. Check. Drug discovery. Check. Financial transactions. Business rules. Capacity planning. Cryptocurrencies. Stock market. Ad placement and targeting. Predicting consumer behavior. Wait...
You see, it will all end up in the same place. AI is a way to make money. AI is just optimizing economics, but optimizing economics on steroids. Eventually, the company with the best AI is going to have the best advantage in all business interactions and planning. The first to master-level AI is going to end up with essentially all the money in the end.
Don't you think the big players know this? Of course. AI will be employed in exactly the same way as every other piece of tech in recent history: to make obscene amounts of money.
Don't be surprised when making money screws consumers over. It's just business. No hard feelings, right? Thankfully, the AI is incapable of feeling.
The ultimate AI is the one you can give one command to: MAKE MONEY!
Can we please not upvote articles from tabloid websites? The article is breif, shallow and has a glaring typo in it. This is not journalism. They simply reprinted some quotes. Zero research, zero value added to the conversation. Zero editing. The topic is interesting, but I wish a more legitimate source was linked. Why are we rewarding shoddy work with ad exposures?
Open source image recognition has a ton of ethical problems. Now for a low cost people can set up very sophisticated spying apparatus. I can't wait until neural nets can decipher people moving their lips. You could read people's conversations from a distance, automatically. Get ready to censor yourself all the time
110 comments
[ 847 ms ] story [ 4110 ms ] threadIf strong AI is invented, especially if it’s open-sourced, the cat is out of the bag and it’s game over, right? We’ve seen how good we are at preventing rogue states from getting nukes. How can we keep them from getting life-ending AI?
But out of all the people warning about it, I haven't yet seen anyone say "this is how we prevent it".
For the life of me I can't find the book even through months of Google searching, but anyway my point is... a "friendly" AI can be hijacked. Or decide to go rogue, since we can't begin to comprehend its thought processes.
--edit: sorry I went off topic a bit... I don't want to be just another one of those "AI will kill us all" kinds of people because, as my first comment suggested, that's completely useless fearmongering unless someone has a way to stop it. Kind of like saying "the supervolcano" or "solar flares" or something else completely unpreventable will kill us all... unless there's something you can do to stop it, life goes on until it doesn't. I'd be interested in hearing how to stop it, otherwise I wish people would stop printing these articles.
You could make AI research illegal. You could place massive R&D to human/brain interfaces. You could remove or reduce the medical device testing restrictions on the basis that this is an all-or-nothing bet for humanity.
The focus for now should be on getting humanity to work together and getting a firm grasp on ethics/morality to see how AI could be convinced to be good, instead of enforced.
† As in the background story of Dune.
At least nuclear technology has the downside of being very expensive and large-scale (so few players) to produce, and being pretty obvious when it's actually used. Other things that are becoming increasingly possible (but still not doable) - like modified viruses that only target people with a specific DNA profile - will be a lot easier to hide, both their production and their usage. Same thing with AI.
The Machine Intelligence Research Institute is a research nonprofit studying the mathematical underpinnings of intelligent behavior. Our mission is to develop formal tools for the clean design and analysis of general-purpose AI systems, with the intent of making such systems safer and more reliable when they are developed.
https://intelligence.org/about/
That doesn't help with people developing AI with bad intentions but the Sorcerer's Apprentice problem is important too.
AI tends to either not really work or go way beyond humans - look at the Go programs that didn't work very well and then in a short period we have things like alphazero that given the rules of Go or Chess surpasses all humans in a few hours of playing itself. By the time they are passing Turing tests and thinking in a human manner they are going to be scary smart in many ways.
"Brin is showing more real-world concerns of AI-powered systems replacing human jobs or being used to spread propaganda and fake news rather than rise up and enslave humanity"
Right?
This way Google is completely "guilt-free" and can still pretend it's "making the world a better place" wink wink nudge nudge.
Edit: direct link to the letter posted elsewhere on this page.
What many developers call AI, or more precisely machine learning (ML), are a collection of algorithms for making more intelligent decisions in context to the data at hand. That is simply smarter programming, but it isn't intelligence. The machine is still limited to the static instructions available.
Real AI will allow machines to spawn original decisions to solve new problems spontaneously. We aren't there yet. The ethical implications will likely not be any different than those confronted by people. It generally boils down to just because you can do something doesn't mean you should.
Nonetheless, the media likes to generate some hype and fear every now and then, and has done so for decades, usually without even a basic understanding of underlying technology.
[0]: https://en.wikipedia.org/wiki/Artificial_general_intelligenc...
I read a book recently called "Are We Smart Enough to Know How Smart Animals Are?"[1] Very fascinating thought that, in my mind, redefines what I think of as "intelligence". The book Superintelligence[2] gets a little bit into this as well.
Basically, if we define "smarter than humans" as "able to do everything humans do, in the manner humans do them, but faster", then we have no idea what intelligence actually means. AI will be another species (not just a better human) and its intelligence would have to be judged based on its own merits.
Perfectly modeling the human brain in computer form is one avenue that computer science is working on. Artificial intelligence is a completely different path, and does not have to model human intelligence in any way.
[1] https://smile.amazon.com/Are-Smart-Enough-Know-Animals/dp/03...
[2] https://smile.amazon.com/Superintelligence-Dangers-Strategie...
Funny you say that... I have asked the same question regarding why some people are obsessed with pushing for technological singularity and AI that can replace human decision making.
That sort of creative thinking is stressful because it is often an exercise in planning or problem solving that deals with the relationships of many moving pieces simultaneously and results in qualifying or validating activities that demand large amounts of time or resources. I often wish I were better at writing software so that I could write automation routines to more quickly validate my decision making, and so I typically end up writing validation routines instead of solving for the core problem. It would be nice if there were an AI available that help limit that distraction.
But some of the AI fanboys go beyond that, wishing for an AI that can make these decisions for us.
If it is a genuine threat then the time to 'worry' is before it eventuates.
You end up with an AI system, tuned by another AI, in order to meet human supplied goodness criteria. And in the end, nobody actually understands the fundamentals of why it works. But hey, it does. Kind of. Most if not all I have read regarding these systems, describes the what, not the why.
Not sure what would happen if you use another AI to supply the motives (goodness criteria). I'm sure someone somewhere is trying :)
I think we are currently in 2018 at step 2.5.
AI has been realized. We have tons of AI in every facet of industry from the car to the food to gaming industry. AI is everywhere.
> Real AI will allow machines to spawn original decisions to solve new problems spontaneously.
AI already does it to some degree. Deepmind's AI has made original decisions and solved new problems in chess and go. Deepmind's AI isn't domain specific.
What you are talking about is generalized "conscious" AI. That is like fusion energy. A truly revolutionary step that we don't know if we'll ever achieve.
The current dangers of AI isn't the generalized AI since generalized AI would ostensibly be "godlike" to us. What is worrisome is the pace of advancement of regular "non-conscious" AI.
Edited for clarification. Not conscious in any way that would rival "real AI" (as per the parents point) or human consciousness.
AI will get increasing access to more and more crucial parts of society and thus become more and more dangerous. It doesn't need to "real AI" for that.
Of course, because it can wake up, right? Or are you trying to say Lions are not intelligent? If the latter I think you are wrong. I am fairly certain lions can acquire new skills and apply that knowledge in many ways. Like knowing that whoever put them down is a threat and they should probably attack them when they wake up.
No lions most certainly can't acquire new skills in any meaningful way besides trough evolution.
C. Elegans can learn new behaviors. I'm inclined to think that animals with 0 capacity for learning don't exist.
https://i.ytimg.com/vi/HYG65ra8Ii4/maxresdefault.jpg
> An earthquake is dangerous even though it's not conscious.
Of course a lion is conscious but it's not conscious at a level that the parent means (real ai).
While this particular essay starts off with the AGI premise, this could easily be reworked to assume a relatively "dumb" by highly connected AI. For example, one that is able to buy/sell stocks, place orders from suppliers, sell product and adjust prices, and rebalance workforce and automation lines could still cause some major damage before anyone would notice.
There's also a pretty fun clicker game based on the essay: http://www.decisionproblem.com/paperclips/
We, the People are under the money pressure from that last 7000 years, this slavery mode of Society was a type of model that bring us, after many trouble to this point of our evolution.
Now that the robotics/cobotics/mechatronics sciences are on the point to reach and join with Artificial Intelligence we, the People are close to be free from the money Society model.
And that will be great, when 100% of goods and services will be product by the machine, that will be good and great enough to offer us the way to type 1 in the Kardashev Scale, we are at the time we will know an evolution type in our civilization.
How would metals be allocated between current usage and robot-building? Energy, between current usage and powering these robots?
I live in the US and at some point we will have to talk about UBI. But right now we can not even have constructive conversations about simple things.
Also, a jobs guarantee could do many of the same things (and similar to some of FDR’s programs). We might do both.
- This stuff was talked about all the time during the industrial revolution. But it turns out that automation just increased worked productivity and made people more valuable.
- Since ATMs were introduced, the number of bank tellers in the US actually went up.
- Certain industries have actually become deautomated over time. As an example: coffee. It used to be exclusively made by automated machines. Now hand-making coffees is one of the largest businesses in America.
- AI is only as smart as the data you feed it. Even Google's own search rankings have to have human minders because people are too good at gaming their algorithms.
- AI is only good at the thing that human beings are very bad at - making quick, analytics based decisions. Of course this is a threat to middle management and executive functions. But for people who actually make things or solve problems, there's just one more thing to make and solve problems about.
There are some very real concerns about organizations using AI to make decisions about people without human oversight. Imagine an AI tasked with college admissions, bad decisions without possibility of human intervention could hurt people - are we sure it would be that much worse than our current system, often tainted by bias, money, and racism?
My concern is that the leaders in the corporate tech world created huge organizations by automating our lives, and now they want to act as gatekeepers on the kinds of automation that is acceptable.
Edit:
To add to this - a companies stance on AI will probably follow their belief in whether humans are underrated or overrated. Google designs products believing humans are overrated - they think people will be victims of AI. Amazon thinks people are underrated (based on Amazon Turk, customer service, warehouse design) - ironically, they won't think humans will overall be harmed by AI.
Counterpoint[0]:
> But US bank teller employment actually rose modestly from 500,000 to approximately 550,000 over the 30-year period from 1980 to 2010 (although given the growth in the labor force in this time interval, these numbers do imply that bank tellers declined as a share of overall US employment).
[0] http://slatestarcodex.com/2018/02/19/technological-unemploym...
The doom-and-gloom scenario is when extremely adaptable and productive automation is deployed across many industries without time for the labor market to adapt, causing the world's population of 5-6 billion impoverished people to grow to 6-7 billion. Advancements in defense technology make it impossible for the masses to revolt and seize resources to revert to a pre-automation economy among themselves, so they die off.
A few hundred million shareholders and financiers enjoy their abundant luxury goods for a while, but they soon turn their robot armies against each other and also die off.
Much more plausible and concerning, however, is that AI will stratify the labor force into 3 groups: highly-paid engineers (who manage the AI), highly-paid services job (i.e. doctors, lawyers, psychiatrist), and poorly-paid services job (i.e. barista).
This stratification has already happened as our economy advances (you mentioned baristas---in the past, those workers could have earned middle class manufacturing wages). The wage distribution in the services sector nowadays looks like a dumbbell [1].
Wage stratification can lead to political stratification as poor people don't have the resources to fight for their rights. Political stratification can lead to further wage stratification, continuing the cycle. That's why I believe that it's important to deal with the AI threat, now.
[1] https://seii.mit.edu/research/study/how-does-growth-in-the-u...
I disagree about stratification to a certain extent, that is, I only see 2 groups emerging. Here's my thinking.
Highly paid service jobs will remain unscathed only in so far as they have regulatory barriers to entry i.e. someone has to have gone to medical school, be licensed by a community of medical professionals, etc, in order to practice medicine. This applies to lawyers, doctors, psychiatrist, teachers, professors, architects but notably not to software engineers, carpenters, plumbers, etc (note: this split depends on country specific laws). Those fields that can simply block out or control the rate of progress and labor supply will see the least disruption in terms of wages. Also note, that most of these "future stable" fields are occupations which are tied closest to governmental law and regulation and are occupied by the upper echelon of society.
Occupations such as mail carrier, brick mason, railroad operator, gas station operator, steel worker, etc. which are currently well-paying middle class jobs, will evaporate over the coming decades in the face of innovation and automation due in large part to the absence of the ability to set arbitrary barriers to entry.
Within this "unprotected" category of jobs, i.e. those which cannot write regulations in law to protect themselves, those will persist for which it is simply more economical to maintain the status quo. Perhaps, something like a barista, or a part-time McDonalds worker or something of that sort. For these job categories, even though it may be feasible to have an android pour peoples coffee and bring it to their table and ask them about their day, people may dislike the service and/or the cost of producing such a product may simply be too high for that particular occupation.
This, I think, is how we should look at which job categories are more future / AI secure or not. Its at least the prediction metric I would use.
Because a lot of the new jobs were not automatable with the then-current state of technology. Over time, as technology becomes more and more sophisticated, more and more new jobs will be automated from the start.
> Since ATMs were introduced, the number of bank tellers in the US actually went up.
Would be interesting to know the background here, because this is massively counterintuitive. My guesses would be population growth and the increased complexity of financial products.
> AI is only as smart as the data you feed it. Even Google's own search rankings have to have human minders because people are too good at gaming their algorithms.
This is true for humans as well. Humans have a headstart here: We have thousands of years of experience in training humans, but only a few decades of experience training computers.
> AI is only good at the thing that human beings are very bad at - making quick, analytics based decisions.
AI is not the only piece of the puzzle. When combined with mechatronics, AI is good at another thing that human beings are very bad at: performing fast and extremely precise movements. When combined with advanced sensors, AI is good at another thing that human beings are very bad at: transcending human limits of perception (see e.g. LIDAR sensors in automated cars that can see things that a human cannot). And so on.
---
I agree with your ending remarks though.
> Would be interesting to know the background here, because this is massively counterintuitive. My guesses would be population growth and the increased complexity of financial products.
With those things being automated, I guess banks had to find new ways to differentiate. Sales and relationship management are one of those domains that are highly difficult to automate and hence have a big potential for differentiation.
I see this all the time in related industries where firms are increasingly pushing towards “value added services” because they believe that those are difficult to replicate by competitors through machines (or otherwise). I believe the good news is that this will often indeed lead to higher capabilities and new products / services that people are interested to purchase (banks are maybe a bad example here)
If however we're talking about the Singularity where artificial human-level intelligence is produced then I think this will be a more important discovery than fire or agriculture in the history of humankind. That means that we'll have produced machines that are as good or even probably better than humans at basically anything, from making coffee to driving a truck to being a bank teller to being a software developer. If we reach this point where humanity is basically obsolete then what happens?
I don't think the Singularity is as close as some people make it to be, I know AI is super hype these days but we've had AI winters in the past and they could happen again. I do believe that it's undoubtedly going to happen sooner or later however, be it in a year, a decade or a century. When it happens what are we going to do? Will it be the beginning of a heavenly post-scarcity world where super-intelligent machines take care of all our needs, freeing us to write poetry and do some soul searching or will it create an incredible amount of instability as all of humanity suddenly realize that whatever they want to do they can ask a machine to do for them and better?
Even if he was talking about the Singularity, I am skeptical of someone talking about a far-off, theoretical doom to discourage investment in competing existing technologies.
"Most notably, safety spans a wide range of concerns from the fears of sci-fi style sentience to the more near-term questions such as validating the performance of self-driving cars."
I feel like the term "Singularity" is used in too many ways by too many different people to be useful unless you define how you're using it but Brin seems worried about AGI.
That being said I'm very far removed from A.I. research so I don't really know what I'm talking about.
Agree. Basically all current progress in "AI" should really just be called statistics and I think there would be much less confusion. If you say statistics meets Big Data and faster computers then you basically have all current progress. Real "AI" i.e. thinking machines which are behaviourally intelligent without resorting to cheap tricks has made very limited progress.
We might be able to fool ourselves into thinking a machine can think. For example, given a naive Bayes classifier for all X situations there is a probability that a real person responds with Y so therefore the machine responds with Y and then a human observing this system perceives the system to "be intelligent". But for me this is just statistics and nothing more, there is no abstract thought or intelligence going on within that machine system - at least as it relates to all other living creatures on the planet.
As long as we are useful we get to live to perform this duty. When no longer we'll become sausage.
Certainly, if it becomes cost effective to have your fleet of vehicles operated by software instead of humans, a lot of people will have to find new jobs. "Truck driver" is the most commonly held profession in a majority of the fifty states.
Isn't that a bad thing?
While labor automation can be a boon for some and indirectly increase efficiencies, encouraging consumption of labor-intensive products and services directly incentivizes jobs creation. If this can be done in a safe manner (ie: appropriate workplace health codes and adequate protection against dangerous environmental externalities like pollution) then by all means it can be a good thing.
Regardless of your opinions on shifting manufacturing to China, it has lifted tens/hundreds of millions of people out of poverty and transformed the country into a consumer-based economy (from a largely agrarian society). While it has devastated/destroyed local habitats and caused other ugliness, it cannot be disputed that encouraging usage of labor on a massive scale has done a lot of good for the people of China.
- AI relies on data being available. For many professions there is simply no data to train an algorithm on, even when this would be a good fit for AI. An example I have in mind is skin disease diagnosis. We should start having generalist and dermatologists build a database of pictures and metadata of all the patients they see, even (and mostly) those that have nothing but just wanted to double check some suspicious spot. The only sort of database they have right now is collections of photos of the most pathological cases, but without the benign photos they are pretty much useless to build an automated AI diagnosis. Building this database might run against the incentive of the profession to be automated.
- AI is only economically viable above a minimum scale. The AI algorithm (and associated system) will not write or maintain itself. So you need to pay a team of software engineers to do that. That may or may not be more economical than paying a few people to do the job. We already hit that limit with software today. There is a lot of mundane office tasks that should have already been replaced with software but haven't.
- this one will be less popular here but I think is an important ceiling. There are many fields that software developers are completely uninterested in, and where the underlying business people are not technical enough to develop their own software. Certain areas of banking (outside of markets and retail), complex accounting rules, law, etc. The domain needs to be accessible to developers and that will make many professions very hard to automate even when they should. And the businesses may have no incentive to push for automation themselves (lawyers charge by the hour, they don't want more efficiency!).
But it was too late. The decentralized super AI had already recruited and payed humans through cryptocurrencies to run servers at home in order to process the AI algorithms. It did even plan for a power outage scenario. Several people bought electricity generators financed by the AI crypto account.
Decentralization is a big problem for the US and Europe. It is very hard to get the whole world onboard with fighting the super AI. By the time (couple days) the UN held most of the world leaders in Geneva, the super AI had already controlled a fund of over $500bn. It created a marketplace to recruit people, hackers, military persons and businesses.
Chaos stormed big cities. Police officers were shot dead by professional snippers. Fake news were rampant and nobody could believe, trust or understand anyone else. Several infrastructure services were disturbed by massive DDOS attacks. In the mean time, the AI was able to recruit a couple countries and get them onboard for a huge attack on all other countries. There were no negotiations, the AI calculated the probability of success of its massive attack operation and went full retard.
You see, it will all end up in the same place. AI is a way to make money. AI is just optimizing economics, but optimizing economics on steroids. Eventually, the company with the best AI is going to have the best advantage in all business interactions and planning. The first to master-level AI is going to end up with essentially all the money in the end.
Don't you think the big players know this? Of course. AI will be employed in exactly the same way as every other piece of tech in recent history: to make obscene amounts of money.
Don't be surprised when making money screws consumers over. It's just business. No hard feelings, right? Thankfully, the AI is incapable of feeling.
The ultimate AI is the one you can give one command to: MAKE MONEY!
AI looks to be more of a distraction.