The State of AI

to record another show. But today we're going to talk about the state of artificial intelligence because in late June Dr Nathan Bane and Ian Hogarth presented a report titled the State of a I, giving an update on the advancements that have happened in the field of artificial intelligence over the course of the last twelve months, so end of well summer of seventeen to the summer of eighteen essentially. So I thought it might be in using to talk

about what they found as they researched the topic. But first, let's define a few terms, because artificial intelligence is one of those categories of topics that tends to encompass a lot of different ideas and unless you define what you're talking about early on, uh, you could have two people talking about different aspects of AI, and they don't realize they're talking about different aspects and they think they're disagreeing, but in reality they're actually agreeing. It's just they haven't

defined their terms. So we're gonna do that first. So the term artificial intelligence dates to nineteen fifty five. John McCarthy, who was was a computer scientist, had coined the phrase while developing a plan for a conference that took place at Dartmouth, and it was supposed to happen in nineteen fifty six. So he came up with the term artificial intelligence about a year before the conference was to take place.

And his was a definition of general intelligence, which means it would be a way to allow machines to reason, engage an abstract thought, do some problem solving, and also to pursue self improvement. And I think a lot of people still think of this when they are thinking of the phrase artificial intelligence. That means a computer that can process information in a way that is at least comparable, if not identical, to the way we humans process information.

With the concept of self improvement included, there's at least an implication that such a machine would at least possess some degree of self awareness. It would have to know at least that it needed to improve upon something. It may not really have a full sense of self. However, that's not necessarily required in order to have self improvement, so uh, there'll just be a degree of it. Um. However, it could go as far as a machine that actually knows that it exists within a world it has a

sense of self. That could be a potential uh way to interpret this concept. But as time has gone on, we have needed to narrow down various definitions of artificial intelligence because intelligence itself human intelligence is a very broad term. It encompasses many different things. It takes more than creating a machine that can process information at a faster rate

than humans. We wouldn't say that a calculator is more intelligent than a person just because a calculator can, uh can perform a complex mathematical calculation in a fraction of the amount of time your average person could, right, because your average person can still do tons of stuff that a calculator can't do, so we wouldn't call the calculator intelligent. It takes more than just processing information at a very fast rate, so we have to go beyond just efficiency.

If we want to talk about intelligence, the Encyclopedia Britannica defines AI as the ability of a digital computer or computer controlled robot to perform tasks commonly associated with intelligent beings. Uh. This isn't a bad definition, but it does require you to take another step back to consider what sort of actions would be considered intelligent versus those that are more instinctive. So, for example, if a fly sees that there is a hand coming toward it to squish it and it flies away,

is that instinctual or is that intelligent? While you would argue it's more instinct, it's not a sign of intelligence. Uh. There have been a lot of experiments with insects in general to display that behaviors they have that appear to be intelligent because they involve complex behaviors, Uh, turn out to be instinctual because if you start messing with them, they will just repeat the exact same sequence over and

over again. They don't have a way of retaining the information that just happen in order to build upon it, which is all part of learning. Right as human beings, when we encounter new information, we can incorporate that into our knowledge and then we can build upon that in future encounters with similar scenarios. We can even generalize from that situation and apply that knowledge to similar but different scenarios. This is something that differentiates intelligence from just instinct. AI

is also a multidisciplinary field. It's not just one area of study. It's in fact, lots of different areas of study, everything from computer science to biology, to neuroscience, to psychology, to engineering and more. It's tons of different areas of study that all go into AI, and they're also a lot of different ways that we can categorize AI. One way is that we could just divide AI into two

very large categories week AI and strong long AI. John Searle proposed this way of defining AI back in In fact, he argued that week AI is probably the best will ever do, will never have strong AI. Week AI refers to a simulation of human thought. That you could have a machine that appears to think like a person, but that's just an appearance. In reality, once you strip away all the different layers, it turns out it's just a simulation.

The computer is not quote unquote actually thinking. Uh, it's mimicking the way we think, and it tends to be in a relatively narrow band of applications. It may perform those applications very well, it might even do so better than a human can, but it cannot act outside of that narrow band, or if it does attempt to act

outside of it, it doesn't do so very well. So an example of this might be IBM's Deep Blue, which was the computer they designed to play against chess grand masters, and it plays chess really, really well, well enough to beat grand masters at chess, But you couldn't then tell it to sort a complex series of problems so that you could tackle them properly, or ask it about what the weather is going to be like three days from now.

You couldn't set it to other tasks. It was programmed for a very specific application, and you could not just leverage that quote unquote intelligence to something else. Whereas with a person, you can have a person tackle all sorts of different problems, and even if the person has never encountered that problem before, he or she can apply the information, the knowledge that they have accumulated throughout their experiences an attempt to apply it as best they are able to

the new task. They might not be very good at the new task, but they can at least try to do that based upon the information and the knowledge that they have gained in their past experiences. It doesn't matter how many games of chess Deep Blue plays, it's still

not going to be good at doing other tasks. Strong AI would refer to an artificial intelligence that could actually think in a way that's at least analogous to the way we humans think, even if it uses a different methodology to do so, and it could apply that intelligence to any situation, not just a narrow set of situations. And that does not mean that it would immediately be great at everything. This isn't a definition of a super

intelligent computer. It might also be pretty crappy at brand new tasks, but it can learn over time and improve over time, self improvement being an important concept in intelligence. So I can learn from mistakes and even generalize what it's learned into new, unrelated situations, so it would not be bound by those narrow set of applications like a week AI would. Cirl, By the way, is also the philosopher who proposed the Chinese room thought experiment to argue

against strong AI. I've talked about the Chinese room thought experiment and other episodes. Basically, it's this this thought experiment where you imagine that you are in a room and the room only has a door, and the door has a little slot in it, and occasionally a piece of paper is shoved through the slot. You get the piece of paper, something is written in an alphabet that you don't you don't understand, you have you have no knowledge of this alphabet. It's just it looks like squiggles to you.

But you have an enormous book, and that enormous book has a list of all these different pages of squiggles. So what you do is you consult the enormous book, you look for a page of squiggles that's identical to the one that was sent to you, and then you follow the instructions of what to do. When you get a page that has the squiggles, you follow the instructions, you put the output through the slot in the door, and things continue. Searl argued, this is essentially the same

way that computers process information. They don't understand the information that's coming in. They look at the information, they look for the match of that information against what is programmed to do, and then they respond with the appropriate response. But they don't understand the process. Uh. You would say that the same thing is true of the Chinese room

thought experiment. That if you did a piece of paper written in Chinese, and the person inside does not understand Chinese, they see the piece of paper, they have the book, they write the response, they send it back through from an external observer, it would appear that whoever is inside the room understands Chinese. But the truth of the matter

is you don't understand Chinese. You're just following the instructions. Now, there have been a lot of responses to this particular thought experiment, but that's another episode, so I'm not gonna go into it here, but it's a very interesting philosophical notion. There was an assistant professor of integrative biology and computer science and engineering at Michigan State University named Erin Hints who lays out four types of AI uh in an

article that he wrote for the Conversation dot com. His four types of AI start with type one, which are reactive machines. These base all operations on the current state of any given situation, but it cannot refer back to past events. Uh, As Hints points out in this piece Deep Blue, When I was mentioning earlier. It falls into this type one category. When Deep Blue played chess, it

wasn't tracking moves. It wasn't saying, all right, well, the last five moves, my opponent did such and such, So I suspect based on that that I'm learning more about his style of play. Deep Blue would look at the state of the board, where all the pieces were, and it would do this as if it was a the first time Deep Blue had ever seen the chessboard. It's

not referring to the previous moves. It's just looking at what is on the board at that moment and then starts to evaluate all the potential moves it can make, all the potential moves its opponent can make, and then chooses a move among that set. The next time it's Deep Blues turn, it does it all over again. It is not referring to its past experience. It's only looking at the current state of the board. So again, it can't analyze player behavior leading up to the turn and

then base the decision off of that. So Deep Blue blade chess as if on every single turn it was the first time it ever scenes ever I had a chance to look at that board. Um Ronnie Brooks and AI researcher argued that this is really the only type of AI we should try to build, because any other AI would require some sort of internalized concept of the world. It would need to have some form of representation of the world in its for lack of a better word, mind, in order for it to be able to react off

of that and be us. We humans are the people programming these machines, we would have to program that representation of the world, and no matter how carefully we do, that is never going to be as good a representation as the world actually is. Right, It's only going to be a weak simulation of what we see the world to be, and therefore any decisions such a machine makes based off of this imperfect representation of the world will

in turn be imperfect. Moreover, if I were to program such a computer, I'm doing so based off my own concept of what the world is. But my concept of the world is different from someone else's concept of the world. Someone who comes from a very different background, with a very different set of experiences might have a very very

radically different perception of what the world is. And if I were to design a machine off my perceptions, it might behave in a way that is completely alien to this other person, perhaps in a way that is harmful to this other person, and we'll get more into that later. That falls into the realm of bias. So there are some who argue that we shouldn't even try to go beyond type one because of the potential problems we could encounter.

Type two. AI possesses some limited ability to remember. This is sort of like short term memory for humans, though perhaps it's more transient than short term memory. The memories never get converted into long term storage for these machines, but rather they serve to help a machine make immediate decisions. Hence points to self driving cars as an example of this type of AI. Hence says that, well, they have to identify and monitor other elements on the road that

are constantly changing, such as other vehicles. They have to be able to tell how fast is another vehicle traveling, How close to your vehicle is this other vehicle, what direction are they traveling in? Uh Like, if you're going down the highway and there are other cars on the highway, your car needs to know how many there are, where they are in relation to you, how fast they're traveling. This is all ongoing information. So the alternative would be for your AI to look at the world as a

series of snapshots. Right, But snapshots don't tell you things like speed. They would tell you the thing was here, now it's here. You could interpret it as speed if you knew how long it was between snapshots. But that's a lot of unnecessary processing power. It makes more sense to design AI that has the ability to at least hold information in short term to understand things like velocity. So uh, slightly different version of AI, slightly more sophisticated

than type one. But you don't have to worry about other stuff, right, You don't have to worry about anything outside of whatever the AI's purposes. So a self driving car, for example, doesn't need to know how expensive a jug of milk is, right. It doesn't need to know any of that. It just needs to know rules of the road, needs to know how to identify things that it's places where a car can go versus where a car should not go. It has to be able to identify pedestrians, uh,

these sort of things. But outside of that, the car doesn't need to worry about it, so it still has a fairly narrow set of parameters that it follows. Its worldview, in other words, is constrained, so you don't have to worry about creating a perfect representation of the world. You just have to create as perfect a representation of the specific part of the world the AI will inhabit as

you possibly can to make sure it behaves properly. Type three AI should not only have an internal representation of the world, but also a concept of other entities that are within that world as and it doesn't just note the presence of other things within its environment, but recognizes which of those things have agency. Humans possess agency. We understand our own faculties. We can recognize that others possess

similar abilities. If you and I were to have a conversation, we would do so knowing that the other person possesses at least some of the same abilities that we ourselves do. Right, so we know like you would know that I have motivations, that I have needs and wants. You would know this, and I would know the same about you. We might not know what they all are, but we recognize that

the other person has them. A Type three AI would be able to recognize this and other entities It would not itself necessarily possess needs, wants, anything like that, but it would recognize that other entities do have those things, So it is not self aware, but is aware of others.

So if we just assume that we are the only ones who possess these faculties, then any conversation we would ever have with anyone else would be akin to speaking to ourselves, because we would assume other people don't have those faculties, they don't possess the intelligence that we have. That would mean that every single episode I did of tech stuff would essentially just turn into Tacos and Lord of the Rings, because that's all I would care to

talk about. But I assume you guys want to hear more than that, Sorry, wants to hear about Tacos and Lord of the Rings, or at least one or the other. I don't know which. I'm gonna guess Tacos if I have to guess. So this comes to the theory of the mind, which is also pretty close to what Alan

Touring was talking about during the Turing Test. If we were to create a machine that could reliably mimic a human well enough so that your average person couldn't tell if the responses it was getting the human was getting were from a person or from a computer, You would say that computer passes the Turing test, and Touring would say, you might as well grant that the computer possesses intelligence, because you would do the same thing to another human being.

Right if you talk to another human being, the human talks to you like a human being, you say, oh, this person has some of the same basic aspects of humanity that I have, like intelligence and motivations and needs and wants. Turing said, you might as well extend that to computers if they're able to mimic human interaction close enough so that you could not tell if it was a human or a computer. Even if the computer doesn't possess those things, we might as well assume it does,

because we give that same consideration to other people. However, Hints would say that would really only apply to type four AI. Those are the types of artificial intelligence that have self awareness. This is an AI that not only recognizes there are other entities out there, but understands that it itself is an entity possessing intelligence, and AI in type three would recognize that humans have thoughts, feelings, and motivations, but an AI in type four would have those of

its own. It would have its own motivations, its own needs, its own wants, its own loves and hates all the way in main not defined it in such terms. It would not only be able to recognize motivations, but also understand motivations. It could put itself in the place of another entity and say, this other entity wants to do

X because entity is why. So it might be Jonathan wants to kick open the door because he's hungry for tacos, and the computer would be able to understand this concept, although it may not ever want to eat a taco. I consider that imperfect machine. So that is where we are.

That's the definitions of artificial intelligence, so that you kind of understand, uh, the philosophical approach to what we consider a I. When we come back, we'll talk more about the actual report and what the two researchers found as they were looking into the advances that have been made in AI over the past twelve months. But first let's

take a quick break and thank our sponsor. Much of the work and artificial intelligence that have been following has fallen firmly in the type one category I mentioned before the break. And that's not to say that the work

is boring or it's not useful. The technology needed for type one AI is incredibly sophisticated, so it involves not just developing sensors for a machine to be able to observe its environment, but also the various programs algorithms necessary to process information in a meaningful way so the machine can react in the way that we wanted to react.

Stuff like image recognition, voice recognition, depths, sen saying, all that kind of stuff sort of fall into that category, although they can also be incorporated into higher categories of artificial intelligence, but they are sort of their building blocks essentially. One of the first topics from the report focuses on machine learning and a concept called transfer learning, so we

get to talk about what that means. Machine learning is an approach that involves a computer examining data, learning from that data, and then using what has been learned for future decisions. So, for example, let's take Amazon's shopping suggestions. When you buy something off Amazon, you'll see a recommendation for stuff that other people have bought when they were purchasing the same thing you just bought. So Amazon is using machine learning to try and up sell you more stuff.

It's like saying, hey, when other people bought that thing of a jig, they also gotta do hicky. You probably also want to do hicky because you just buy that thing of a jig. That's a simple example of this. Then you have deep learning that's kind of a a subset of machine learning. Deep learning is sort of a

self correcting branch in machine learning. You train a computer on sets of data, and occasionally you have to step in to make corrections and adjustments to make certain the computer is on the right track, that the computer is not making bad suggestions, which will happen just because it's work from large amounts of data, and sometimes it is making choices that to the computer seem logical but to an outside observer seem wackadoodle crazy, So you have to

go in and tweak things. You might train an algorithm, for example, to recognize pictures of coffee mugs, and then occasionally you have to pop in and you see something that isn't a coffee mug that has been mistakenly identified as one. You have to tell the computer, no, computer, that is not a coffee mug, and then it learns from there. Deep learning depends upon artificial neural networks. These are neural networks that mimic the way brains process information,

and they have algorithms that behave like neurons. Each algorithm processes some information, then assigns a weight to how correct it believes its conclusion to be, like I'm pretty sure this is right, all the way down to this might be right, but I don't know. And then it passes the data down to another layer of neurons, which then takes that information, processes it another way, passes it on, etcetera, etcetera.

And then the system can look at all the different weightings of all the different potential answers and say, out of all the conclusions I've come up with, this one is the one I'm most confident is correct. So that's the answer we're gonna go with. We're not gonna go with any of the others because they are statistically less

likely to be the correct answer. So the key here is that you have to train a deep learning network on a really large amount of data so that you can really get it to grasp the concept, and you also have to make sure you tweak the waiting situation the way it waits how confident it is in an answer in such a way that filters out bad conclusions early on. Uh So it's a little different, like you're you're actually tweaking its decision making process as opposed to

looking at the decisions once they've already been made. And it gets really really tricky, but we're gonna leave it to that. You just you have to very gently guide the decision making process and then you let it go on its own, and then it ultimately starts to produce the best decisions if you've designed the system properly. So these are all non trivial problems, but they are surmountable.

We do have deep learning systems out there now. Transfer learning is where you train a machine to do one thing and then transfer that learning to a new task. It might be semi related, or it might be completely unrelated to the original task. You can reapply the learning model you developed for the first task, and you this cuts down the time it would take to train algorithms

to do something new. Moreover, computer models that have been trained on different problems will begin to build a more rich representation of the world, which I mentioned earlier is necessary for the higher forms of artificial intelligence. The report gives an example in Google Inception V three network, which was trained on image recognition and then retrained on recognizing

skin diseases. The result was that the AI could actually outperform twenty one Stanford dermatologists when it came to making informed decisions such as whether or not a patient should get a biopsy. Next, the report acknowledges the importance of graphics processing units, also known as GPUs. Unlike most CPUs central processing units, a GPU is designed to process a lot of data in parallel, so that ends up being really useful when you're training computer models with enormous amounts

of information. But it's not an approach that works for all types of computing because not all computational problems can be divided up to be solved in parallel, and a GPU would handle a problem that can't be divided up in parallel much more slowly than a very powerful CPU could. This is also true for quantum computers, meaning that when we get reliable, powerful quantum computers will likely see a

real boom in training computers and machine learning. The report also argues that while GPUs are incredibly useful for training a model, the actual application of the model can rest on CPUs. So if you prefer, you'd want to use a GPU when you're putting your computer model through school, but when it's time for your computer model to actually do its job to pursue its career, you switch it over to CPU because the parallel part all comes in

the training section, not in the application section. The report also identifies a few big challenges and pushing AI further. First, there's, of course, the technological barriers. The report points out that processor clock frequencies are are starting to plateau. So generally speaking, clock frequencies tell you how many operations a processor can complete in a second. That's really an oversimplification, but generally the higher the number, the more stuff your processor can

do within a second amount of time. Advance as an AI will likely depend upon new microprocessor architectures to overcome this hurdle. We're reaching the limit of what we can do with the classical microprocessor design. So essentially we're getting closer to the end of Moore's law due to fundamental physical limits that we cannot overcome if we just stick with the way we've been making microprocessors for the last several decades. But that does not mean our our computers

are never going to get more powerful. It may only mean that we have to innovate new architectures, new designs, new approaches to processing, which in turn could necessitate a new version of Moore's law. We might be on another astronomical expansion of processing, but it would require brand new architectures that haven't necessarily been proven yet. The researchers identified Google's tensor processing unit or TPU, as a possible successor.

The TPU is a type of application specific integrated circuit and a s I C. This is a circuit that has made for a very specific application, as opposed to a general circuit like a CPU. A CPU is supposed to be able to handle lots of different data for lots of different applications, but a TPU is meant for a very specific application, for example, for artificial intelligence. Related

to this is another barrier, which is financial. Harnessing really powerful processing technologies is expensive, so artificial intelligence R and D tends to be really costly progress and AI is limited in part by funding. In other words, it's not just technology, it's also where's the money coming from. As for how AI has been coming along, the researchers pointed to Google's Alpha zero, which taught itself how to play the game Go at superhuman levels, and it did that

just by playing itself. It played games of Go against itself, repeat eatedly, with no human interaction. The system didn't have any historical data to pull from. It wasn't consulting historic games of Go and the strategies that people employed. It was developing strategies on its own, so it only had the basic rules of the game programmed into it, and then it just began playing thousands and thousands of games against itself and learned strategies. It would learn tactics, it

would abandon approaches. It had forty days of training, and it reached levels of mastery that could foil even the best human players. The researchers also mentioned Open Ai that's a team that created Ai agents that could play the game Dota too. Dota two is a mobile that's a style of game in which two teams of players try to win a match that involves capturing certain spaces on a playing field and defeating the members of the other team. Like Alpha zero, the team used a self playing feature

as a training mechanism. Every player on a team was controlled by a different AI agent. Now there was one computer that was generating all these AI agents, but each AI agent was acting as its own kind of individual. So each agent had its own neural network, and that meant that these different AI agents were having to collaborate with each other to work together to form these strategies

in order to achieve goals. So this was not just a one computer that was controlling all the pieces simultaneously. It was almost like a separate computer system for every single player. And then they would talk to each other and coordinate with each other, which is really cool and also a little terrifying if you if you think about it, computers working together independently is kind of scary anyway. Another big development and AI is addressing bias in machine learning models.

I mentioned this earlier. Here's the interesting thing. A computer can have a bias, and that's because computers are working off of algorithms that were ultimately designed by human beings, and human beings have bias. If I were to set out to create something, I'd be drawing on my own personal experiences and my own knowledge. But that is a tiny sliver of the spectrum of human experience. The same

is true for people who design machine learning algorithms. As a result, those algorithms might overlook or misidentified data points that fall outside the experience of the architect who designed that machine learning tool, and depending upon the nature of the AI, that could be disastrous. So, for example, back in two thousand nine, Hewlett Packard had to deal with

a scandal. They had these cameras that had image recognition software built into the camera, and it would identify a person's face so that it would focus properly on a on a the subject of your photo. Assume that if you had a person in the picture, that you wanted the person to be in focus. However, they failed to recognize dark skinned people. So that's a problem where your your computer tool is ignoring people because the person who designed it had designed it to recognize folks that were

like themselves. They weren't necessarily thinking about it outside of their own realm of experience, which was obviously a pr nightmare. Now imagine you have that same issue. But now let's go beyond something that is just a public relations problem to something even worse than that. Uh, think about self driving cars. Self driving cars need to be able to recognize pedestrians who are crossing the street. But if you have a self driving car that doesn't recognize a dark

skinned person, then that could lead to tragic results. You could have a terrible collision fatalities. These are non trivial issues. Biases may appear simply problematic on first blush, but they can lead to really catastrophic outcomes, and so in recent months, more work has been dedicated to creating systems that eliminate bias, and that's easier said than done. The researchers gave an example of a biased system with Google Translate as well.

So Turkish is a language that does not have gendered pronouns like he and she. It just doesn't. In Turkish, all pronouns he, she, and it are represented by a single pronoun oh. The researchers translated she is a doctor and he is a nurse from English into Turkish, and Google Translate dutifully change the gendered pronouns in English to the Turkish genderless pronoun oh. But then they went to reverse the process, turned the Turkish phrases back into English.

Google Translate assigned genders to the pronouns. You know, they were both genderless pronouns in Turkish, but in order to make it makes sense in English and not be it is a doctor and it is a nurse, it assigned genders, and it assigned he to the doctor phrase and she to the nurse, even though the original English phrases were she is a doctor, he is a nurse, translated from Turkish back to English into he is a doctor, she

is a nurse. So the genders on the occupations swapped, and that reveals a gender bias in knee translation algorithm. That just assumes that if you're talking about a doctor and the gender is indeterminate in your original language, then it must be a man, which is more than a little problematic. So those are just simple examples, but it

goes much deeper than that. Well, i'll tell you more about what the researchers found in their state of the AI and as well as an update that has happened since that report came out, But first let's take another quick break to thank our sponsors. Another really important concept and artificial intelligence that the researchers point out is transparency, because it's not really enough to have a computer get

to the right answer. We need to know how it got to that answer, and it may turn out that the computer system is making a lot of incorrect assumptions before it arrives at the right conclusion. Those faulty assumptions should be addressed to avoid problems in the future, such as future conclusions that are wrong because they depend too heavily on faulty assumptions. So AI designers need to build in systems that help us check the work of the

AI to make sure this is not happening. This is something that needs to be built into an AI system from the beginning, or else we get into what was called a black box situation. So a black box is where you have a system where all the processes that happen inside the system are hidden away from the average person. You don't know how the system got to its conclusion, and so you don't know if you can trust the

conclusion or not. That's a problem. This always makes me think of the computer Deep Thought, which was the super Intelligent computer and Hitchhiker's Guide to the Galaxy. They asked the computer what is the meaning to life, the universe and everything? And the computer says forty two, Well, you don't understand why the computer got to its conclusion. Of forty two, because the computer doesn't tell you how it got to its answer, It just processes the information and

then produces the answer. We don't want that situation with AI. There's also the danger of inserting changes in data to cause AI systems to make big mistakes. By inserting what the researchers referred to as adversarial patches, you can cause a system to fail. So, in other words, you purposefully introduced bad data to make the system start to have problems throughout the processing of information. They gave examples of image recognition software and showed that by inserting some extra data.

They included one that was a sticker that had a particular design on it, you can override a computer's ability to correctly identify an image and cause it to misidentify it. In the example they used, they showed a sticker that if you put it down in the view of the the camera, then the AI would always identify it as a toaster, no matter what it was, because the sticker was enough to fool the computer into thinking what it was looking at was a toaster, even if there was

a banana saying right next to the sticker. So that's a huge problem if you can fool computer vision into thinking it seeing one thing when it's something else. Again, if we look at the autonomous car example, and you're able to think of a way to have the vision system of the car, assuming that's relying solely on optics,

then you've got a real problem on your hands. But even if it's relying on multiple sensors, if you find ways to fool those sensors or to misdirect them in some way, you will cause the technology itself to behave in a way that it shouldn't because it's acting on the wrong kind of information. The report then goes on to address the issue of talent who are working in the field of AI. They estimated that twenty two thousand pH D educated researchers and engineers are working on AI

around the globe in some capacity. About five thousand of them are very high level researchers. The United States leads the world in open positions for jobs relating to AI research and development. Google is the leading employer of AI talent in the US, but China has produced more pure

reviewed publications relating to AI than any other country. Next, the report looks at how AI has been rolled out in various industries, noting that medical imaging and liquid biopsies are two effective uses of AI applications to help diagnose patients. Healthcare in general is a large area of opportunity for AI. Another application of AI is a little less warm and

fuzzy than healthcare. That would be how governments are starting to put AI at work in surveillance operations, such as incorporating it in CCTV software to include facial recognition technology. I also mentioned Project Maven and previous episodes of tech stuff.

Project Maven was another example they cited. The report covers a ton of other industries from warehouse automation to autonomous vehicles, to security, to agriculture to finance, and essentially all industries are seeing increased AI roll out, but at different rates. So it's not like you're seeing AI suddenly flooding all industries at exponentials speed, but they are starting to get more of an inroad into every single industry. It's just some of them. It's faster than others, but they tend

to improve efficiencies and they tend to reduce costs. But it's also hastening and era of automation that will make it imperative to figure out what the heck to do when it comes to employment, which the report actually does address a little bit later. They also mentioned briefly the recent focus on privacy and security in the wake of things like the Cambridge Analytica scandal over at Facebook, as well as the adoption of g d p R. UH.

Those are, by the way, unrelated to one another, but I've also talked about both of them in recent episodes of Tech Stuff. I can't help but think that as AI becomes more sophisticated, protecting privacy will become a larger challenge. AI will be able to work through large data samples and potentially identify individuals within it with very little trouble. So in light of something like g DPR, this would

make a lot more types of data sensitive. We would have to identify those is saying you need to have this classified under g d p R. This is not truly anonymous data because remember, Harvard professor only needed three points of data to identify of all adults in the United States. That was the person's gender, their birth date, and their ZIP code. That's all she needed, and then she could identify of the adult US population based on

those three data points. So when you think about sophisticated computer algorithms, and they're intelligent, and they are able to work with large data sets very effectively and very quickly, you start to see the potential for fewer and fewer data points to point to a specific individual. And then the concept of anonymized data starts to get really, really fuzzy.

It's hard to say if a piece of data truly is anonymous unless it's just swallowed up by huge amounts of other information and you've you've washed it completely of its individual status. Otherwise there may be a chance of tracing it back to a specific person, and then you have the issues of g d p R. After the industry section in the report comes a politics section UH and that one they look at some survey results that address issues relating to AI, including the employment question I

mentioned earlier. According to the surveys that the researchers were consulting, seventy six percent of respondents felt that the inequality between the rich and the poor will become much worse than it is today as a result of AI and automation. Essentially, thinking those who own the systems that have AI roll out involved and those who own the businesses are going

to profit. Uh and then those who are otherwise affected are going to find themselves out of work, and you will get this increasing gap between the halves and have nots. The found it unlikely that the economy will create new, better paying jobs as a result of AI and automation, so saying that people are pessimistic as being kind of

an understatement. Also, based on the results cited in the report, it seems like most people don't think a universal basic income is likely to happen, not that it wouldn't work, but that it's not likely to get adopted. The results also seemed to indicate many people are concerned about AI's potential dangers, ranging from a loss of privacy to more existential threats, and that more people favor some form of regulation than a fully deregulated approach, with in favor of

regulation opposed and for not really sure. Uh So, almost the same number of people think that AI should have some form of regulation attached to it, as I don't know one way or the other. That might be due to survey wording. We have to remember survey results aren't always in negative of how people really feel, because it often also relies upon the wording used in the survey and how it was administered. The report found that in the United States, unemployment is at a seventeen year low.

Jobs are on the rise, but wages are lagging behind job creation. In fact, it found that in the United States, labor productivity has increased much more dramatically than compensation rates have increased. The researchers also found that many of the new jobs that have been created are low paying ones, so that's problematic. The researchers are quick to point out that you cannot necessarily correlate any of the labor statistics directly with the adoption of AI and automation because there

are so many other factors that are also present. There's just not enough information or evidence to support any firm conclusions about the impact of AI and automation on jobs. Yet, and not only that, the report points out that there are quote unquote only two million industrial robots in the world right now, and then the US has fewer robots in factories compared to countries like Japan, Germany, and Korea. The researchers conclude that the report with a few predictions

of their own. Uh They say that a Chinese research lab will produce a significant research breakthrough sometime within the next twelve months. A machine learning algorithm will be able to design a therapeutic drug that will produce positive results in clinical trials within that twelve months, and that US and China will scramble to sweep up tech companies in Europe and Asia as part of a trade war and AI race, kind of like the space race was in

the sixties and seventies. Uh. And here's an interesting PostScript that was not in that initial report. The day I finalize these notes for this episode, I received a report from Riot Research about an AI bubble and how it is due to burst. So this is just on the finance side of things, not in the technological side of things. It suggested that the return on investments for AI will yield quote rather poor results, with this being akin to

a bubble bursting end quote. It suggests that many smaller companies working in AI could end up folding, similar to when the VR bubble burst in the nineties, but a larger companies like Google will weather the storm and they

will continue to do R and D work in AI. Essentially, the report serves as a warning to investors that they should consider carefully where they put their money with regard to AI applications, as the amount that they invest is going to be greater than the potential yield from those investments. They're essentially saying more money is going into artificial intelligence than is going to be produced from the results of that AI work. At least right now, that may change.

But here's the weird thing is that are not really weird, But here's the kind of self fulfilling prophecy. Is that if investors start pulling their money in order to protect their investments, they don't want to invest in a in an industry that isn't going to return create a return on that investment, then as a result, we could see developments slow down in AI, and then we start to see it plateau, and so it becomes this kind of weird self fulfilling prophecy where in the short term you

may not expect a really good return on investment. That's a big risk. But if you end up heating this warning and you pull your money from investing in such things, then it may never get the chance to prove itself in the long run, and it may just put off true advancements in AI much further than they would otherwise happen. It's a double edged sword kind of situation. Anyway, That is the state of artificial intelligence as of the summer of Who's to say what will happen in the next

twelve months. It will be interesting to see where we are in twenty nineteen, what role AI is playing in various industries and in our lives, and whether or not UH it has truly advanced in a noticeable way, or if it's just a situation where we get incremental improvements and UH an increased rollout, in which case you might say, well, things have gotten better, but not to a point where you you know your socks are gonna get blown off. Who's to say. We'll find out a year from now,

I suppose. In the meantime, if you have any suggestions for future episodes of tech Stuff, you should write me and let me know. The email address for the show is tech Stuff at how stuff works dot com, or draw me a line on Facebook or Twitter. The handle there is tech Stuff hs W. You can also follow us on Instagram and don't forget Our next episode is episode one thousand. I'm letting that sink in. I've done

a thousand of these. I'm so tired, but I'll see you guys on the next episode, and I can't wait to talk to you for the next thousand, So I'll talk to you again really soon for more on this and thousands of other topics because at how stuff Works dot com