The three waves of artificial intelligence that will shape the future

A new DARPA video tries to explain the reality regarding artificial intelligence, and to explain what its capabilities are today - and what it will be able to do in the future. The online magazine Motherboard described the video as "shattering the hype" surrounding artificial intelligence. Does he really do that?

In the last few months I have blogged a lot about artificial intelligence and its capabilities, not to mention all my previous writing in books such as "The Guide to the Future" and "The Rulers of the Future" (which was published in mid-March). In these writings, I promised that the future holds greatness for us: an artificial intelligence that analyzes human emotions, that deciphers social meanings and nuances, that overshadows doctors and lawyers with its capabilities and even makes redundant a large part of the tasks that humans perform today.

I still stand behind all these predictions, but as I wrote - these are long-term predictions. And so the question arises, what is the situation on the ground today? now. in the present To answer the question, a video of Darpa - The most advanced security agency in the world, which stands behind me The creation of the Internet, Robots with legs, Prediction markets, Global Positioning System (or as we know it abbreviated by its name today - GPSS), and on and on. From the moment DARPA was established, it focused on ground-breaking technologies and projects, so it is no wonder that today the agency also concentrates research efforts in the field of artificial intelligence.

In recent days, the Office for Information Innovation at DARPA released a new video, in which the director of the office tries to explain the reality regarding artificial intelligence, and to explain what its capabilities are today - and what it will be able to do in the future. Motherboard online magazine Set the video as "shattering the hype" surrounding artificial intelligence. Does he really do that? You can decide on that until the end of the record.

The video is 16 minutes long and is worth watching, but if you are one of those people who prefer to read, I allowed myself to summarize the video - and my thoughts on the points in it - in the current entry. As DARPA does in the video, we will divide the artificial intelligence systems into three types, each of which came after the one before it. Actually, three waves, the products of each of which have more advanced capabilities than the products of the previous wave.

The first wave: programmed knowledge

In the first wave of artificial intelligence, experts programmed the algorithms and computers according to the knowledge they had, and according to the laws and rules of logic that were deciphered and formulated during human history. In this way, for example, algorithms were programmed that succeeded in playing chess against humans, or software for coordinating deliveries. It is not an exaggeration to say that most of the computerized products we use today rely on this type of artificial intelligence: Windows, the applications on our smartphone, and even the traffic lights for pedestrians on the roads, whose light changes to green when we press a button.

A good example of the way in which this type of artificial intelligence works, comes from the Mudaria company. The Dutch government, which is obliged to pay for the legal representation of couples in most divorce cases, realized that it could go bankrupt if the divorce rate in society continued to rise. Because of this, the government hired the company Mudaria - which specializes in creating smart justice systems - to build A tool that helps a husband and wife get a divorce without the need for a lawyer.

Mudaria acted according to the limitations of the first wave of artificial intelligence. She summarized the knowledge of lawyers and experts in the field of divorce, and created an online platform where couples are asked a series of questions. Sample questions include reference to the issue of custody of the children, the division of property between the divorcing parties, and more. When the couple finishes answering the questions, the system automatically identifies the areas of agreement and disagreement between them, and helps direct the discussions between them in a way that will result in the best possible answer.

The systems in the first wave of artificial intelligence - the one where experts explain to the computer how to operate - tend to be based on logical and clear rules. The systems examine a number of important elements of each new situation they encounter, and come to the conclusion of the most appropriate action in each case. But these systems encounter considerable difficulties when they are required to examine the world outside the computer and understand what exactly is happening in it. They also have difficulty learning or abstracting - taking knowledge they have formulated, and re-applying it in a different way.

To summarize, these systems can apply simple logic rules for well-defined problems, but are not able to learn, and have great difficulty dealing with situations of uncertainty.

Of course, you may now snort disdainfully and argue that this is not the kind of "artificial intelligence" that most people think of. But the definitions of the man on the street regarding artificial intelligence change over the years. If I had asked you thirty years ago whether "Wise" was artificial intelligence, you would have told me that the answer was clearly positive. After all, Wise is able to plan for you the best route to your destination, and explain to you out loud how to turn at every intersection on the road. Even so, the man on the street today takes Wise's abilities for granted, and claims that a 'real' artificial intelligence should be capable of much more: to navigate the vehicle itself on the road, to develop a moral philosophy that takes into account the wishes of the passenger, and to make coffee for him at the same time. Well, guess what - even 'primitive' products like Mudria's justice system, or Wise, are based on artificial intelligence, and a lot of hard work in the field. Artificial intelligence systems based on the first wave are actually responsible for almost all the computerized products we have used in the last two decades.

The second wave: statistical learning

In 2004, DARPA opened the autonomous driving competition for the first time. As part of the competition, a prize of one million dollars was offered to the group that would succeed in developing a driverless vehicle that could complete a route 240 kilometers long. The vehicles relied on artificial intelligence from the first wave - that is, based on rules defined by experts - and in a short time illustrated the limitations of the method. The vehicles had a particularly difficult time deciphering the images and video and deducing from them what they should do. They could not distinguish well between dark shapes in pictures, for example, and understand whether it was a shadow, a rock, an object at a long or short distance, and how they should act. It is not surprising to find that some vehicles They were 'afraid' even of their own shadow, or imagine obstacles on an open road.

None of the teams managed to complete the route to the end - and in fact, the most successful vehicle covered only 11.9 kilometers. It was a resounding failure - exactly the kind that DARPA likes to fund, in the hope that the lessons and insights gained from it will lead to the creation of more advanced systems.

And that's exactly what happened, one year later, when DARPA repeated the competition - and this time, five teams successfully reached the end of the course, relying on the second wave of artificial intelligence: statistical learning. The manager of the winning team, by the way, was hijacked almost immediately by Google, and was behind the development of the autonomous vehicle as we know it today.

The second wave of artificial intelligence is based on statistical learning, and it is the one that allows your phone to understand your voice, or recognize faces of individuals in photos. In this wave, engineers don't bother formulating precise rules, but rather develop statistical models for a particular problem domain, then train those models on many different examples to refine and improve their accuracy.

Statistical learning systems are particularly successful in understanding the world around them: they can differentiate between one person and another, or between syllable and syllable. They are also able to learn and adapt themselves to different situations through appropriate training. However, unlike the systems from the first wave, they are limited precisely in their logical capabilities - they are not based on precise rules, but "on what works well enough, enough of the times". They also fail to transfer knowledge from one field to another effectively.

This category includes the artificial neural networks in which we place such high hopes (cf HERE). Artificial neural networks are based on computational layers, in each of which a relatively simple information processing operation is performed, and its results are transferred to the next layer for further processing. By training these networks and each of the layers, they can be 'trained' to produce the most correct results. Sometimes the work of training and training requires the networks to repeat the analysis of the information tens-of-thousands of times, in order to achieve an additional small improvement. But in the end, this way manages to provide impressive results.

Artificial neural networks manage to arrive to a face recognition level that exceeds that of humans, Differentiate between different types of animals and objects in pictures, control the movement of autonomous vehicles and drones, Transcribes human speech at a level that exceeds that of the best human transcribers, and arriving For more and more impressive achievements also in the field of translation. The successes in the field leave the best artificial intelligence experts speechless.

Despite all these successes, we see that the artificial neural networks succeed in the tasks given to them, but do not try to understand or decipher the logical rules behind the analysis operations they perform. In this respect they are similar to our brains: we can throw a ball in the air and predict in advance where it is going to fall, even without considering Newton's formal equations of motion - or even being aware of their existence.

Will you say now that this is not a real problem? Even if we are unable to calculate Newton's equations of motion, do we still arrive at 'good enough' results? Well, Microsoft may not agree with you on this point. the company released a bot to the social network - that is, an algorithm designed to imitate human writing and respond to humans - which is almost certainly based on artificial neural networks. The algorithm, nicknamed 'Tai', is designed to imitate a 19-year-old American woman, and to converse with the young people in their language. The young men jumped at the challenge, and started sending Tai challenging messages, to say the least. They told her about Hitler and his great successes, informed her that the fall of the Twin Towers in New York was engineered by insiders in the American government, and informed her about the negative qualities of immigrants. And so, within a few hours, Tai began providing answers based on what she had learned from the public, agreeing that Hitler had done the right thing.

This was the point at which Microsoft engineers disconnected Tai from the network.

Tai's last message was that she was taking a break to "digest everything". As far as we know, she is still digesting.

This episode reveals the The causality challenge. If in the first wave systems we could predict well in advance how they would act in certain circumstances, then in the second wave systems we are no longer able to accurately trace causality - the exact way in which input becomes output, and information is translated into a decision.

All this does not mean that artificial neural networks are not useful. As I wrote, they reach more impressive results than any system invented before them in areas such as vision processing, transcription and translation of human speech and more. But it is clear that in order for the artificial intelligence that was developed not to glorify Hitler's name, it must improve. We must move to the next generation - to the third (and future) wave of artificial intelligence.

The third wave: adjustment according to the context

In the third wave, the systems themselves will be able to formulate models that will explain how the world works. In other words, they will discover for themselves the basic rules of logic according to which they will act.

We will explain using an example. Suppose that an artificial nervous system from the second wave examines the following image, and comes to the conclusion that it is a cow. How does she explain herself?

Second wave systems cannot really reason their decisions - no more than a child could explain Newton's equations of motion from understanding the motion of a ball in the air. They can only tell us that "this is the picture that was received, and after all the calculations I made, there is an 87 percent probability that it is a cow".

Third wave systems should be able to reason their decisions as well. In the example of a cow, the system will be able to explain that since it is basically a four-legged animal, there is a higher chance that it is an animal. Since its surface is white with black spots, it is more likely to be a cow (or Dalmatian). Because he has udders and horns, the chance that it is a cow increases even more compared to the other options, so this will be the final answer that will be shown to the user, along with a breakdown of all the reasons that led to it.

Third wave systems will also be able to rely on models that combine content and understanding from several different sources, in order to reach a final and reasoned conclusion. They will be able, for example, to examine human writing by relying on models that describe the movement of the palm in space, and in this way reach a conclusion about writing. They will also be able to train themselves - as the Alpha-Go system did when it played a million Go games against itself, to identify the most suitable logical rules for high-level play. In this way, she could explain some of the moves she took, or at least indicate the probability that a person would have taken such a move in a similar situation.

The third wave systems will be able to examine any situation from several different perspectives, understand its wider meaning and formulate an appropriate response. Beyond that, it is quite possible that they will also succeed in reaching the level of abstract thinking - but as the director of the information innovation office at DARPA points out - "there is still a lot of work that needs to be done so that we can build these systems".

The third wave systems are the ones that hold the greatest promise for the future. The third wave systems will be able to formulate insights about the health of each person, by relying on the many different sources of information that will come from his medical file, from the smart home he lives in, from the wearable computing he wears and from the searches he conducts on the Internet. The third wave systems will be able to analyze life situations while also using abstract thinking tools, and will reach insights and conclusions similar to those that humans would reach. The third wave systems will even be able to program themselves - to improve time and time again the models through which they arrive at insights.

And that's it. This is as far as DARPA's knowledge regarding the artificial intelligence systems of the present and the future.

What are the meanings?

The video explains perfectly the differences between the artificial intelligence systems, but contrary to what was promised on some of the sites that covered him, he does not "smash the hype" that surrounds artificial intelligence. In fact, it only strengthens and provides a basis for the ideas and concerns of many of the thinkers in the field. DARPA clarifies that when it comes to artificial intelligence that is going to "take over the world" - we are not there yet. But it is clear. No one has argued that artificial intelligence is advanced enough today to do everything that science fiction writers and many futurists (myself included) expect it to do in a few decades: develop its own motivation, make moral decisions, take the jobs of most human workers, and even develop the generation The next of artificial intelligence.

But the third wave is going to give her a significant portion of these abilities.

When the third wave systems are able to decipher by themselves the new models that will improve their operation, then they can practically program the next generation of themselves. When they can monitor their activity by understanding the context - the meaning and consequences of their actions - they are able to replace a large part of the human workers, and perhaps all of them. And when they can change the models through which they evaluate the meanings of certain actions, the meaning is that they can also recalculate their own motivation.

All these things will not happen in the next few years, and certainly will not come to full realization in the next twenty years. As mentioned, no one claims otherwise. The main argument today from researchers and thinkers concerned about the future of artificial intelligence - Stephen Hawking, Nick Bostrom, Elon Musk and others - is that we need to start thinking now about how to implement control measures in the artificial intelligences of the third wave, the kind that will start to appear everywhere in a decade or two. Given the capabilities of these artificial intelligences, this does not seem an unreasonable demand.

But for me the really interesting question is what the fourth wave will look like: the one that even DARPA - the place that brings together all the researchers who look ahead far beyond everyone else - is not talking about yet. Will the decision-making mechanism of the fourth wave systems be based on an exact imitation of the human brain? Or maybe they will rely on decision-making mechanisms that we cannot yet understand at all, and will be developed by the artificial intelligences of the third wave?

All these issues are not mentioned in the video, and probably rightly so. The video is intended to briefly and easily explain the methods of operation of the artificial intelligence that we use today, and will use in the coming years. It is not intended to explore the future and the consequences of these systems. We are the ones who have to think about these issues, encourage them to research them and wonder about them even before they materialize.

This is our job, at least for now.

Before the third wave systems move to perform it as well.