In this text, we will not talk about the Terminator at all. And not about how to prevent the uprising of the machines. We will talk about the things that developers and engineers have to deal with in order for artificial intelligence to become truly massive.
Big data monopolization
Big data is needed to train artificial intelligence. Big Data is often referred to as the "oil of the 21st century". And now most of the "oil reserves" are concentrated in the hands of a few corporations - Google, Facebook, Amazon, Microsoft, IBM. Even the users who generate them have no access to them.
When buying equipment (for example, the most affordable Android smartphones) or using services for free (Chrome, Gmail), in fact, users pay for these benefits with data about their behavior and habits. Corporations are now using this in ad targeting, but some have gone further. Giants like Google and Facebook are already modeling reality for their users with algorithms to rank the results of information. The more data, the stronger the control, and the more likely it is for some companies to create advanced artificial intelligence, while others (the ones that are deprived of access to an array of user behavior data) are left out of work. Corporations buy the most promising AI startups to use their algorithms or gain access to their Big Data. Or just to destroy a potential competitor. But as long as Big Data is concentrated in the hands of a few players, the risk of artificial intelligence monopolization is high. The remedy for monopoly is the open market, where both Big Data and processing algorithms can be bought and sold freely. Work is already underway in this direction (including by Ukrainian or Russian startups).
Data errors
Artificial intelligence uses the data that a person has generated for "learning". So (aside from virtually “infinite” memory and the ability to process large amounts of data quickly), AI is only as smart as the data we give it. For example, neural networks that learn to draw "master" the styles of existing artists and create paintings, repeating human creativity.
The Yandex neural network, before writing poems in the spirit of Yegor Letov, studied his songs, and, in fact, did not bring anything of its own to the style. From time to time, they say that AI will replace the profession of a journalist: it will be able to master large amounts of data faster and write like a living person. But along with human knowledge, AI also absorbs human errors. Roughly speaking, if the neural network learns the style of a certain writer, it will also learn his typical mistakes. However, if problems with style in modern journalism are a tolerable thing (after all, the Internet is full of illiterate texts), then when it comes to medicine, law or unmanned vehicles, the price of such mistakes is too high.
Fear of the "breaker"
One of the greatest physicists of our time, Stephen Hawking, says that artificial intelligence can become both the best and worst invention of mankind. Among the risks of AI, the scientist calls its use by a minority to exploit the majority. And - a classic - a direct conflict of AI with humanity. Science fiction of the last 50-70 years is literally saturated with this idea: there is a theory that, having gained consciousness (that is, having realized itself), artificial intelligence will feel threatened by humanity. Since it is in the hands of people that there is a “breaker” that can turn it off. And, of course, will try to destroy the threat.
Conscious life forms tend to deal with potential threats, that's right. And although we are still very far from creating a highly developed AI (the one that is able to realize itself) (and in general, it is not a fact that it will be created), there is one more good news. Some experts in the field of AI, in particular, the Ukrainian Maxim Orlovsky (PhD in neuroscience, head of BICA Labs, a laboratory dedicated to solutions in the field of AI) believe that the creation of artificial intelligence without a "breaker" as such will be a panacea for potential conflict. And blockchain technology will help in this, which just implies distributed work and the absence of a single control point. Now Maxim and his BICA Labs are working on creating such an AI.
The wrong hardware
The best minds in the world are struggling to teach AI what a small child can do. Object recognition, for example. But in order to “think” like a human, artificial intelligence must work like a human brain. And our brain simultaneously processes information from different sources and is not overloaded from this. For him, this is a daily routine. Neural networks run on hardware that is not designed for so many parallel threads. Graphics processing units (GPUs) are more suitable for such tasks than classical ones (CPUs), but both of them do not work perfectly both in terms of speed and power consumption. Fundamentally different principles are needed, and the work of scientists does not stop on them.
In recent years, several companies have announced the development of chips for neural networks. In 2014, IBM introduced SyNAPSE, which processes many signals simultaneously, like a biological brain, while consuming minimal resources. But with SyNAPSE, not everything is so simple: the chip turned out to be incompatible with the existing computing architecture, and the company even had to write its own programming language for it. Nevertheless, the development turned out to be promising, which is confirmed by the fact that even DARPA invested in it. In 2016, developers from MIT introduced Eyeriss, a prototype 168-core processor for working with neural networks in low-power devices. Other companies are also talking about announcements. So far, we are not talking about mass production and mass availability of such microcircuits. That is, the technology will become available and popular very soon. As previous experience of technology development shows, since the appearance of the first models, several generations of device solutions have been improved and refined. Now we are at the very beginning of the path of creating a real hardware for artificial intelligence.
Unavailability of infrastructure
Artificial intelligence is including unmanned technologies. Self-driving cars, drones, assistant robots that will take over everyday routine or help you navigate in an unfamiliar place — today their work is built in such a way as to avoid collisions and dangerous situations within the already developed infrastructure.
You have probably heard about the dilemmas for self-driving cars about how the car should behave in an emergency - to rescue passengers or pedestrians, for example. Ground delivery robots are trained to avoid obstacles so as not to become a source of problems for pedestrians. Solving this kind of problems takes a lot of resources from developers. But, it is quite likely that it lies on a different plane - the plane of creating a completely new infrastructure for unmanned objects, where they interact only with each other, and where there is no chance that a child will run out towards you or a truck of the color of the sky will leave (if it is an unmanned truck, it will will be able to give signals to other objects on the road that are part of the Internet of things).
GO TO FULL VERSION