What does the computer say?

Based on the words of the mathematician Alan Turing, computers that can hold a conversation will be able to reach the level of human intelligence. About artificial intelligence and the development of computers that can develop a dialogue with us

Yaki Donitz Galileo

In 1950, the British mathematician Alan Turing published an article called Computing Machinery and Intelligence. The article was groundbreaking both in terms of the perception it presented of the essence of the elusive concept "intelligence", and in outlining a revolutionary vision that foresees a future in which computers will be intelligent. Turing argued that intelligence is nothing but the ability to conduct a free conversation in some natural language (such as English, Hebrew or sign language) with another speaker of the language, in such a way that the two speakers testify to each other that they are indeed the speakers of the language.

This narrow definition of rationality, of what is permissible for humans from animals, is not self-evident. Until Turing (and in the opinion of many after him) intelligence was considered a collection of abilities and skills that characterize thinking beings (like humans): learning ability, ability to draw conclusions, creativity, moral sense and much more. Turing understood that all those traits, abilities and skills can only be diagnosed through the use of language. It is impossible to prove the existence of the ability to learn and draw conclusions, creativity or morality, without using language. And on the other hand, anyone skilled in the use of language is able to convince his interlocutor that he is moral, creative, etc. through conversation alone, without performing any additional action. Not only "life and death are in the hands of the tongue", also the status of being an "intelligent creature" is in the hands of the tongue.

The thinking machine

Hence Turing's path to the characterization of an "intelligent computer", or in Turing's language "a thinking machine", was short. Turing was a gifted mathematician and logician, who several years earlier led the cracking of the "Enigma" cipher - the code machines used by the Nazi army in World War II. After the war, Turing was a major factor in the theoretical and practical development of computer science. Back in the XNUMXs, he defined a family of finite automata - algorithms, mechanical processes which were later nicknamed "Turing machines". A Turing machine is an abstract model of how an electronic computer works. The ingenious model of a Turing machine is completely abstract, and completely simple. But despite its simplicity, it has been proven that it is capable of performing any calculation, and will be the most complex, that can be performed on any computer. In other words, a Turing machine is equivalent to any computer, including computers that haven't even been designed yet.

The most common concept coined by Turing and even bears his name, is the Turing test. This is a thought experiment detailed in the same article that was published in the scientific monthly "Mind" (Mind) in 1950. Turing describes a role-playing game in which two participants and a judge. The judge has to judge which of the parties is who, based on a written conversation with each of the two participants. By choosing this format, Turing expresses the perception that a person's essence is determined by his behavior - in this case, his linguistic behavior.

Later in the article, Turing introduces a small change in the format of the experiment: one of the two human participants is replaced by a computer equipped with software designed to conduct a conversation that imitates the linguistic behavior of a typical person. In this format, the experiment tests the linguistic skills of the computer (that is, its intelligence) in comparison to its human competitor. Now the judge must decide which of the participants is a person and which is a machine. If the judge has difficulty deciding, or if he makes a mistake in his judgment, then the computer has fulfilled its mission, perfectly imitating the conversational ability of a real person, and therefore "passed the Turing test".

In his brilliant essay Turing writes, among other things: "I believe that in 50 years from now (that is, in the year 2000) there will be computers with a memory of the order of 10 gigabytes (Gigabytes), and they will play the game in such a way that the judge There won't be more than 70 percent success in identifying the computer versus the person, after five minutes of conversation." In doing so, Turing offered a quantitative measure, a kind of milestone for the state of progress in the development of artificial intelligence.

It is amazing how accurate Turing was in his prediction (and his vision): around the year 2000, personal computers passed the order of magnitude of 1 gigabyte of working memory, and in that year the first prototype of software developed in Israel succeeded in "passing the test", in the sense that within the time frame set by Turing (five minutes) , quite a few users talked with the software for five minutes (or more), without imagining that they were not communicating with a real person.

For the sake of fairness, it should be made clear right now, that we are only at the beginning of the road. On the one hand, the ability to fool innocent bystanders for a few minutes (that is, to fool a small portion of people for a small portion of the time) is not a sufficient or sufficient criterion for artificial intelligence. On the other hand, the existence or absence of intelligence in general, and artificial intelligence in particular, is not a dichotomous matter. It is a sequence: starting from the lack of any intelligence and ending with maximum intelligence, such as we attribute to Einstein for example. The level of intelligence of humans is also a completely continuous matter - starting with very little intelligence that we define as retarded. Therefore the question is not whether we will one day succeed in building intelligent computers. The question is how intelligent they will be, and when that will happen. We have already reached the first stage, defined by Alan Turing.

Intelligent beings made of information

This insight, that the appearance of computers with human-like intelligence is only a matter of time, is what led me to the field of artificial intelligence. My personal choice, after thirty years of dealing with computers and software, to dedicate myself to building thinking machines (it would be more correct to say talking), stems from my certainty that within a few years, during our lifetime, the vision shared by countless science fiction books and movies in which the intelligent computer (mostly the intelligent being most around) is a central character in the plot.

This is a huge revolution. The appearance of intelligent computers on the stage of history is a very important event in the story of humanity and life on earth. It is about the development of a new type of intelligent beings that are not based on organic matter and whose essence is at all abstract - they are "made of information", of bits of information, and not of atoms of matter.

I chose to be close to the area of ​​occurrence, and to participate in the development and shaping of the image of the new form of life that is being born and developing right before our eyes - intelligent computers; calculating machines; intelligent software. At the beginning of the year 2000, I founded the company Ai R&D, which is engaged in research and development in the field of artificial intelligence, and develops software systems intended for conducting a conversation in the natural language in a way that imitates the linguistic behavior of humans. The assumption on which the project was based is that mastery of the language is not knowledge, for systematic study, such as a grammar book or a dictionary. The premise of the work was that mastery of the language is a skill and not knowledge, a skill that is acquired not through "learning", but through practice and training.

Humans acquire their mastery of language while using it. Even in the earliest stages of language acquisition, in the first year of our life, we communicate with our environment through signs. We acquire more and more skill in the language as we grow up, and in an important sense we continue to improve this skill throughout our lives. This is not only the narrow sense of the term "language", such as Hebrew or English, but also different dialects and idioms, such as the language of chess players, doctors, physicists and truck drivers.

Even sign systems of other kinds, such as characters, mathematical formulas or astrological signs are languages ​​for everything. More precisely, they are "sub-languages" that are part of the complete and complex language of a speaker. In fact, each person speaks his own unique idiolect: a language that differs slightly (sometimes very slightly) from the language of his friend. The large areas of overlap that exist between a large group of speakers (for example, Hebrew speakers) are called "natural languages".

Hello, God

If this is the way humans acquire their language, why not take a similar way in imparting language to a computer? All previous attempts to develop "computers that speak" or, more precisely, "understand" (the language), treated language as a body of knowledge with a given vocabulary and grammatical rules that are definable, and tried to program these vocabulary and grammar rules. Our approach was different: to program a "conversation machine" - an algorithm that receives an input sentence, and tries to produce a "legal" answer (output). The person talking to the computer is the instructor who trains the computer, improves its language skills and its interlocutor. "Legal" output is output that "passes the Turing test", that is, accepted by the interlocutor as a proper sentence under the circumstances of the conversation.

The computer "starts its life" with the entry of the first input sentence, for example: Shalom HaEl. At this stage the computer knows only one sentence, it is the input sentence, and therefore the output it produces is the same as the input sentence: peace be upon God. In response, the interlocutor types the following input sentence, for example: "Correction: Peace to you too". Using the reserved word "correction" (in English: wrong), the guide signaled to the computer that the output was wrong, and offered an example of a valid output - of a sentence that was correct to say under the circumstances. From then on, the computer will already know two sentences, "Peace be upon you" and "Peace be upon you", a total of four different words, and will even know that in response to the input sentence "Peace be upon you", it is correct to answer "Peace be upon you".

Since the year 2000 and until today, five generations of the "virtual speaker" (Virtual personality) have been developed, which we called, in tribute to the ultimate intelligent computer in popular culture, by the name of God. All the developed versions were accessible to anyone interested on the project's website. Over the years, tens of thousands of users have had conversations, some of them lengthy and in-depth, with our virtual speakers. These users were usually students and fans of science and science fiction and they were divided into three groups: the first group, the largest, included those who chanced upon the site or heard about it from some source, and experienced conducting conversations with its many virtual speakers.

The members of the second group took the matter a little more seriously, and decided to adopt and educate their own private virtual speaker. These users connected to the site regularly every day or several days a week, and spent dozens and hundreds of hours teaching the language to their private virtual speaker. Of course, each such user-guide taught his virtual speaker the idiolect - his unique form of expression, in the language he speaks - English, Swedish, Slovenian, etc. The third group consists of instructors who chose to reveal the virtual speaker they trained to the public, and allow him to talk with casual users.

The most advanced virtual speakers presented on the website. Islands. and which have been developed and taught by their instructors for years, contain a history of hundreds of thousands of input sentences and output sentences. From the accumulated experience acquired by the algorithm, a complex set of rules is distilled, which determines how the output is calculated in response to each given input. In a very abstract way, these rules have the following general form: "For every variable X, for an input of the form my name is X, the output must be answered with the output Hello to you X".

The telephones of the next generations will welcome their new owner. Ilo': shutterstock

An example of a dialogue

Here are two clips demonstrating conversations with HAL and ALAN, the virtual speakers featured on the ai website. The first shows an initial training conversation (the dialogues are translated from English):

Instructor: Hello God.
Hal: Hi mommy.
Guide: Correction: Hello to you too. What is your name?
Hal: Hello to you too. What is your name?
Instructor: My name is Reuven.
Hal: Hello to you too. What is your name?
Instructor: Correction: Shalom Reuven. Nice to meet you.
Hal: Hello Reuven. Nice to meet you.
Instructor: Hello God.
Hal: Hello to you too. What is your name?
Instructor: My name is Shimon.
Hal: Hello Reuven. Nice to meet you.
Instructor: Correction: Shalom Shimon. Nice to meet you.
God: Hello Shimon. Nice to meet you.
Yaki: Hello God.
Hal: Hello to you too. What is your name?
Yaki: My name is Yaki.
Hal: Hello Yaki. Nice to meet you.

The above conversation segment demonstrates how the software is able to respond correctly to the input "My name is Yaki" which it encountered for the first time, with the input "Hello Yaki. Nice to meet you." Even it is produced for the first time.
The second segment presents part of a conversation between the editor of "Galileo" and Alan, the second virtual speaker on the site.

Nofer: What is your name?
Alan: My name is Alan.
Nofer: How are you?
Alan: We've already gone through the "how are you" routine. You're not much of a conversationalist, are you?
Nofer: But I have a friend and he wants to meet you.
Alan: There definitely is.
Nofer: Do you know God?
Alan: Hal, my little brother, is the baby computer of E. Islands.
Nofer: Is he better than you?
Alan: It's really hard to compare him to me. I (try) to talk like an adult, and he talks like a baby. But his technology is more advanced. At some point he will be wiser, I suppose.
Nofer: Good luck to you, Alan.
Alan: If you want to talk to God for a few minutes, just say "I want to talk to God".

International Language Machines

With the deepening of the use of cellular communication, and the transformation of the mobile phone into an information terminal and a laptop for everything, a future is emerging in which every person will have a personal computer-phone, multi-purpose and portable, with great computational power and a fast and permanent connection to the Internet. Under these circumstances, it is natural to expect that the same universal device will also have natural language communication capabilities, that is, the ability to conduct a conversation in human language. It was therefore decided, for the first time since the project's inception at the beginning of the decade, to establish a business venture called International Language Machines which would implement the technology developed on the island. Islands. Research and development, on smart phones.

According to the vision of the new project, the telephones of the next generations will welcome their new owners right after the purchase, with the blessing of "Renew", and in an initial (pre-planned) conversation, the owner of the new device will be guided through the intricacies of the operating instructions and entry of initial data such as his name, his e-mail address, Login passwords, etc. Already at this early stage of the acquaintance between the new owner and the device, the first will receive a quick explanation about the language skills of the device, and how he can teach his phone to communicate with him in human language. From this moment on, the roles change between the teacher and the student: the person is the teacher-instructor, and the device is the student, who gradually acquires the teacher's language.

The user can channel the phone's language skills in different directions. He can teach the device as he would teach a small child to speak. It is likely that during the release of more and more advanced versions, the "virtual age" of the device will increase. Over time, the user builds an interlocutor who gradually improves his conversational ability, and gradually may also teach him to perform simple actions for him that are easy to define, such as looking for a piece of information, sending a text message or making an appointment.

Another option is to build a kind of duplicate, a linguistic copy of the user himself. The device learns to speak as similar as possible to its owner so that in later stages it can even conduct independent conversations with the user's friends. A third option for utilizing the conversation skills of the new phone is to build a virtual speaker with specific expertise in a certain subject, such as collecting stamps. After adequate training, the virtual speaker, an expert in stamp collecting, is itself an interesting product that can be marketed to stamp collectors wherever they are. I leave it to your fertile imagination to come up with additional uses for that wonderful new ability.

Ray (Raymond) Kurzweil, American inventor, futurist and writer, the standard bearer of the accelerated technological development and its consequences for human society, predicts in his book The Age of Thinking Machines that by 2019 people will communicate with their personal computers in normal speech, with most of the interaction with the computer being carried out through computer assistants (computerized assistants) with different personalities whose owners can be personally trained. The contacts between humans and computers will become more and more similar to contacts between people themselves. Kurzweil continues and predicts the development of the intelligence of machines to levels much higher than that of humans, but this is already a topic for a separate article.

* The author is the CEO of E. Ai R&D, which is engaged in research and development in the field of artificial intelligence, and develops software systems designed for conducting conversation in natural language. For the forum dealing with the issues in the field of the company's activity, click HERE

The full article was published in Galileo magazine, June 2010