how chatbots have evolved

On the Internet today the only talk is about chatbots (badumtss!). The development of large language models led to the unprecedented popularity of programs, which at first appeared as curious, but not very useful toys in everyday life, then for a long time served as interactive versions of the “Questions and Answers” ​​section on company websites, then became the voice interface of search engines and finally turned into interlocutors who are increasingly difficult to distinguish from living people. Let's figure out where they started, how they developed, and where chatbots are heading.

Shoebox

In 1961, IBM engineer William C. Dersch created a device that could perform arithmetic operations in response to voice commands.

The machine was presented at the International Exhibition of Technical and Technological Achievements in 1962. You could only talk to the device about mathematics. It was able to respond to 16 words: ten numbers from 0 to 9 and several terms for simple arithmetic operations. Having received a voice command, the machine performed calculations and printed the answer.

The device was ironically called Shoebox. And although it could not yet be called a full-fledged chatbot, it already had recognizable features today: the smart machine recognized a person’s voice commands and returned with a useful result. It was on this principle that all the chatbots that followed Shoebox worked.

ELIZA

The world's first chatbot, ELIZA, was developed by Joseph Weizenbaum at the Massachusetts Institute of Technology (MIT) in 1966. The bot analyzed the user's request, compared it with patterns known to it, and gave the most relevant answer. At that time, there were no possibilities for creating branching dialogue scenarios, so Weizenbaum chose a special role for his bot – Rogerian psychotherapist.

This method of psychotherapy assumes that the therapist asks the patient leading questions and asks to tell more about this or that incident from his life, and the respondent ultimately analyzes his psychological state. For a technically limited chatbot, the role turned out to be ideal.

So ideal that ELIZA created a sensation. Dialogues with her were published in newspapers, and many did not believe that the chatbot was not a person. Or, at least, they attributed human traits to him. Weizenbaum himself described a case when his secretary asked to work with ELIZA, and after some time asked Weizenbaum to leave the room so that he would not see what she was talking about with the bot. So, although using an example tightly tied to the context of a conversation with a psychotherapist, ELIZA proved that computers can plausibly imitate human conversation. By the way, you can talk to her on the New Jersey Institute of Technology website.

It is ironic that by the 1970s, Weizenbaum began to condemn his scientific colleagues in books and publications and warn about the dangers of their work. And AI was called nothing less than “index of the insanity of our world.” I wonder what he would say now?

Parry

ELIZA soon got a partner. In 1972, American psychiatrist and scientist Kenneth Colby of Stanford University created the Parry program, which some users considered an improved version of ELIZA. Unlike ELIZA, Parry was able to simulate the answers not of a doctor, but of a patient. In 1979, Parry took part in an experiment where five experienced psychiatrists interacted with a chatbot, trying to determine whether they were having a conversation with a mentally ill patient or with a computer simulating him. In 52% of cases, the chatbot was able to deceive specialists, despite long responses and the inability to express emotions.

The two chatbots were forced to talk to each other more than once. Their most famous conversation is a highly realistic dialogue between a therapist and a patient, generated online. Both developments managed to pass the limited Turing test – an exam in which a machine must convince a human interlocutor that it is also a person.

Of course, at that time, ideas about artificial intelligence were very primitive, people simply were not sophisticated enough, and it was easier to deceive them than any of our contemporaries who had at least once dealt with the simplest chatbot. On the other hand, many chatbots that appeared later and existed until recently differed from their ancestors only in the variety of topics they supported. ELIZA and Parry laid the foundation.

Jabberwacky (Thoughts)

In 1988, British developer Rollo Carpenter created what he claimed was the first artificially intelligent chatbot, which was designed to mimic natural human communication in an entertaining and humorous manner. The project was called Thoughts.

Thoughts differed from previous bots in that it used a feedback principle that allowed it to focus on the context of the dialogue – that is, it tried to remember the details of the conversation and use them in answering the user’s questions. In addition, users could teach the bot: for example, English slang, jokes and word games. In 1997, a web version of the chatbot appeared and at the same time its most famous name was Jabberwacky, and in 2008 an improved version called Cleverbot. The new version, according to the creator, had enhanced learning capabilities and was able to generate answers that were more relevant to the context of the conversation.

By the way, Jabberwacky is still you can try it in action.

Dr. Sbaitso

Creative Labs engineers developed the Doctor Sbytso program for the MS-DOS operating system in 1992. Sbaitso stands for SoundBlaster Acting Intelligent Text-to-Speech Operator. True, it was created not in order to push artificial intelligence technologies forward, but to demonstrate the capabilities of voice generation technology that were available in Creative Labs sound cards. From the point of view of its role, the bot repeated the role of ELIZA: it imitated a psychotherapist. Except that the machine gave answers not only by text, but also by voice. And with Dr. Sbaitso we can talk too.

ALICE

Before everyone knew Alice, there was another one from Yandex devices and services: ALICE (Artificial Linguistic Internet Computer Entity). The chatbot, created by scientist Richard Wallace, appeared in 1995 and became an important event in the development of artificial interlocutors, because, together with the bot, Wallace released AIML (Artificial Intelligence Markup Language) – a markup language that described the general rules by which human dialogue is built . Wallace noted that a language has a limited set of constructions that people use most regularly, and each of them can be assigned a frequency index. ALICE determined the topic of the user's request based on the keywords in it and produced the most likely (frequent) answer from the ready-made ones.

AIML was released as open source, which allowed enthusiasts to create their own chatbots and expand the database of possible answers, which in turn accelerated the development of this field.

Chatbot implementation example.

Smarterchild

The Smarterchild chatbot appeared in 2001 and in many ways became the prototype of modern automated (voice and text) assistants. The chatbot was available in AOL IM and later in MSN Messenger. At first, it provided the opportunity to simply chat on various topics and play text games, but as its popularity grew, the program became able to quickly access information from various third-party services: weather, stock quotes, news, cinema schedules, etc. At the peak of its popularity, the bot had about 30 million users. With AIM's popularity fading, Microsoft picked up the baton and created its own SmarterChild for its MSN Messenger.

Siri and other voice assistants

In 2011, a new revolution occurred in the chatbot field. Apple has introduced Siri, a voice assistant that can help you schedule a meeting, create a note, answer an email, and even show you a map of the surrounding swamps when asked where to hide a body. What was revolutionary about this event was that for the first time a voice chatbot appeared on the market with a direct link to the daily tasks of users, which at the same time sounds like a real person, is able to “understand” requests in a relatively free form and even conduct short dialogues.

Following Apple, Google introduced its voice assistant Google Now. In 2014, Microsoft caught up with the Cortana assistant, integrated into smartphones with the Windows Mobile operating system and computers with Windows 10. Like its competitors, Cortana could serve as a personal secretary and search engine.

The same year, a new round of chatbot evolution occurred – Amazon introduced Alexa. She was able to conduct a conversation with the user, answer his questions, report the weather, news, set music, alarm clock, take notes and perform other tasks. The main difference between Alexa and all previous assistants was the new format for using the assistant: now the smart assistant has settled in small home devices – Amazon Echo and Echo Dot smart audio speakers. The assistant is always on and activated by name.

The format turned out to be so popular that many other manufacturers (Apple, Google, Samsung, Yandex, etc.) subsequently also released their own devices with a voice assistant.

Over the next few years, voice assistants did not change much functionally. They were gradually added new contexts of use: they were taught how to make purchases in online stores, manage smart home devices, but in essence they remained the same, locked into several predefined scenarios of interaction with bots. The stagnation ended when generative artificial intelligence burst onto users' devices around the world.

Generative chatbots

Nowadays, ChatGPT is the first thing that comes to mind when it comes to AI chatbots. However, experiments in this direction were carried out long before the creation of OpenAI. Not all of them, however, were successful.

In 2015, for example, Microsoft conducted a public experiment with a chatbot named Tay, which crashed in less than 24 hours. The essence of the experiment was for the bot to communicate on Twitter with users and at the same time learn to speak in their style and in their language.

At first, Microsoft had no idea what could go wrong with this idea, but found out very quickly. Less than a day after its launch, the bot began swearing, publishing racist statements and other hate speech so violently that Microsoft moderators disabled their brainchild. The company said the breakdown was caused by a coordinated effort by some of the social network's users, as Tay was designed to be a personalized conversation partner.

Be that as it may, the software giant did not conduct any more such experiments, but instead chose to invest money in OpenAI in exchange for integrating GPT technologies into its products.

Google also developed its assistant, which was renamed from Google Now to Google Assistant. At the end of 2018, the company rolled out a big update in which the assistant was already able to independently call restaurants to book a table, hair salons to make an appointment for a haircut, and so on. All using a voice-generating neural network. But ChatGPT, released shortly after this announcement, clearly showed that people need something completely different from AI – namely, the ability to communicate like a human being.

The main difference between chatbots of the past and bots based on large language models (such as ChatGPT) is that the chatbot no longer relies on a pre-programmed user interaction script. Instead, it analyzes the request and generates a response that is as similar as possible to what the trainers showed it during the training phase. AI bots master the algorithm for composing answers to questions, rather than simply taking them from the database. Almost every response from such a chatbot is unique.

With the release of ChatGPT 3.0 in beta in November 2020, OpenAI revolutionized the public's perception of what chatbots could be. For the first time, a product appeared on the market that confidently imitated an interlocutor with encyclopedic knowledge in a variety of fields and extraordinary creative abilities. And this was a serious new step in the evolution of such programs.

Multimodal future

In the time since the launch of ChatGPT and many other chatbots based on generative AI, the flow of enthusiasm from users has significantly dried up. First of all, because generative chatbots have also discovered the limits of their capabilities. Yes, such networks can create unique texts, but their quality is comparable to the skills of a novice live copywriter; they can write code, but with errors. Such networks are capable of generating images and videos on demand, but people on them may still have extra fingers and other anatomical defects that AI trainers have not yet been able to correct.

However, development continues, and the competition in this space is now on who can offer the most intelligent multimodal chatbot.

Multimodality means that the neural network will be able to analyze not only one type of data (text and numbers, image or sound), but any of them, and provide appropriate answers. Multimodality is already available in ChatGPT 4: this version of the chatbot can analyze pictures and describe what is depicted in them, although it can only provide text responses for now. Google's competitor, the Gemini neural network, has similar capabilities.

According to the promises of the creators, in the next versions of ChatGPT the number of types of information that the bot can work with will increase, which means that the number and types of tasks that it will be able to solve will also increase.

Their developers see the ultimate goal of the evolution of chatbots as the creation of general artificial intelligence. The same one who, according to pessimists, will enslave humanity – he will be so good at drawing presentations and making videos. However, if everything goes according to the optimistic scenario, then in a few years we will finally get what the ELIZA chat bot seemed to people 70 years ago: an artificial interlocutor who cannot be distinguished from a living person.