- AI
- A
A brief history and terminology of artificial intelligence
In modern realities, there are practically no people left who have missed the "neural network" noise. For some, it has even become a fundamental tool in their work, and some even equate its importance with the internet.
Neural networks are beginning to tightly integrate into our lives, fortunately, as a friendly tool that helps improve the accuracy of analytical conclusions. They are used by ordinary people for simple tasks (help plan the day or edit a letter), as well as by scientists in laboratories for diagnosing, checking the compatibility of various biological components, etc.
In today's information noise, it is difficult to focus on the history of things that are secondary to you, so even the most active user of artificial intelligence may not know where the "roots grow" - but it would be useful!
In my professional activity, this term often appears, and working circumstances force me to closely monitor the market to understand the pace of development of new technology and determine the right time and product to introduce this innovation to clients.
In this article, I will talk about the development of B2C (for individuals) AI solutions, moreover, I will touch on two markets: Russian and Western. We will talk about the development of technological giants in the field I am considering, where they started, what they produced, and for what purposes. But let's start with the foundation.
First conversations
In the mid-20th century, the first concept of a neural network began to form: researchers tried to create a machine that mimics the functions of the human brain. In 1943, Warren McCulloch and Walter Pitts proposed a mathematical model of a neuron, and by the late 1950s, Frank Rosenblatt introduced the perceptron - a simple machine learning model created to help computers learn from different volumes of data. It can be called the first practical implementation of a neural network.
In the period 1980-2000, the first algorithms for learning, comparing, and analyzing data began to be developed and appeared, and in the 21st century, the direction began to develop rapidly. In the 2000s, powerful graphics processors appeared, and large amounts of data became available, which prompted the community to start developing Deep Learning algorithms - a set of machine learning methods based on learning representations rather than on specialized algorithms for specific tasks.
The interest in embodying human thinking "on a board" appeared almost immediately as computing technology became available to most people. In the 20th century, people were able to achieve certain results fairly quickly due to algorithms that were invented by humans, but they could not advance further due to the insufficient development of technologies.
Development of modern neural networks in the West
OpenAI
In December 2015, the non-profit organization OpenAI was founded - an American research organization engaged in the development of artificial intelligence.
On April 27, 2016, the company released a public beta version of OpenAI Gym, a platform for developing and comparing reinforcement learning algorithms
This was a kind of machine learning task. The essence is that the user was offered an environment with configured rules and a body capable of acting within this environment. Users developed and compared their algorithms on this platform.
In 2017, AI bots for playing Dota 2 were introduced
An unexpected but proven product. Indeed, OpenAI managed to create a bot for a popular competitive game that does not act according to a described algorithm, but makes decisions independently based on a previously loaded dataset (a dataset for training neural networks)
On May 28, 2020, a group of researchers from OpenAI led by Dario Amodei published a scientific article with a detailed description of the GPT-3 algorithm.
At this stage, the world was presented with the principle of operation of the GPT-3 generative neural network. The presentation did not make a splash, because there was no finished product, only the general idea was shown. However, this did not prevent attracting key investments to the project, which helped to release an innovative product to the market!
On November 30, 2022, ChatGPT was launched - an artificial intelligence chatbot developed by OpenAI and capable of working in dialogue mode, supporting natural language queries
This is the already well-known ChatGPT. By simply writing one sentence in the chat, the user could get: a simple software algorithm, a recipe, a recommendation, an idea, advice, and much more.
A historic day, after which all those who position themselves as IT giants could not help but accept the market challenge and start working on their counterpart.
In March 2023, a more advanced model was integrated – GPT-4.
In this release, the model was trained on a significantly larger volume of data, learned to recognize and work with images. The quality of the answers has improved significantly. It was about GPT-4 that I began to hear positive feedback from programmers: "now it can really automate my routine processes almost flawlessly."
On February 15, 2024, Sora was introduced — a neural network designed to generate short videos based on text descriptions.
This product was also well remembered by users. The internet was filled with videos of amazing quality with a note of their neural network origin. Someone even generated full playthroughs of computer games, by the way, quite successfully.
On May 13, 2024, the GPT-4o model was released, capable of multimodality, faster and more resource-efficient content generation.
There were no revolutionary updates, but the volume and quality of the response also increased.
On July 18, 2024, GPT-4o mini was introduced — a smaller version of GPT-4o, providing faster and more economical operation.
This was the "Golden mean" between GPT-4 and GPT-4o
On September 12, 2024: o1-preview and o1-mini models were launched, designed to improve accuracy in science, programming, and logic tasks.
The latest update from OpenAI at the moment. Better performance, huge datasets, verified answers that the neural network double-checks, internet access – all this is GPT-o1.
On February 6, 2023, Google introduced Bard — an AI chatbot based on the LaMDA language model, designed to answer users' questions in the search engine.
On March 21, 2023, Google provided access to the chatbot to selected users from the USA and the UK.
Little is known about this release, as Google decided not to show the first version to the whole world (as OpenAI did), but to give it only to some users. In fact, an obvious move for the release of such a level of product, because an unsuccessful version can bring significant reputational losses, which, on the scale of Google, are expressed in quite substantial volumes.
In April 2023, Bard developers reported that the service gained the ability to write, debug, and explain code. The chatbot knows 20 programming languages and is connected to other Google products.
After testing on a closed group of users, the company shared brief notes on the functionality of the solution. Taught by the information background around OpenAI, Google focused on promoting capabilities in the field of programming.
In May 2023, Google announced that it had made Bard available to 180 countries worldwide.
And here is the long-awaited moment, the first AI from the long-established IT giant Google is available to all users within the ecosystem.
In December 2023, Bard switched to a new language model — Gemini in the Pro version. In addition to text generation, it can create images.
After long observations of user feedback, fixing shortcomings, etc., the company finally expanded the functionality, taking the next step — image generation.
On May 15, 2024, Google integrated Bard into its main products, including search, providing users with more intuitive and conversational answers.
At this stage, the product was implemented in all key components of Google: assistant, search, call applications (e.g., transcription), etc.
In February 2023, Microsoft presented the updated Bing with AI based on the ChatGPT chatbot at a presentation. The new Bing uses generative AI in its web search function to return results that look like longer, written answers compiled from various Internet sources, rather than a list of links to relevant sites. In addition to the search function, the chatbot can maintain a conversation, generate coherent texts, and answer complex questions thanks to its extensive language model.
On March 14, 2023, Microsoft officially confirmed that the chatbot runs on the GPT-4 language model.
Many wondered: why not their own model? But it is worth considering that Microsoft has entered into active cooperation with OpenAI and supports the company with very solid investments.
On May 4, 2023, Microsoft abolished the waiting list and opened the Bing chatbot to all users.
The company, almost simultaneously with Google, released its AI chatbot to the public. This "added fuel to the fire" in the competitive race and forced companies to actively engage in work and add new functionality faster.
In May 2023, Microsoft announced that it plans to expand the chatbot's capabilities soon: chat history, the ability to upload images to the chatbot, and work with third-party plugins within the chat will be added.
A fairly standard scenario that both OpenAI and Google have applied.
On September 21, 2023, Microsoft introduced Copilot, a universal AI-based tool integrated into Windows, Edge, and other products, providing users with enhanced interaction capabilities.
In this aspect, Microsoft has already distinguished itself from its competitors. They have integrated their neural network into the Copilot AI assistant within the Windows operating system. This allowed for operations on text, browser, and office packages, which, when used correctly, significantly increases the efficiency of working on routine tasks.
On October 3, 2024, Bing Chat was updated using the DALL-E 3 model, allowing users to generate images based on text requests.
Thus, in the West, there are 3 key players in the field: OpenAI and, quite expected candidates, Google and Microsoft. At the moment, the main driver and innovator is OpenAI, they are the first to implement previously unavailable functionality and introduce key innovations to optimize users' routine activities.
Google and Microsoft are more likely to pick up already invented features and implement them on their platforms. Would I say this is bad? – not at all! Creating competitive struggle and different approaches to solving the same tasks – all this together is the fuel and stimulator of the technological race.
These companies also have huge ecosystems, and ecosystems have an incredible number of systems in which it would be very useful to implement artificial intelligence technologies.
Therefore, I am sure that the backlog of tasks for each of the manufacturers is already scheduled for years to come!
Development of modern neural networks in Russia
Yandex
In February 2023, Yandex announced that it was developing its own version of the generative neural network ChatGPT as part of the development of the YaLM (Yet another Language Model) family of language models. The project received the preliminary name YaLM 2.0, which was later changed to YandexGPT.
On May 17, 2023, the company introduced a neural network called YandexGPT (YaGPT), adding a special skill to its virtual assistant "Alice" that allows interaction with the new language model.
Here I would like to draw attention to the date and realize that the domestic "Yandex" is not so far behind the release dates of Google and Microsoft. The difference was only 2 months. But the model could only answer one question and did not take into account the context of the chat.
On June 5, 2023, the Yandex press service reported that the neural network had learned to remember the context of the conversation and ask clarifying questions.
A small but important update that allowed the neural network to keep the context of the current conversation with the user and, when forming a response, rely on the user's past questions.
On June 15, 2023, Yandex added the YandexGPT language model to the "Masterpiece Room" image generation application. Thanks to this, its users were able to create meaningful posts with text, a title, and a relevant illustration.
At this stage, we see the already mentioned thread of the scenario - the creation of a neural network for image generation. "Yandex" did not stand aside.
In July 2023, the YandexGPT neural network became available to businesses for creating virtual assistants and chatbots, as well as generating and structuring textual information.
And this is an important step for the industry as a whole, because "Yandex" gives its API to businesses for implementing the technology in production. Companies are actively using, implementing, and further training the model. Yandex-GPT was implemented as a search engine for the internal knowledge base, included in analytical processes, and other important production tasks. Customers actively shared feedback and results.
On June 27, Yandex reported that the neural network had learned to summarize articles from the internet. The function works with Russian-language texts and articles from the internet up to 30,000 characters in length.
At this stage, Yandex itself began to implement its development into its ecosystem, starting with the browser. The pilot implementation was the ability to summarize (highlight the main points) the content of web pages.
On September 7, 2023, Yandex presented a new version of the language model — YandexGPT 2 at the Practical ML Conf conference. Compared to the previous one, it is capable of solving more types of tasks, while the quality of its answers has improved. According to the developers, YandexGPT 2 answers user queries better than the first version of the neural network in 67% of cases.
After testing the first version and receiving feedback from users and businesses as a whole, Yandex improved the model to version 2. They increased the volume of the dataset and improved the accuracy of the answers.
On September 14, 2023, Yandex began testing YandexGPT for creating quick answers to queries in "Search". The neural network in real-time finds several of the most suitable sources for the user's query, makes a brief summary of each of them, and then selects the most useful answer from the resulting ones. It is displayed under the search bar along with a link to the source.
"Yandex" continues to actively implement the neural network in various aspects of its ecosystem, this time, a specific answer to a question from the search bar.
Since October 6, 2023, YandexGPT can create brief summaries of Russian-language videos on the internet. It summarizes videos ranging from two minutes to four hours in length that contain speech.
One of the revolutionary releases in the Russian neural network market: live translation of foreign videos. Such an implementation, one might say, broke the language barrier and opened up a huge world of English-language content to Russian-speaking users. It is important to mention that this functionality became a completely free part of Yandex Browser and worked excellently even in the first versions.
March 28, 2024: The third generation of YandexGPT language models was announced. The first model in the lineup — YandexGPT 3 Pro — became available on the Yandex Cloud website, improving the processing of complex queries and the accuracy of answers.
"Yandex" continued to further train the fundamental model, releasing a new version 3, but already adding a paid Pro edition (a similar scenario can be observed with OpenAI's paid subscription). A common story with monetizing the flow of private users.
May 28, 2024: A lightweight version of the neural network — YandexGPT 3 Lite, designed for tasks requiring high response speed, such as chatbots and spell checking, was launched.
This was a narrowly focused model that was strongly recommended for specific tasks (spelling, simple dialogue with the user). It solved them faster and more accurately.
October 24, 2024: The fourth generation of language models was introduced — YandexGPT 4, including a powerful Pro model and a lightweight Lite version, with improved command generation and the ability to integrate into Yandex Cloud.
"Yandex" continues to improve the accuracy of answers, the volume of data for training models, and functionality. This time, it was one big release: the basic client version, the Pro client version, the Lite client version, and integration into Cloud for business.
Sber
In April 2023, "Sber" introduced GigaChat — a neural network that can simultaneously generate both text and images based on textual descriptions.
In 2023, GigaChat was created - a generative language model developed by Sber. It was trained on a vast amount of textual data and was designed to assist users in solving various text-related tasks.
Sber also caught the wave of artificial intelligence, starting to develop its neural network based on the ruGPT-3.5 model. The first release was also a generative neural network for working with text.
In March 2023, a more advanced model was integrated – ruGPT-4.
In March, in response to the release of Yandex GPT, Sber released a more powerful version of its product based on the 4th generation model.
In 2023, GigaChat was updated and gained the ability to create images based on textual descriptions.
The company also did not forget about image generation, and within the framework of GigaChat, the Kandinsky bot was introduced, which, by the way, produced high-quality generations from the first release.
In 2023, GigaChat was updated and gained the ability to perform arithmetic operations and solve mathematical problems.
After releasing the main functionality, Sber began adding capabilities to the current generative model, teaching it mathematics and logical thinking.
In 2023, GigaChat was updated and gained the ability to help users write texts such as letters, articles, stories, etc.
The prompt window was increased, the dataset was expanded, and the model was brought to the level of working with full-fledged articles and large volumes of text.
As a result, two players from Russian IT companies have seriously taken up the development of AI solutions: Sber and Yandex. It is important to note that the manufacturers showed a quick response on their part and today they are "breathing down the necks" of Western leaders.
More actively at the moment, Yandex GPT is developing, the company is actively spreading the development into many areas of life: for business (in Yandex Cloud), for the user in the form of a chat (skill in Yandex.Alice), for the user in the form of integration into the browser (Yandex.Browser). Colleagues collect extensive feedback at once and actively continue to correct errors, make new implementations and models.
Sber also shows rapid growth and good performance in the form of GigaChat. The company actively supports the spirit of competition in the market, not only having personal motivation for development, but also broadcasting it to competitors.
Conclusion
The development of neural networks is a rapid and impressive process that is actively taking place both in the West and in Russia. Starting with the early concepts of neural networks in the mid-20th century and simple models, we have moved on to creating complex and powerful deep learning algorithms. The main drivers of progress in the West are OpenAI, Google, and Microsoft. In the Russian market, the development of neural networks is led by products from companies
Yandex and Sber. Russian manufacturers have been able to quickly join the trend, and are already catching up with the level of functionality that Western counterparts have. In the history of modern neural networks, it is important to pay attention to the frequency of updates: it indicates rapid growth and is striking in the volume of updates per year.
Write comment