Did you know? The 5 Most Impressive AI Achievements in 2023

The boom in new language models

‍

Large language models (LLM, large language model) learn from huge amounts of texts, find patterns and logical connections in them that are inaccessible to humans, and later use this experience to generate new texts and even make predictions. ChatGPT is the most illustrative but not the only example of a generative language model.

In 2023, many tech giants first introduced their own language models or released chatbots that run on LLM.

‍

Claude/Claude 2

In January 2023, independent startup Anthropic introduced its own chatbot Claude, which was immediately named one of ChatGPT's main competitors. The first Claude worked as a digital assistant and could program in several languages, but at the same time surpassed the competition in terms of the size of the context window — text that he could process at a time. In July 2023, Claude 2 was released, which can be used not only through the API, but also through the website, plus the chatbot's performance has increased.

‍

Bard/Gemini

In May 2023 Google introduced its own chatbot Bard, which uses Palm 2's own language model. Unlike Claude, it supports multimodal queries that contain not only text but also images. In February 2024, Bard became Gemini.

‍

ERNIE Bot

In March 2023, Baidu, the leader among Chinese search engines, introduced its own ERNIE Bot chatbot. Back then, the service was called a Chinese alternative to ChatGPT: it has a similar web interface, a similar plugin system, and similar functionality. At the end of the first quarter of 2024, ERNIE Bot already had more than 200 million requests per day.

‍

Q

Amazon introduced its generative chatbot called Q. It is interesting because it is built into the Amazon ecosystem by default and is available to users of AWS cloud services. This largely determined the specifics of his work: the chat answers questions from developers, provides instructions and other documentation, helps fix problems with server errors, and can be integrated with a code editor or IDE.

‍

GPT-4

OpenAI didn't plan to fall behind the competition and released the fourth model of its chatbot in March 2023. It supports plugins that search for news on the Internet and external services, plus the context window has grown significantly (up to 100,000 words in English). By the way, ChatGPT-4 is available by default to users of Bing Chat, Microsoft's search engine.

In parallel with creating new language models and improving existing ones, major market players are working to reduce the number of machine hallucinations and limit the generation of unethical content.

‍

Extending local language support in LLM

The most popular generative AI models were initially adapted to English queries. Many of them support answers in other languages, but they are often less accurate. Although existing models cover more than 100 languages, this is not enough: the world uses 7000+ languages and adverbs (along with those that are on the verge of extinction).

In 2023, a major step was taken towards solving this problem. Meta AI announced the launch of the Massively Multilingual Speech (MMS) project and has begun building a dataset for training language models in more than 1,100 new languages. This should make it easier to access information and use digital devices in your preferred language, even if it is only spoken by a few thousand people.

It is expected that in the near future, one model will be able to solve several speech problems for all languages, which will lead to an even greater distribution and increase in the performance of generative chatbots.

‍

Dissemination of multimodal language models

2023 was a breakthrough year for multimodal AI models, although they were first talked about loudly in 2022. Multimodal AI combines and collaboratively processes various types of data: text, graphics, audio and video, using unique neural networks for each type of content. Multimodal models are considered to be more accurate, and the ability to analyze data from several sources makes them a universal tool in human hands.

Some of the generative language chats listed above have multimodality support. For example, a Gemini multimodal model can be shown a photo with a plate of cookies and asked to write its recipe without giving details. At the exit, you will get a cookie recipe whose names you don't even know.

Multimodality in digital assistants makes it possible to improve the personalization of responses, use advanced visual search, and train AI models using different data formats. This is already in demand in medicine, transport, when creating modern multimedia content and, in general, for analyzing data from various sources.

‍

The first serious step towards reading minds using AI

In 2023, several major studies were conducted that addressed the topic of visualizing human thoughts. For this purpose, data obtained as a result of brain scans and processed using artificial intelligence were used.

In May, the University of Texas at Austin (USA) introduced a system to help people who cannot speak but are mentally healthy. Scientists measured human brain activity during an MRI study and decoded their thoughts, and with fairly high accuracy. For example, if a person said in their head the phrase “I don't have a driver's license yet,” the smart system decoded it as “She hasn't even started learning to drive yet.” Although the results can hardly be described verbatim, this was a big step forward compared to what was before.

Scientists at the University of Sydney (Australia) are moving in the same direction. At the end of December 2023, they developed a portable system that uses the same approaches used to study the brain using electroencephalography (EGG). A special portable device reads and then converts brain signals into language units, without the need to scan the brain with an MRI or implant electrodes into the brain, as was done before.

The scientific journal Nature published an article about how scientists finally managed to combine human brain tissue with electronic components. The research resulted in accurate speech recognition and even solving mathematical problems. For this experiment, human brain tissue was specially grown from stem cells in the laboratory. The materials obtained were used to create the Brainoware device, which works much more accurately than a computer with artificial intelligence support.

In addition, IBM has announced an analog chip for AI workloads. The uniqueness of this 64-core processor is that it is created by analogy with the human brain.

‍

Restoring limb sensitivity using AI

Researchers at the Feinstein Institute (USA) were the first to use a double neural shunt, which helped restore limb sensitivity in people suffering from complete or partial paralysis of the extremities. In a unique clinical study, impressive results were obtained. AI microchips were implanted into the brain of a paralyzed person, which made it possible to re-connect the brain to the body and spinal cord, which helped restore arm mobility and sense of touch, and the effect persisted even outside the laboratory. This happened for the first time in medical history. The operation to implant the microchip took place for 15 hours on the open brain of a paralyzed person. Its results give hope that artificial intelligence technologies can significantly improve the quality of life of paralyzed people.

‍

What's next for AI?

According to Collins Dictionary, “artificial intelligence” has become the most prominent concept of 2023. It is obvious that this technology will continue to dominate in the future, and the next few years will bring several more revolutionary discoveries made with the help of AI. 2023 was a turning point in the evolution of AI and laid the foundation for the emergence of new, better intelligent tools.

Although it is too early to sum up the results of 2024, the world will already remember it for its new achievements in the field of AI. For example, it is predicted that in just one year the number of visual effects created with its help will increase significantly. Self-learning AI systems will learn how to make scientific discoveries and develop new drugs, and previously unknown companies that develop AI systems will become one of the world's top tech giants.