• Turing Post
  • Posts
  • Russia's Stride in AI: A Deep Dive into 2023's Challenges and Breakthroughs

Russia's Stride in AI: A Deep Dive into 2023's Challenges and Breakthroughs

Unveiling Yandex’s YaLM, Sber's Gigachat, Brain Drain, Hardware Constraints, and the Murky Waters of Censorship and Regulation

Previously: we published the 6-month reports about China (the most famous US rival); India (with the second-largest talent pool for AI and ML); and the UK (which tends to lead Europe). This Foreign AI Affairs series was a huge success!

Today we continue the series with the 8-month report (from January 2023 to August 2023) detailing the transformative impact that the launch of ChatGPT has had in Russia.

Russia

In 2023, Russia's relationship with AI has been multifaceted and complex, especially considering Russia's reversion to the worst aspects of the 20th century, including positional warfare, repressions, denunciations, and paranoia. While the country has been somewhat isolated due to its exclusion from international platforms like ChatGPT, this isolation has not halted local developers.

Let's examine the developments in AI, challenges, business implementation, government relations, and regulatory state:

Local Large Language Models (LLMs) and their Usage

While ChatGPT was breaking records all over the world, Russia was excluded from the AI revolution — the website couldn't be accessed without a VPN, and even if a user managed to reach the website, Russian phone numbers weren't allowed to register. This has led to a cottage industry of services in the spirit of 'import substitution' and 'gray import' – registration of foreign phone numbers, digital credit cards with exorbitant currency rates, and a whole slew of bots that supposedly give access to ChatGPT, Midjourney, and other AI services through a Telegram chat. They're slower, less feature-rich, and no one knows who might be collecting user data, but for many, it's the only way to keep up with modern developments.

Local alternatives exist, with IT giants like Yandex (also known as “Russian Google”) and Sber (the tech behemoth that sprung from the largest national bank of the USSR) releasing their own analogs. The most prominent include:

  • Yandex’s YaLM 100B model, based on GPT with 100 billion parameters, processes text and has been trained on a vast amount of English and Russian texts. YaLM is one of the largest open-sourced GPT-like neural networks (for more info check their github and Medium). The YaGPT (YaLM 2.0, not open-sourced) model has received improvements, including high-quality pre-train data, the use of fine-tuning, a large alignment dataset, and sequence training, resulting in improved model quality, performance, and understanding of context; YaGPT is integrated into the Alisa voice assistant;

  • GigaChat (a reference to the GigaChad meme) from Sber, a multi-modal system integrated into their Salut assistant; Sber Gigachat is based on the NeONKA (NEural Omnimodal Network with Knowledge-Awareness) model, which combines multiple models and possesses multimodality. It can work not only with text but also with other types of information, such as images. Currently, work is being done on models that will enable image analysis and generation in a conversational format.

  • SistemmaGPT, based on "research from Stanford”, which suggests that it has Alpaca at its core, only released for research purposes, with commercial use prohibited.

Most of these technologies seem to have been in development before the breakthrough success of ChatGPT but received extra resources after AI became a global phenomenon at the end of last year. There was also, probably, a lot of yelling involved.

The rapid launch of this service was primarily driven by the availability of developments in the area of transformer models. In recent years, we have trained a number of our own fundamental models, such as ruGPT-2, ruGPT-3, ruGPT-3.5, ruT5, ruBERT, ruRoBERTa, ruDALL-E Malevich, Kandinsky, mGPT, FRED-T5, etc.

For the Salut family of virtual assistants, we have developed a range of dialogue models based on modern generative language models. While working on these models, we have extensively researched RL techniques in NLP, developed mass markup tools and processes, accumulated large amounts of high-quality data, created an infrastructure for training and inferencing large models and, most importantly, organized strong research and development teams. All of this, combined with our company's culture of agile development and focus on technological innovation, allowed us to realize an ambitious new project within a very short timeframe.”

Sergey Markov, Chief of the R&D Division at SberDevices, specially for TuringPost

Challenges

The vast amounts of resources being deployed to advance AI in Russia are confronting several key obstacles that threaten the future development of the field, particularly concerning resource-intensive general-purpose public models. These challenges include:

  • Brain Drain: Various estimates suggest that between 500,000 and 1.2 million individuals have emigrated from the country, a significant number of whom are from high-tech sectors.

  • Hardware Constraints: The lack of access to bulk ordering of AI-specific GPUs cripples the pace of innovation and experimentation.

  • Censorship Constraints: The models must be designed to consider censorship, leading to inherent limitations. Most Russian chatbots, for example, refuse to discuss the war (or the "special military operation," the official euphemism for the Russian invasion of Ukraine), do not respond to questions about opposition figures, and exhibit biases such as refusing to render images of Ukrainian soldiers while agreeing to draw Russian ones.

These stumbling blocks not only limit the potential of AI in Russia but also cast a shadow on the ethical integrity of the nation's technological advancements, calling into question the objectivity and freedom of the technology within the country's borders.

If you find this post insightful, please consider sharing it via your social networks: https://www.turingpost.com/p/russiaai2023 Copy and and share! 🤍 

AI in Business

First, a few words on the state of business in Russia by the second year of the war. Essentially, the government is engaged in a war that is costly both financially (Reuters estimates that the expenditure for 2023 will amount to $100 billion, about a third of the entire public budget) and in terms of human capital (it's hard to get a trustworthy estimate due to the 'fog of war', but the US Joint Chiefs of Staff assume 300,000 casualties on the Russian side, with somewhere in the ballpark of a million leaving the country).

It's the extremely adaptable Russian entrepreneurs who are keeping the economy from fully spiraling out of control. AI is also used as a tool to counterbalance the lack of human resources.

Retailers like X5 use AI to replace human data analysis at basically every level, from logistics and warehousing to user experience and personal promo. They claim that if the quality of store replenishment forecast grows by just 1%, then due to the effect of business scale, the introduction of new technologies will already pay off. Digital storefronts like Ozon (for businesses) and Avito (for individuals) use AI to moderate product descriptions, leaving the more difficult cases for human moderators.

One of the key industries where AI is most actively being integrated is medicine. According to the National Center for the Development of AI, 16% of healthcare organizations in Russia are already implementing AI, but 78% of them are in the pilot project stage. The report cites a variety of trends, including data extraction from real-world data; the development of mobile health (mHealth); predictive analytics and the creation of digital twins for modeling dynamic processes in the patient's body; the use of AI to train nursing staff; the development of surgical robots for healthcare; the use of AI to develop individual health insurance programs and facilitate the routine work processes of medical workers.

It is hard to ascertain whether these trends are significant or even real, as this information comes from official government sources, who have a vested interest in promoting this.

Another popular use of AI is in the financial sector, where lending approval or refusal has been mostly automated, taking minutes instead of days, according to Vedomosti.

When considering all the above-mentioned tools, it is unclear whether there are safeguards in place to protect marginalized communities from confirmation bias. There doesn't seem to be much conversation on the topic, and official comments rarely touch on this.

Government and Surveillance

The Russian government has shown interest in AI for many years now. In 2017, President Vladimir Putin stated that whoever becomes a leader in AI will rule the world. Two years later, Putin signed a national strategy for the development of artificial intelligence. In 2021 alone, the government allocated 1.4 billion rubles for grants to AI developers. At the same time, plans were announced to create a directorate for the development of artificial intelligence under the Ministry of Defense (and, apparently, it really worked). The head of the ministry, Sergei Shoigu, said that they were producing combat robots capable of working without operators.

Back in April, an analytical summary of current digital repressions published by the Re:Russia think tank stated that there were several systems in development for monitoring the Internet for suspicious activity:

The ‘Vepr’ system will look for ‘information pressure points’ and forecast where and when protests might erupt. ‘MIR’ will handle fully automated searches for prohibited content. ‘Oculus’ is designed to identify calls for protests in images and videos, and to recognize the faces of demonstrators. However, according to the very same ‘leaked’ Roskomnadzor papers, none of these systems are yet to be operational. Their development has been hampered by sanctions and a lack of skilled specialists who are willing to work for the Russian government.

The ultimate goal seems to build an ‘indestructible’ authoritarian state through the use of AI-enabled digital and real-world surveillance.

There are more quirky applications of AI in Russian politics. The Liberal-Democratic Party of Russia (neither liberal nor democratic in reality), recently lost their long-time leader, Vladimir Zhirinovsky. He had played the role of a token oppositional figure, using inflammatory rhetoric to act as a lightning rod to gather lower-class protest votes in every presidential election since 1994. At the most recent St. Petersburg International Economic Forum, which used to be a Russian Davos but has deteriorated over the last two years, the NeuroZhirinovsky neural network was presented. It is a chat model trained on the vast amounts of texts and speeches of the right-wing politician, and there's even talk of accepting the AI as an official member of the party. The simple reality is that after losing their leader, the party couldn't find a worthy replacement and had to fall back on this toy to bring themselves back into the spotlight.

There are some who suggest that this is in line with President Putin's own dreams of digital immortality, an alpha-test of sorts, but the closed-off nature of the Russian state makes it impossible to ascertain whether this is true or merely a conspiracy theory.

AI Regulations

In Russia, there exists a concept for the development of regulations in the field of AI and robotics until 2024. It encompasses issues of data circulation regulation, legal liability, information security, and the potential application of AI in various sectors of the economy. Among the proposed regulatory measures, there are the regulation of AI-related activities, the introduction of labeling for content created by neural networks, the determination of legislative responsibility for the use of AI, considering data circulation in the usage of AI and robotics, as well as ensuring information security.

However, there are unresolved questions that require further study and development. The issue of ownership rights for content created using AI remains open. It is also undetermined whether content labeling will be useful for rights holders and how it should function. Questions arise regarding the ownership of copyrights for content created using AI and the need for amendments to legislation to update existing norms and establish accountability for the use of AI for criminal purposes. It seems though, that since the war in Ukraine, the development of legislation and regulation in the field of AI is somewhat put on hold.

Conclusion

This pause in legislative activity, paired with the accelerated pace of AI integration across various sectors, creates a nebulous scenario. The implementation of AI solutions continues to grow, and their real-world implications are playing out. Yet the infrastructure for understanding, governing, and critically evaluating these developments lags behind, laying bare a future that is simultaneously promising and uncertain.

Written by Vasily Sonkin with additional reporting by Varvara Novozhilova

UPDATES:

Sept, 2023:

Lomonosov Moscow State University has launched its MSU-270 supercomputer, boasting a peak computational power of 400 'AI' PetaFLOPS. Optimized for AI and high-performance computing, it is built on unspecified "latest graphics accelerators" and is part of MSU's 2030 advancement plan. The machine will support interdisciplinary research and will be pivotal in training AI specialists. Given U.S. export rules on GPUs and Russia's lack of GPU production, the hardware origin remains a topic of speculation.

Reply

or to participate.