Applied AI: Development Challenges and Risks for Russia

The leaders in the race for the best-performing AI, measured by maximum accuracy and minimum flawed responses, are constantly changing. In the author’s assessment, OpenAI’s ChatGPT and Anthropic’s Claude currently share first place for most applied tasks. Chinese models are playing catch-up, and according to information available to the author, even Chinese software developers often use U.S. solutions. Russian solutions are in an even weaker position. This lag is primarily due to the technology gap and stronger competition in the U.S., while Chinese and Russian companies follow the lead of U.S. companies, often borrowing their solutions from publicly available academic publications and open-source projects. In addition, U.S. export controls and sanctions limit access to the highest-performing graphics processor unit (GPU) models for competing states, causing them to fall further behind.

Russian developers are facing tightening U.S. restrictions on advanced AI models, with OpenAI and Anthropic systematically blocking Russian user accounts. Additionally, September 2026 saw a rise in access limitations in respect to abroad AI services enforced directly from within Russia. To circumvent these restrictions, many Russian users are turning to unauthorized software to access U.S. services. On the one hand, this complicates production processes and creates a risk of missed project deadlines if restrictive measures become more stringent. On the other hand, the widespread use of foreign models amounts to nothing less than a mass, unauthorized cross-border transfer of Russian entities' internal information and intellectual property to businesses affiliated with the U.S. government and military institutions.

The high energy and resource demands of advanced cloud-based models and the risk of information leaks can be addressed in part by switching to domestic cloud-based alternatives, available “lightweight” versions of “open-weight” models, and specialized “open-source” solutions. In the latter case, the risk of information leaks can be minimized by deploying models in an isolated environment with strict adherence to cybersecurity protocols. Unfortunately, the cost of this approach would be a substantial decline in process quality and performaтсe.

In many scientific fields and in the development of new technologies, including hardware and software, advanced AI models have become an indispensable tool for increasing productivity, reducing severalfold the time spent on intellectual work per unit of scientific output—provided, of course, that users retain the ability to critically assess AI-assisted results. These capabilities continue to grow: in August 2026, Anthropic introduced a standard, tested in practice, for integrating agents based on large language models (LLMs) with a wide range of laboratory and manufacturing instruments to autonomously carry out the full research cycle in the physical world.

AI is also transforming applied fields such as software development. The main conclusion emerging from discussions at industry conferences is that, in the future, code will be written exclusively by AI agents, while people will be responsible for defining tasks, overseeing agents, and reviewing code. Accordingly, the software development industry will need managers and analysts to define tasks, as well as experienced programmers to review code. Similar conclusions are drawn by the authors of the recently published report “The Hidden Cost of AI: The Software Development Process in the Age of Artificial Intelligence.” As one participant at a recent conference put it: “A revolution in software development has just taken place. We weren’t expecting it, and we don’t know what to do next.”

A number of researchers, including Russian researchers E. Vityaev, S. Shumsky, and A. Kolonin, point to the possibility of creating AI architectures that are more resource-efficient than LLMs and also offer interpretability, unlike those based on deep neural networks and transformers. This latter property may be particularly important for building “transparent” or “trustworthy” AI systems for critical infrastructure facilities and tasks. This requires applied research to develop and test such domestic alternative architectures, including hybrid “neuro-symbolic” architectures, as potentially cost-effective and interpretable alternatives to existing foreign LLMs.

Import substitution in software and hardware development requires the creation of domestic frameworks based on LLM agents, certified and included in the Russian Software Registry, that support the full software development cycle, including specification development and validation, code and test generation, as well as code verification and review. These frameworks must be able to use both cloud services, where justified by the balance between efficiency and security, and local models deployed in an isolated environment to prevent cross-border information leaks.

Ensuring the security and reliability of both neural network models and software written with their assistance requires the development of formal verification methods for both software and AI models themselves, building on domestic advances in this field. Developing this technology could substantially ease the burden on the “bottleneck”—the highly qualified specialists needed to review and validate code written by AI agents—and potentially make AI models themselves more “transparent” and “trustworthy” by using tools to identify their areas of competence and applicability.

To prevent information and data leaks in industry and business, unauthorized use of foreign cloud-based AI services, including through tools that circumvent access restrictions, should be treated as a cybersecurity threat. Addressing this threat requires either providing competitive domestic and locally deployed alternatives, where available, or arranging special routing and filtering for traffic that is formally unauthorized but necessary for manufacturing and operational tasks. Until the first problem identified in this list—import substitution for LLM agents—is resolved, it is advisable to avoid additional country-wide restrictions on foreign AI services on Rissian side. AI security issues should be addressed through targeted measures on a case-by-case basis at the level of individual companies and institutions.

What Types of Applied AI Are Being Discussed?

The main areas of development in modern applied artificial intelligence (AI) and their associated threat vectors can be classified into four main groups.

First are centralized cloud-based systems built on so-called “large language models” (LLMs)— being developed in U.S. (OpenAI ChatGPT, Anthropic Claude, Google Gemini, Llama, Grok) and China (DeepSeek, Qwen, Kimi, GLM), whose full capabilities are usually available only through web services hosted outside of Russia. According to certain performance rankings, the Russian systems being developed by Sber (GigaChat) and YandexGPT (Alice AI) currently lag behind foreign services. The main part of this article will thus focus on the revolution that has taken place in this segment over the past few years and the opportunities and risks it presents.

Second are autonomous robotic systems capable of carrying out assigned tasks without an operator; for example, under conditions of interference or when communication with a control center is impossible in principle or impractical in a given situation. Such systems and their requirements were discussed in an earlier review. The relevance of this field is underscored by recent news—in particular, the September 2026 announcement that production of a new type of autonomous attack drone was beginning in Germany. Based on the limited information available, such systems remain far from perfect, but they are being actively designed and developed despite the unsuccessful efforts of several advocacy groups to establish an international ban on the development of lethal autonomous weapon systems (LAWS). Currently, the main obstacle to their development is the resource demands of modern neural network-based AI architectures, which require substantial memory and consume large amounts of energy. Overcoming these constraints through new computing architectures, such as neuromorphic architectures, will make the challenges posed by autonomous systems even more pressing.

Third are so-called “swarm” systems built on the autonomous systems mentioned above. These systems allow autonomous agents (for example, aerial, ground, surface, and underwater drones) to carry out assigned tasks not only autonomously but also in cooperation with one another, without operators. According to publicly available information, both autonomous and swarm systems are being actively developed, primarily in China, which is clearly counting on them in likely future conflicts.

Fourth are hybrid systems, in which the limited intelligence of systems in the second and third groups is compensated for or supplemented by control from a centralized system in the first group, provided there is a connection to the controlling data center, with or without human involvement. Most systems today fall into this category. The best-known system using advanced LLM technology is the U.S. system Palantir.

Anton Kolonin:
Growth Paths for the Coming Superintelligence

Which LLMs Do We Know?

Now to examine the current state of the first category of systems, which use “transformer” architectures based on deep neural networks. The highest-performing foreign models can contain hundreds of billions of parameters—more than the number of neurons in the human cerebral cortex (approximately 100 billion), but fewer than the number of synaptic connections, which is several orders of magnitude higher.

Efficiently processing queries and generating responses with neural networks, especially during training, requires powerful clusters of specialized graphics processing units (GPUs) capable of highly parallel computation. Only U.S. companies—Nvidia, AMD, and Intel—produce these processors with the necessary technical specifications. Chinese companies such as Huawei are attempting to develop alternatives but are still trailing in this race. Nearly all well-known and widely used AI systems today, including Chinese systems, rely on GPUs from a single U.S. company—Nvidia.

The leaders in the race for the best-performing AI, measured by maximum response accuracy and minimal flawed responses, are constantly changing. In the author’s assessment, OpenAI’s ChatGPT and Anthropic’s Claude currently share first place for most applied tasks. Chinese models are playing catch-up, and according to information available to the author, even Chinese software developers often use U.S. solutions. Russian solutions are in an even weaker position. This lag is primarily due to the technology gap and stronger competition in the U.S., while Chinese and Russian companies follow the lead of U.S. companies, often borrowing their approaches from publicly available academic publications and open-source projects. In addition, U.S. export controls and sanctions limit access to the highest-performing processors for competing countries, causing them to fall further behind.

The high energy demands of neural network models also affect the cost of services for end users: using cloud-based models can cost anywhere from tens to tens of thousands of dollars per project, depending on usage patterns (see the estimates for intensive model use presented at the AGI Russia seminar).

Russian developers face U.S. restrictions on access to advanced models: OpenAI and Anthropic systematically block Russian user accounts. In September 2026, cases of access restrictions imposed from within Russia were also recorded. To circumvent these restrictions, Russian users of U.S. services are compelled to use unauthorized software tools on a large scale. On the one hand, this complicates production processes and creates a risk of missed project deadlines if restrictive measures become more stringent. On the other hand, the widespread use of foreign models amounts to nothing less than the mass unauthorized and illegitimate cross-border transfer of Russian entities’ internal information and intellectual property to businesses affiliated with U.S. government bodies and military institutions.

The high energy and resource demands of advanced cloud-based models and the risk of information and data leaks can be addressed in part by switching to domestic cloud-based alternatives, available “lightweight” versions of “open-weight” models, and specialized “open-source” solutions. In the latter case, the risk of data leakage can be minimized by deploying models in an isolated environment with strict adherence to cybersecurity protocols. Unfortunately, this approach would come at the cost of a substantial decline in the quality and performance of the processes involved.

What Do These LLMs Offer?

A previous review, published two years ago, mentioned an analytical report by former OpenAI employee Leopold Aschenbrenner predicting that AI technology would reach the threshold of self-improvement within three years —a stage at which AI could develop its own code without human assistance, more efficiently and at an incomparably greater speed (see the detailed discussion of the report). Once AI reaches this stage of development, countries—or national corporations—with the relevant technologies will gain an unquestionable geostrategic advantage. Former U.S. Secretary of State Henry Kissinger and Google founder Eric Schmidt, now one of the leading executives involved in developing military AI in the U.S., made a similar prediction even earlier in their book.

Notably, judging by its style, Aschenbrenner’s report appears to target decision-makers in the U.S. Department of State and the Department of War. It proposes not only ending the open publication of AI research findings altogether—a step leading U.S. companies have already partly taken—but also restricting foreign travel by specialists in the field, which China implemented in 2026. Moreover, to secure a U.S. strategic advantage in AI, the report proposed lifting restrictions on domestic oil production, a measure already being put into practice, alongside prioritizing the development of nuclear energy to power data centers.

Some experts and commentators are skeptical of such forecasts, pointing to the so-called “AI bubble” and predicting that it will “burst,” much like the “dot-com bubble” in 2001. Others emphasize the energy inefficiency of existing solutions, describing them as “dinosaurs doomed to extinction” that will be replaced by more economical and advanced technologies.

However, it is already clear that in many scientific fields and in the development of new technologies, including hardware and software, advanced AI models have become an indispensable tool for increasing productivity, reducing severalfold the time spent on intellectual work per unit of scientific output—provided, of course, that users retain the ability to critically assess AI-assisted results (see the discussion of modern AI’s capabilities for automating scientific work). These capabilities continue to grow: in August 2026, Anthropic introduced a research preview for the Model Hardware Standard (MHS) for integrating LLM-based agents with a wide range of lab and manufacturing instruments, to autonomously carry out the full research cycle in the physical world.

AI is also transforming applied fields such as software development. The main conclusion emerging from discussions at industry conferences this year is that code will be written exclusively by AI agents in the future, while people will be responsible for defining tasks, overseeing agents, and reviewing the code. Accordingly, the software development industry will need managers and analysts to define tasks, as well as experienced programmers to review code. The authors of the recently published report “The Hidden Cost of AI: The Software Development Process in the Age of Artificial Intelligence” reach similar conclusions. As one participant at a recent conference put it: “A revolution in software development has just happened. We weren’t expecting it, and we don’t know what to do next.”

Anton Tamarovich:
Rules or Practice: New Turning Points in Global AI Governance

What Problems Does This Create?

Now to examine the potential risks and threats in software development and extrapolate them to electronic product development, analytical and scientific work, and decision-making in business, public administration, and other fields. Suppose AI increases productivity severalfold, eliminating the need for engineering and technical personnel with low or intermediate skill levels in production processes, while requiring only highly skilled specialists capable of assigning tasks to AI agents and accepting their completed work. But in such a situation, how will engineers and technical specialists—programmers, for example—develop advanced expertise if they have not worked their way up from entry-level development roles and gained the corresponding hands-on experience?

A familiar problem at software development companies is the shortage of staff who remain enthusiastic about reviewing, correcting, and refining other people’s code—most specialists would rather “write their own.” When LLM-based agents begin generating vast amounts of code at tremendous speed, with quality exceeding that of the average programmers the company lets go, the remaining skilled staff face a “backlog” of project reviews, acceptance checks, and revisions. The shortage of highly skilled staff creates a “bottleneck” in project work. In 5, 10, or 20 years, when the remaining specialists leave the company for various reasons, there will be no one to replace them, because those who could have progressed to that level from junior positions will already have lost their skills.

The situation is compounded by the persistent problem of so-called “hallucinations,” or flawed solutions proposed by LLM-based agents, even those using advanced models. One practical solution is iterative development, in which an “agent” produces successive versions of the code while a “senior programmer” controls the process toward a flawless version. Even with this approach, however, the speed and volume of code variants generated by the “agent” make the “senior programmer”—the reviewer—the “weak link.”

The opacity and lack of interpretability of existing neural network models pose a more general problem for their use. On the one hand, AI is becoming so convenient and practical that it is starting to replace people in manufacturing processes. On the other hand, the boundaries of what neural network models know and do not know remain unclear in practice, and their “hallucinations” cost businesses tens of billions of dollars a year. One approach to addressing this problem is the formal verification of neural network models or a code generated by them, a field actively developing in the West. In software development specifically, formal code verification methods may help address hallucinations—an approach being pursued by Amazon, among others.

Another problem is cross-border information leakage. This is probably less critical for academic research, since “leaked” ideas, documents, and code become part of the public body of scientific knowledge in any case. Even in science, however, there is a potential risk that the “other side” could claim priority in a public scientific and technological development.

When handling sensitive corporate or government information, a Russian employee or official faces a choice: either a) deliberately reduce their productivity by forgoing new technologies or limiting their toolkit to domestic models or “lightweight” versions of U.S. or Chinese models, or b) use “questionable” (from security perspective) tools to circumvent access restrictions and work “at the cutting edge,” while “leaking” work documents and code to servers in U.S. or China.

A similar situation emerged two decades ago with social media, when personal data aggregators, primarily in the U.S., learned to monetize personal data and content generated by social media users worldwide by selling user behavioral profiles. It took two decades for the significance of the problem to be recognized internationally and for individual states to begin taking steps to protect user data, seeking to shield their citizens from the commercialization of their data by transnational corporations and from potential manipulation of their behavior through these online services. Now, any professional activity may become a source of data for aggregation by U.S. or Chinese corporations, including work performed by employees of companies and government institutions using internal documents and technical information in the workplace.

Notably, enterprise licenses for cloud-based AI services in the U.S. include commitments by providers not to use user data and prompts to train subsequent versions of their models. However, the current practice of accessing these services in Russia on an individual basis rules out confidentiality protections even in theory, let alone in practice.

Arevik Martirosyan:
Global Governance of Artificial Intelligence: The Birth of a New Architecture or Fragmentation of the Digital Agenda?

What Should Be Done?

Back in 2017, Vladimir Putin stated the need to take a leading position in artificial intelligence. Today, simply keeping pace and avoiding technological dependence on counterparties in the “AI race” requires import substitution for AI technologies as critical infrastructure—including hardware, software, and the models themselves—a task that has long been necessary.

A number of researchers, including Russian researchers E. Vityaev, S. Shumsky, and A. Kolonin, point to the possibility of developing AI architectures that are more resource-efficient than LLMs and also offer interpretability, unlike architectures based on deep neural networks and transformers. Interpretability may be particularly important for building “transparent” or “trustworthy” AI systems for critical infrastructure facilities and tasks. This requires applied research to develop and test such domestic architectures, including hybrid “neuro-symbolic” architectures, as potentially cost-effective and interpretable alternatives to existing foreign LLMs.

Import substitution in software and hardware development requires domestic frameworks based on LLM agents, certified and included in the Russian Software Registry, that support the full software development lifecycle, including specification development and validation, code and test generation, and code verification and review. These frameworks must be able to use both cloud services, where justified by the balance between efficiency and security, and local models deployed in an isolated environment to prevent cross-border information leaks. One such framework was proposed specifically for very-large-scale integration (VLSI) chip design at the Microelectronics Forum in 2025.

Ensuring the security and reliability of both neural network models and software developed using them requires further development of formal verification methods for software and AI models themselves, building on domestic work in this field. Advances in this technology could substantially ease the burden on the “bottleneck”—the highly skilled specialists needed to review and validate code written by AI agents—and potentially make the AI models themselves more “transparent” and “trustworthy” by using verification tools to identify their areas of competence and applicability.

Improving the efficiency of interaction between AI agents and expert analysts within the framework mentioned earlier requires further practical development and broader application of software development methods based on formal specifications—semantic models of applications—as well as technologies for executing these specifications directly. This approach is widely used in autonomous process control systems (APCS). Specifications based on the IEC 61131 and IEC 61499 standards can be used both to generate code and for direct execution by an interpreter; this approach can also be applied in other fields. Russia has specialists with experience developing such standards and systems for autonomous process control systems (APCS), business applications, and other domains.

To prevent data leaks in manufacturing and business, unauthorized use of foreign cloud-based AI services, including through tools that circumvent access restrictions, should be treated as a cybersecurity threat. Addressing this threat requires tow things at the same time: providing competitive domestic and locally deployed alternatives, where available, and arranging special routing and filtering for traffic that is formally unauthorized but necessary for manufacturing and operational tasks using foreign services, if permittable.

Until the first problem identified in this list—import substitution for LLM agents—is resolved, it is advisable to avoid introducing additional nationwide restrictions on the use of foreign AI services. AI security issues should be addressed through targeted measures on a case-by-case basis at the level of individual companies and institutions.

About the Author

Anton Kolonin, Leading Researcher at the Novosibirsk State University (NSU) AI Research Center, Ph.D. in Technical Sciences, and Academic Advisor to the Russian Academy of Engineering. He has been developing artificial intelligence software systems for various purposes in both foreign and domestic companies for over 40 years. Since 2013, he has been a member of the international community for the development of Artificial General Intelligence (AGI Society). In 2025, he chaired a program committee of the leading international conference on this topic. In 2015, he organized a Russian-speaking online community of AGI developers (AGI Russia) and has been conducting weekly online seminars on the subject for more than six years. Since 2018, he has been preparing annual AI reviews for RIAC. From 2022 to 2024, he served as the technical lead for a project to replace all imported software in autonomous process control systems (APCS) stack for the company’s critical infrastructure.