How AI helps to write software. Overview of one of the most promising technologies of the future

Terminology

AI-augmented software Software Engineering is a term that was introduced Gartner in 2020 to describe the process of using technologies (artificial intelligence for example, machine learning natural language processing, etc.) to speed up application development cycles and. DevOps A number of vendors have already released products for this type of work. Among them are Codota, Deep Code,, Kite Google,,, Mendix Microsoft OutSystems,. Parasoft

by 2022, at least 40% of new software development projects will use the resource of virtual AI developers.

However, artificial intelligence and machine learning technologies have been used by software companies and departments around the world for about five years. In 2020, Gartner first included this concept in the list of the most promising technologies of the future - Hype Cycle for Emerging Technologies (technology maturity cycle).

Areas of application of artificial intelligence in software development

AI is used in the development of software solutions at the following stages:

Collection of technical requirements

Digital assistants analyze documents with collected requirements, indicate disagreements in the text, inconsistencies in numbers, units of measurement, amounts and offer possible solutions.

Rapid prototyping

Converting business requirements to code typically takes months or even years. However, machine learning greatly reduces this process, allowing those with less experience to use natural language or visual interface development techniques to create a prototype.

Coding

In the process of writing code, an AI-based autocomplete system offers recommendations for completing lines of code. Intelligent assistants reduce code creation time by 50%. In addition, they can recommend referring to related documents, best practices and giving code examples.

Error Analysis and Handling

The virtual assistant can learn from past experience to identify typical errors and automatically flag them during the development phase. Machine learning can be used to analyze system logs to quickly and even proactively detect errors.

Automatic code refactoring

Pure code is required for collaboration and long-term maintenance. As the company develops, software solutions can change, and the question of how to modify the code for better application performance arises. Machine learning is used in this case to analyze the code and automatically optimize the code for easy interpretability and performance.

Testing

Automated testing systems use AI not only to start the testing process, but also to create test cases.

Commissioning

Sometimes errors in the program code become explicit only after the software is put into operation. But AI tools prevent such situations by checking the statistics of previous releases and application logs.

Project Management

Software development sometimes goes beyond budget and schedule. Advanced analytics systems make it possible to use data from a large number of software development projects to predict technical tasks, the necessary resources and time for the project. Machine learning can extract data from past projects, such as user histories, feature definitions, estimates, and actual conditions, to more accurately predict workloads and budgets.

2022: Definition of the boundaries of interaction between product management and AI: expert view of Nikolai Prokofiev

One of the main trends that can change product management, like many other areas of life, has become the use of artificial intelligence. Nikolai Prokofiev, the initiator of the creation of a service for online payment Tapper, which made it possible to automate the restaurant business, talks about what role artificial intelligence plays in the life cycle of creating and accompanying a product and how much it can be trusted. Read more here.

Development of AI-based tools for software development

According to a 2020 study by Deloitte, there is a great interest in the market in solutions that use artificial intelligence tools in the field of software development:

• In recent years, dozens of products have appeared on the market that use AI to improve the efficiency of software solution development.
• The volume of annual investments attracted startups that offer AI products for software development amounted to $704 million USA by September 2019.
• The global custom software development market is projected to grow to US $61 billion by 2023.
• Driven by growing software demand, the number of development engineers will grow by 21% in 2028.

Market Experts Opinion

2025

4x more code, but 10x more security issues. How neural networks affect programming

Programmers using AI assistants to write code show three to four times higher performance compared to manual colleagues, but the number of security problems in their code increases tenfold. Apiiro, which specializes in application security analysis, analyzed tens of thousands of repositories with code from Fortune 50 developers and identified critical risks of using artificial intelligence in programming. Apiiro product manager Itai Nussbaum described the situation with the phrase: "AI fixes typos, but creates bombs with clockwork." This became known on September 5, 2025.

According to The Register, the study covered program code created Claude using the Code AI tools from, Anthropic-5 GPT from OpenAI and 2.5 Gemini Pro from. Google The analysis showed a sharp increase in the number of security problems in the code - from December 2024 to June 2025, their number increased tenfold.

Neural networks significantly increase the amount of program code, while repeatedly increasing the number of potential vulnerabilities.

Apiiro understands security problems not only as direct vulnerabilities, but also the use of risky programming practices. These include using open source code as dependencies, using insecure patterns, and storing sensitive data, including API keys, directly in program code.

Nussbaum noted that artificial intelligence generates not one specific type of vulnerability, but all types of security problems at the same time. According to him, company leaders need to understand: the introduction of the mandatory use of AI when writing code should be accompanied by the parallel introduction of AI tools for analyzing the security of applications.

The study revealed positive aspects of using AI assistants. The use of neural networks allows you to reduce the number of syntax errors by 76% and logical errors by 60%. Automated systems effectively fix typos and help programmers avoid simple technical errors.^[1]

Schaller experiment: Russian developers assessed the prospects for AI in transferring old code to new technologies

At the end of July, Christian Schaller, who heads the Fedora Desktop Team and the desktop development group at Red Hat, announced^[2] the results of an experiment to study the capabilities of artificial intelligence (AI) Claude.ai to port the graphical components of the GTK library to modern cross-platform technologies such as Vulkan. As Christian Schaller noted, many developers from the open community are now skeptical about the codes that artificial intelligence generates, so he wanted to show that an AI assistant can be used to solve routine problems.

Christian Schaller, Head of Fedora Desktop Team and Desktop Development Group at Red Hat

I thought it would be a good example of an application if I found something really old and asked Claude.ai to help me adapt it to modern technologies, "Christian Schaller explained in a blog post. - I suddenly remembered the xtraceroute program - an old application originally written in GTK1 and OpenGL showing route tracing on a three-dimensional globe, and found that since I last viewed it, it has been updated to GTK2, but it has not been touched for 20 years. And I thought it was a great test case. So I took the code and uploaded it to Claude, asking it to provide me with a modern version of this app for GTK4 using Vulkan.

View of the xtraceroute application using a legacy library GTK2

GTK (short for GIMP ToolKit) is a cross-platform library of GUI elements, on the basis of which a Linux window manager called Gnome and several others are built: AfterStep, IceWM, Sawfish. Initially, GTK, which was called GTK +, was based on OpenGL technology (an analogue of DirectX for Linux) and provided interaction with the graphics subsystem. However, in January 2024, a version of the GTK4 was released, which has already been rewritten on the Vulkan cross-platform technology.

At the same time, the GTK4 is not backward compatible: that is, you can't just take the application codes for GTK3 and rebuild them for GTK4. In some cases, significant code refinement is required.

Vulkan technology provides the development of applications that can look the same on operating systems such as Windows, Linux, MacOS, Android. Vulkan is developed by the Khronos Group consortium and provides developers with direct access to the hardware capabilities of graphics coprocessors, providing high performance and flexibility in managing graphics rendering. This makes it easier to create graphical applications that can run on multiple operating systems.

View of the GTK4/Vulkan-based xtraceroute application upgraded using the AI assistant

As a result of the experiment, Christian Schaller managed to transfer the application written in GTK2/OpenGL to the new GTK4/Vulkan platform in 4-5 hours and 10 attempts. Moreover, he managed to reuse the codes obtained in the process of work to write a new application for Vulkan, which displays the location of Red Hat offices on the globe.

In general, he praised the modern capabilities of AI assistants for working with source codes in complex projects and suggested using them at least to test the usefulness of additions to Gnome Shell and automatic generation of fixes in various projects with a large number of changes.

Christian Schaller has developed a new application based on xtraceroute AI-assistant ported Vulkan codes

The experiment using generative AI to transfer an application from GTK2/OpenGL to GTK4/Vulkan is really indicative, - Roman Khazeev, director of digital technologies at Rostelecom, told TAdviser. - It is important primarily because it is brought to a complete working result. But from a professional point of view, this is no longer news: such approaches have been actively used in the industry in recent years, including Rostelecom. We have long used AI capabilities to refactoring code, migrate between programming languages, and also to switch to new technological frameworks. For example, we have our own development - Vasilisa, a corporate analogue of Copilot, which is able to work with tasks of this level.

For Russian developers, the task of transferring the source codes of the programs developed by them to other basic technologies is also very urgent. An experiment with porting codes to Vulkan as cross-platform software for the graphical interface, and especially its recognition as leaders in the development of open source projects, one of which is Red Hat, can be useful to Russian developers.

Christian Schaller's experience is more than applicable to Russian developers, especially in the context of working with open technologies and migrating legacy projects to modern platforms, "Daniil Smirnov, senior lecturer at the laboratory for the development and promotion of cybersecurity competencies at Gazinformservice, shared his opinion with TAdviser. - An example presented by Schaller aims to overcome skepticism towards AI assistants for software development.

For Russian developers, the use of AI assistants to switch to modern platforms is very important, since it allows you to transfer graphics applications from outdated Windows to more modern operating systems and other basic components.

Reengineering technologies lie at a different level of abstraction, so they can be used for all libraries and software components, "Vadim Sabashny, CEO of LANIT-Terkom, explained to TAdviser. - Even the lack of source code does not become an insurmountable obstacle. For example, our company is now using AI assistants for database software developers: we are solving the problem of translating stored code between the most popular DBMSs - PostgreSQL, Oracle and MS SQL Server.

However, now Russian programmers have rather routine tasks in transferring codes from Windows to domestic operating systems and basic software. Using AI assistants can greatly simplify these tasks.

When we talk about porting software from Western platforms to domestic ones, the Fedora Desktop Team experiment is not just a beautiful demonstration of the capabilities of language models, but a direct guide to action, "explained Evgeny Nikiforov, head of the development group of the Java development department" Reksofts. " - Today, most customers in Russia work with domestic Linux distributions. However, the vast majority of applications for them are still assembled for imported frameworks and require significant adaptation. Or they generally require migration from Windows in order to fully replace and work on Russian software. Manual migration lasts months, and sometimes years, because it requires a meticulous analysis of outdated code, the collection of expertise and details of the hardware layer.

However, AI assistants also have certain limitations. If they can still perform routine migration operations, it is difficult for them to develop an application architecture or optimize codes, since these tasks often require creative approaches to creating optimal solutions.

Christian Schaller's experiment is interesting, but it is important to understand its context, - said Sergey Lunegov, Product Director of Axiom JDK. - It demonstrates the potential of AI assistants in the tasks of migrating graphics applications, where errors and instability are possible, and this is not critical. But when we talk about a library or system component on which the infrastructure of applications in industrial operation depends, the requirements for quality and predictability are completely different. The applicability of such approaches in the Russian context directly depends on the level of project maturity and acceptable risk.

Thus, Russian experts confirm the overall effectiveness of the experiment of Christian Schaller, who recognized the promise of using AI assistants for routine operations with source codes, such as modernization and refactoring (changing the internal structure of the code without changing its functional features).

AI copes well with "mechanical" migration and autogeneration of templates, but architectural solutions and verification must be left to people, - Dmitry Zboroshenko, ML-architect of Softline Solutions, warned TAdviser readers. - It is possible to conduct such experiments with almost any code that (a) has not been updated for a long time, (b) is well documented and (c) is covered with tests in a sufficient amount - 80-90%, depending on the purpose of the code base.

Russian companies now have a lot of work that can be optimized using large language models. In particular, it is necessary to modernize all domestic solutions that were based on outdated technologies.

Looking at our landscape, the first candidates for modernization are open source projects where there are outdated graphics components, "Konstantin Kudryashov, director of AI infrastructure at Ainergy, explained to TAdviser. - For example, scientific imaging tools like Gnuplot or MathGL, outdated desktop environments, and in general any niche or industrial software stuck on GTK1/2, Qt3 and the like.

Moreover, Konstantin Galuzin, developer of databases and AI tools of Simpl Group, notes that there are still projects that are written on outdated technologies. Among the legacy platforms, in particular, the expert attributed jQuery, AngularJS, Backbone.js, PHP without modern frameworks. He told TAdviser that AI assistants can be used to solve the following problems:

• Automatically convert the code to React, Vue, Svelte or Solid.js.
• Generate TypeScript types for untyped JavaScript.
• Refactor legacy code by replacing legacy APIs (such as XMLHttpRequest with Fetch or Axios) with more modern ones.

The use of AI assistants can significantly accelerate the digital transformation of the Russian IT market, especially in terms of the modernization of outdated systems, says Konstantin Galuzin. - Today, a significant share of organizations, including in the public sector and banking, still work on databases like Delphi, 1C 7.7, Java 8, PHP 5. Migrating to modern technologies (Python, Go, Kotlin, React) is an expensive and time-consuming task. Automatic refactoring will reduce the cost of supporting older systems.

The use of artificial intelligence to transfer solutions developed for foreign platforms to the Russian technology stack will accelerate the solution of the problem of digital sovereignty and the development of a completely Russian software development ecosystem.

It is important to be aware of the difficulties associated with the optimality of the code (the article gives a negative example) and ensuring data security, - said Kirill Semion, General Director of the National Center for Competencies for Holding Management Information Systems (NCS ISU). - Potentially, this is a good opportunity to get a tool for import substitution in the IT sphere. For example, to transfer applications from Windows to Linux or from a foreign mobile platform to a domestic one. But, like any tool, it must be used rationally and "with open eyes."

Sergey Rybakov, head of the advanced projects department, MyOffice"" believes that the experience of Christian Schaller is applicable in Russia with two reservations. The Russian IT industry is now focusing on security. And the use of AI without manual code validation, built-up testing and acceptance processes threatens the emergence of vulnerabilities with all the ensuing consequences in the form of financial and reputational damage. Also, the most current proprietary models and services are not available to Russian developers. This pushes them to either create their own solutions or use an open model, the experience of which was demonstrated by Christian Schaller.

The introduction of AI assistants will radically reduce the time of modernization from months to hours, ensuring exponential growth in development efficiency, "Stanislav Yezhov, director of AI at Astra Group, told TAdviser. - Integration of safe and trusted language models into domestic development platforms will catalyze technological independence and speed up import substitution in conditions of limited access to international projects with open codes.

Russian companies are faced with the task of quickly replenishing the software solutions of software developers who left our country, and in conditions of a shortage of personnel. Using AI assistants allows you to solve this problem

The use of AI assistants will reduce dependence on foreign software, "explained Daria Usacheva, data management engineer at SimbirSoft. - This gives old projects a chance to integrate into current technology stacks. It is important to perceive AI assistants as a tool in the hands of the developer, and not its replacement: responsibility for the quality and support of the code remains with the person. When working with complex projects, AI has not yet coped with full structural analytics, but in tandem with the developer it significantly increases the efficiency of the process.

However, experts note that Western AI assistants are now mainly used, which are still in the public domain. Using them in the process of introducing new technologies can tie modernized projects to foreign AI infrastructure. Therefore, it is better for Russian developers to rely on Russian AI assistants in their work, which are already being developed and offered.

Any AI assistant is based on a language model and a set of software tools, "Avenir Voronov, head of AI at Dar, explained to TAdviser. - Today, most Russian companies use Western solutions, and local developers creating LLM are focused mainly on large corporate projects. Individual developers have almost no chance of creating their own LLMs or assistants. Therefore, I am skeptical about the prospects for a noticeable development of the domestic software market. There are likely to be small niche solutions from strong teams, but globally Russian developers will have to compete with China and the United States, where teams are given resources and time to develop AI assistants, and it will be extremely difficult for Russian developers to update old technologies.

AI blind spots: Why software testers will never be out of work in Russia - TA opinions

As explained by Kirill Smelovets, chief architect of artificial intelligence and machine learning services Cloud X, there are already AI agents on the market that can independently raise the test environment, interact with the user interface, collect logs and generate reports. However, real control of the desktop and application is much more difficult than mechanical reproduction of steps: models need to understand the logic of business processes, typical scenarios for using the product and the variety of Russian environments, he continues. In addition, users often behave illogically: enter typos, change the sequence of actions, open unexpected menu branches.

Inaccessible AI areas: Why Russian software testers will always be in demand

Smelovets is confident that in the future the role of the classic tester will be transformed: the specialist will tune agents, train them on fresh datacets, validate conclusions and make sure that the model does not degrade. The final decision to "release or not release" the product will still remain with the person.

Nikolai Kirichenko, head of testing at MD Audit (Softline Group of Companies), doubts that AI will be able to completely replace manual testing - especially in Russian realities, where they often work with non-standardized interfaces, rapidly changing requirements and industry specifics. A human approach is needed where contextual assessment, common sense, UX and understanding of business logic are important. AI can speed up the routine, but critical scenarios and final verification will remain with the person, the interlocutor explained to TAdviser.

The founder of the Chad AI neural network, Arthur Koltsov, claims that artificial intelligence is not capable of performing long chains of actions, so there can be no talk of a complete replacement of manual testing. For example, if a task consists of 10 steps, and AI performs each of them with an accuracy of 90%, then the probability of passing the chain is about 34%. For AI to replace testers, the accuracy of each step must be at least 99.9%, while the new approach must be more economically profitable, Koltsov notes. Neural networks speed up the execution of routine tasks (helps with documentation, analyzes errors, even prompts possible bugs), but they are not yet smart enough - a person must configure them, he stressed.

According to Tatyana Konyukhova, head of the sales group of industry solutions "First Bit," the main problems of AI in testing are technical limitations:

• false positives: AI may misinterpret test results
• dependence on data quality: efficiency directly depends on the completeness of training data
• complexity of configuration: integration of AI tools requires special competencies
• limited interpretability of solutions: the complexity of explaining the logic of modern machine learning algorithms for end users
• risk of bias: systems inherit errors from historical data.

The author of the Solar appScreener product Daniil Chernov draws attention to the fact that software quality verification should also include information security testing of software products at all stages - from development to production. In this direction, AI is unlikely to be able to replace platforms for comprehensive security in the development of applications that include technologies for static and dynamic analysis, analysis of the composition and supply chain of software (SAST, DAST, OSA), the expert believes.

Andrei Myskin, head of testing practice at Reksoft, agrees that AI will not completely displace manual testing either in Russia or in other countries. According to him, testing is largely a creative process that requires non-standard decisions and views. In addition to solving complex engineering problems, a testing specialist needs to present himself as users with completely different levels of training, habits, and approaches to work. He needs to think in a special way, presenting potential risks and limitations, thinking through the most unusual and non-standard situations. By 2025, AI models will not be able to work out all such situations, Myskin added.

Igor Volynets: "The introduction of artificial intelligence is not possible in every project"

Despite the benefits that AI offers, human participation is still critical. Neural networks, like any tool, need control and adjustment by the developer, since they can offer code that does not correspond to the specifics of the project or even contains errors, explains Igor Volynets, a member of the hackathon jury, an expert in the field of software quality assurance. How the capabilities of AI have changed, whether the technology is applicable on all projects, what restrictions can be encountered during its implementation and how to overcome the team's resistance - Igor Volynets said in an interview. Read more here.

2024: Nvidia CEO: Don't teach kids programming - AI will replace developers

At the end of February 2024, Nvidia CEO Jensen Huang shared his vision of the prospects for the introduction of artificial intelligence in the field of software and services. In his opinion, there is no longer a need to study programming, since in the future AI will replace developers when writing code.

Automated software creation is considered one of the most promising areas of application of generative AI (Genia). Such tools will not only improve performance, but will also help improve the quality of program code and reduce the number of possible errors.

Jensen Huang shares his vision of the prospects for the introduction of artificial intelligence

According to Huang, as Genia systems improve, the need for programming specialists will rapidly decrease. And therefore, the head of Nvidia does not consider it necessary to teach children the skills of writing code - in the future these tasks will be able to fully take on AI. It will be enough for a person to formulate a task in a natural language.

Our task is to create a computing environment in which no one has to program. Everyone in the world will become programmers. It's an artificial intelligence miracle, "Huang says.

Against the backdrop of the rapid development of Genia, according to the CEO of Nvidia, people should focus on acquiring knowledge and skills in other important areas, such as biotechnology, education, production and agriculture. Huang's statements contradict the well-established view that "if a young man wants to succeed in the IT industry, he must learn to program." However, as of the end of February 2024, there are various AI-based programmer assistants that generate code in response to text requests and provide support in solving certain problems.^[3]

2017: Andrey Karpati: machine learning and neural networks will help develop Software 2.0

Andrei Karpathy, director of Artificial Intelligence at Tesla, [1] pointed out in a 2017 article that the time is coming for new software - Software 2.0. It will be an order of magnitude more complicated than existing developments and will be helped to develop machine learning and neural networks.

Machine learning models find important functions and patterns in data, he believes, and areas that benefit most from 2.0 software include computer vision, speech recognition, machine translation, games, robotics and databases.

A recent analysis by Forrester Research indicates that the custom development market ON has grown to 1.25 trillion, and the dollars USA University of Cambridge adds that 50% of the work is spent on code compilation and bug correction. With the help of artificial intelligence, this part of the work can be significantly reduced.

Artificial Intelligence in Mobile App Development

In the field of mobile development, AI provides developers with new opportunities. First of all, this is due to the fact that the use of AI helps to attract more users to use the application.

AI tools automatically execute certain algorithms to ensure that more users start using this application. For example, AI can track the patterns and preferences of its users, predict their future decisions and choices, and work accordingly. This gives developers the ability to make changes to new versions of applications very quickly.

The growing popularity of smart devices has spurred the growing use of artificial intelligence in the development of operating systems and user interfaces. And the need for the availability of mobile applications that create a personalized user experience is growing day by day. Artificial intelligence acts as a personal virtual assistant, the application can record a large number of user actions.

Thus, artificial intelligence becomes a valuable source of feedback for developers who can implement the identified needs of users in a short time.

Vibe coding

Main article: Vibe coding

AI Software Development Tools

A growing number of AI-based tools support software development processes. Some of these solutions are available for free. Leading technology vendors use such tools and offer them to their customers in the form of additional products (plug-in).

Facebook uses a recommendation service to fix bugs and improve code. Recent IBM Mono2Micro and Application Modernization Accelerator (AMA) projects provide application architects with the tools to update legacy applications and reuse them. And Microsoft in 2021 announced that it is integrating artificial intelligence technologies with its Power Fx programming language, which is used in the development of applications on the platform. Power Platform This will allow the company's customers to create programs almost without the need to write code.

B Russia actively uses AI to create software products. In Sber July 2021, Sber AI registered Rospatent in a program that allows artificial intelligence to recognize and analyze objects in, virtual reality follows from the department's materials.

According to a Forrester survey, 37% of respondents acknowledged that they use AI for a more efficient testing and development process.

File:Ии для по.png

^{Tools for developing and testing software based on artificial intelligence. Source: Forrester Research, 2021}

However, there is another side to AI products for development. ON Teams that use tools to improve code may experience a drop in productivity first, as AI products require skills and the ability to handle them. Only after deep immersion, they are able to issue accurate recommendations for optimizing software development.

2025

Microsoft will replace hundreds of developers on AI who helped create it

In mid-July 2025, it became known that Microsoft was laying off about 200 employees of the King gaming studio, which is behind the popular mobile game Candy Crush Saga. These developers will be replaced by artificial intelligence, in the creation of which they took part. Read more here.

Indian IT specialist for 4 years managed to work in 80 companies - juns and neural networks worked for him

Soham Parekh, a programmer from India, managed to find a job in eighty different technology companies over a four-year period, while not actually performing any work duties and delegating tasks to artificial intelligence. The specialist received wages without doing real work, and after being fired for inefficiency, he immediately settled in new organizations according to a similar scheme. Information about fraudulent activities became known on July 3, 2025. Read more here

MTS has launched a free service for writing computer code. It uses a large language model with 1.5 billion parameters

In June 2025, MTS launched the free AI service Kodify Nano for generating program code, built on a language model with a volume of 1.5 billion parameters. The product is aimed at developers who need a local tool without access to cloud power. According to the developers, their solution allows you to test basic scenarios directly on user devices and is included in the strategy for the development of domestic digital products. Read more here.

Amazon developers complain that work after AI implementation becomes monotonous "like warehouse employees"

Amazon is actively introducing artificial intelligence technologies, including in the field of writing program code. However, the company's developers complain that this approach makes their activities monotonous "like warehouse employees." This is stated in an article by The New York Times, published at the end of May 2025. Read more here

Google has released a neural network that creates web applications

On May 6, 2025, Google announced the Gemini 2.5 Pro (I/O Edition) artificial intelligence model, which has improved software code generation capabilities, especially in terms of creating interactive web applications. Access to the neural network is provided through the Gemini API, as well as through the Vertex AI and AI Studio platforms. Read more here.

AI service released to generate full-fledged applications without programming skills

At the end of April 2025, the startup Adaptive Computer launched a new service based on artificial intelligence, allowing you to generate full-fledged applications without programming skills. The company raised $7 million for development as part of the "sowing" round of financing. Read more here.

Started the site generator by text description - it copies the existing interface by link

At the end of March 2025, the Same.new Internet service was launched based on artificial intelligence, which allows you to generate websites by text description. In this case, the interface of an existing resource can be copied using a user-specified link. Read more here.

Head of AI company Anthropic: In a year, 100% of the program code will be written by artificial intelligence

Artificial intelligence will be able to fully take over the work of writing the main program code in 2026. This statement was made in mid-March 2025 by Dario Amodei, CEO of Anthropic, which specializes in generative AI technologies.

Programming is one of the fastest-growing fields for AI, he said. Amodei believes that within three to six months (by June-September 2025), artificial intelligence will be able to generate up to 90% of the program code within the framework of the task, and six months later this figure will reach 100%.

Dario Amodei

However, AI will not be able to completely replace human specialists. Programmers will still play a key role in developing the concept of the application, determining the necessary functionality and making final decisions. Amodei, who previously served as vice president of research at OpenAI, emphasizes that the participation of a human programmer in the software creation process will remain necessary, but AI will gradually take on many of the tasks that people used to perform. It is assumed that this approach will expand the possibilities in terms of software development, as well as solve many problems associated with the human factor. In addition, in the future, the programming process may be translated into a natural language, which will change the methods of creating software and the requirements for specialists.

In general, Amodei calls for a review of the definitions of "utility" and "futility" of artificial intelligence. He believes in the possibility of building a world in which AI will complement a person's capabilities by taking on time-consuming and routine operations. At the same time, people will be able to focus on more creative tasks.^[4]

GPT-4o and other neural networks do not cope with most programming tasks - OpenAI study

Large language models (LLMs) greatly simplify and speed up the writing of program code, but they are not able to independently cope with most programming tasks. This is stated in the OpenAI study, the results of which were published in mid-February 2025.

OpenAI specialists have developed a benchmark called SWE-Lancer to assess the capabilities of LLM to solve real problems in the field of software development. The test showed that AI models are able to correct errors, but they are not able to understand why these bugs occur. As a result, neural networks can make their own mistakes during operation.

OpenAI study finds that GPT-4o and other neural networks are unable to effectively solve most programming-related problems

The study analyzed the AI models OpenAI GPT-4o and OpenAI o1, as well as Anthropic Claude-3.5 Sonnet. These neural networks were tasked with solving 1,488 tasks for developing software from the Upwork freelance platform for a total of $1 million. All tasks were divided into two categories: individual tasks for performers (elimination of errors or implementation of functions) and management tasks (where the model plays the role of a manager who selects the best offers for solving problems).

Benchmark showed that none of the models was able to fully solve the problems and earn $1 million. The best result was shown by Claude 3.5 Sonnet, which could potentially get about $208 thousand, coping with 26.2% of the tasks of the performers. The researchers emphasize that most neural network solutions are incorrect. The test shows that basic AI models as of February 2025 cannot completely replace human programmers.

The results demonstrate that the real-world freelance tasks in our benchmark remain challenging for advanced language models, the OpenAI study said.[5]

2024

Russia took 5th place in the use of AI in programming

Software developers from the United States most often in the world use artificial intelligence-based tools to generate code. Russia in this rating was in fifth position. Such data are contained in the study, the results of which were published in mid-June 2025.

According to The Register, the authors of the work are Simone Danielti, Johannes Wachs, Xiangnan Feng and Frank Neffke. They developed a special machine learning model to analyze commits on GitHub. After that, about 80 million code samples added from 2018 to 2024 were analyzed.

Russia took 5th place in the use of AI in programming

It turned out that 30.1% of all functions written in the Python programming language and added to GitHub by developers from the United States in 2024 were generated through AI. In second place in the list is Germany with an indicator of 24.3%, and France closes the top three with a result of 23.2%. This is followed by programmers from India with 21.6%, while the indicator of developers from Russia was 15.4%. The result of China is 11.7%.

According to the authors of the study, after the share of the code generated by AI in the total mass exceeds 30%, the quarterly volume of commits on GitHub will increase by 2.4%. In other words, AI allows you to speed up the work of programmers. Moreover, the potential economic benefit of increasing efficiency indicators using AI is estimated at $64 billion to $96 billion per year.

At the same time, the researchers recognize that their counting methodology has certain limitations. This is, in particular, the assessment of commits exclusively in Python without taking into account other programming languages. In addition, statistics on other code repositories are not considered.^[6]

A special laptop has been released for a programmer who writes and checks the code

At the end of November 2024, Mechanical Revolution announced the Code AI laptop, designed specifically for programmers. The laptop includes an artificial intelligence assistant designed to automatically write code. Read more here

20% of Russian IT companies use generative AI in software development

According to the Russoft association, about 20% of domestic software developers in 2024 use generative artificial intelligence when creating software. This indicator demonstrates steady growth from 7.9% in 2022 and 15% in 2023, which became known on November 20, 2024.

According to Kommersant, the growth in the use of AI occurs against the background of a shortage of personnel in the industry, which can reach 1 million people in 2024. The average wage in the sector increased by 12.5%, reaching ₽123 thousand.

20% of Russian IT companies use generative AI in software development

President of Russoft Valentin Makarov notes that artificial intelligence has reached a level that allows you to technically replace entry-level developers. This becomes economically feasible for small companies due to the low qualifications of novice specialists and the general growth of salaries.

Technical Director of RTK IT Kirill Pikhtovnikov emphasizes that development using AI and automation of processes has become a priority area for investments by Russian IT companies. Rostelecom is introducing similar technologies to optimize routine operations.

At the same time, in Russia, the demand for platforms with ready-made tools for creating IT projects increased by 40%. According to the agencies Smart Ranking and Nocodecircle, the volume of this market by the end of 2024 could reach ₽3,5 billion.

Softlogic CEO Denis Loginov points to potential risks when using AI in development. If the tasks are complicated, problems may arise with the interpretation of their neural networks, which will affect the integration and security of the final code.

The Ministry of Digital Development of Russia considers the trend towards the use of AI as a natural stage in optimizing and introducing new technologies, without directly linking it with the personnel deficit of the industry.^[7]

More than a quarter of Google code writes artificial intelligence

Google is actively developing artificial intelligence technologies, forming a comprehensive infrastructure that covers a variety of areas - from data centers and hardware solutions to search optimization and automatic code generation. Google CEO Sundar Pichai spoke about the achievements in the relevant area on October 29, 2024. Read more here.

MTS launched a neural network to generate the Kodify program code

On October 22, 2024, it became known that MTS AI presented a publicly available version of the Kodify neural network service for automatic program code generation. The service allows developers to create and test code fragments using artificial intelligence technologies for auto completion. Go in detail.

Yandex launched a service for generating software code Yandex Code Assistant

On September 12, 2024, Yandex announced the launch of a new Yandex Code Assistant service designed to automatically generate program code based on artificial intelligence. The service, designed to simplify the work of programmers, is already available in test mode on the Yandex Cloud platform. Yandex Code Assistant is expected to be part of a broader platform designed to build, deploy and accompany digital products. Read more here

Google launched a free AI tool for writing code

On April 9, 2024, Google introduced the free artificial intelligence-based Gemini Code Assist tool, designed to help write code. This system has received Automatic Row Completion. Read more here.

Launched an open neural network that writes program code on voice commands

On March 12, 2024, startup Cognition Labs announced Devin, a specialized neural network capable of independently performing complex software development tasks. It is claimed to be the industry's first software engineer with artificial intelligence tools. Read more here.

2023

The volume of the global market for generative AI for software development for the year grew to $2.1 billion

In 2023, costs in the global generative artificial intelligence (GENI) market for software development reached $2.1 billion. This is almost a third more compared to 2022, when expenses were estimated at $1.6 billion. Such data are given in the Market Research Future study, the results of which were published in early December 2024. Read more here.

10% of software developers in the world use AI programming assistants

As of 2023, approximately 10% of software developers globally use various artificial intelligence programming assistants. And approximately 63% of organizations in the world test or implement AI code creation tools. This is stated in a study by Gartner, the results of which were published on April 11, 2024. Read more here.

Notes