Sberbank and Yandex create analogues of GitHub Copilot
[ad_1]
Sberbank, which introduced the GigaChat generative neural network this year, has developed a development tool and code completion GigaCode. The possibility of creating a service for developers based on artificial intelligence (AI) is also being considered by Yandex. Market participants note that such tools are still able to add only small code fragments. The result of the generation may coincide with someone else’s code, but lawyers do not see any risks in this either for the owner of the service or for its users.
Sberbank has registered in the Rospatent registry the GigaCode / JARVIS program, designed, according to the description, for auto-completion and code generation in different programming languages and requests using AI. The registration was completed on August 8, and on August 3 the bank registered the gigacode.ru domain, which hosted the JARVIS login page at the time of publication. Sberbank confirmed the creation of the program to Kommersant. Now the service is used by several thousand Sber specialists, in the coming months GigaCode will be available to external developers.
In the spring, Sberbank introduced the GigaChat neural network, which is capable of generating code fragments in the course of a dialogue with the user. GigaCode, as Sberbank said, “is developing in parallel with GigaChat and has a number of intersections.”
The new service can integrate with popular development environments: IntelliJ IDEA and PyCharm from JetBrains, Visual Studio Code from Microsoft, etc.
The generative neural network YandexGPT was also presented by Yandex this year. The company clarified to Kommersant that they are considering creating products for developers, but now they are focused on integrating YandexGPT into services for mass users.
The first mass AI code generation service was GitHub Copilot, introduced in 2021 and based on OpenAI products (developed by ChatGPT). Microsoft, which owns GitHub, provides access to the service by subscription: $10 per month for individuals and $19 per month per user for organizations. Microsoft and OpenAI have blocked access to GitHub Copilot and their other neural network services in Russia. Sberbank did not specify the conditions for access to GigaCode.
The speed with which the language model is able to generate code is essential for developers, “now they write only a few paragraphs of text with sufficient efficiency,” says Alexey Sergeev, head of the Machine Learning and Artificial Intelligence practice at Axenix (ex-Accenture). This, he says, makes the tools suitable for writing individual sections of code, methods and functions.
Softline Digital ML/AI architect Dmitry Zboroshenko sees interest in code generation tools among clients, but so far “there is no way to judge how widely they are already being used.” Such products can speed up development, but their mass introduction will increase the threshold for entering the profession, adds Nikolai Sokornov, director of the software development department at Reksoft: “Newbies are most often assigned routine tasks – for example, write several times a piece of code according to a model. Tools like Copilot eliminate the need for such actions.”
Brad Smith, Microsoft President, on AI August 29 (quote from TASS):
“The biggest mistake people can make is to think that this tool will allow people to stop thinking altogether.”
In 2023, Microsoft announced a new version of GitHub Copilot that can generate descriptions for proposed changes to a project, and began using the Copilot brand in relation to the AI functions of its other products – in particular, the Microsoft 365 office suite. At the same time, a lawsuit against the creators of the service from a group of people on whose work the neural network underlying GitHub Copilot was trained.
There is a possibility that the result of the generation will repeat the once written code, but whether this will be a violation is a question, intellectual property lawyer Anastasia Skovpen believes: “It all depends on whether the copyright holder can prove that his work was in the training data and that it is illegal reproduced or reworked.” So far, there have been no such cases, adds Vladimir Ozherelyev, head of the DRC intellectual property practice: “The legal risks of the owner of code generation tools are minimal, and large companies will probably also actively use them.”
[ad_2]
Source link
