Last June, Microsoft's GitHub and openai launched copilot, a service that provides advice on the whole line of code in Microsoft's development environments such as visual studio The Codex model can provide billions of additional code lines as the extension of the existing code. According to the Codex model, the Codex can be used as an extension of the existing code line. Copilot can also use its knowledge base and current environment to emerge a method or solution for the description that developers want to complete.
Microsoft announced at the build 2022 conference that although copilot used to be a technical preview, this auxiliary tool will be fully available sometime this summer. Copilot will also be available free of charge to students and "proven" open source contributors. On the latter, GitHub said it would share more information later.
The experience of using copilot will not change much. As before, developers can cycle through suggestions from python, JavaScript, typescript, ruby, go, and dozens of other programming languages, accepting, rejecting, or manually editing them. Copilot will adapt to the editing of developers, match specific coding styles, automatically fill in templates or duplicate code patterns, and recommend unit tests that match the implementation code.
In addition to visual studio In addition to code, the copilot extension will be available for noevim and JetBrains, or in the cloud of GitHub codespaces.
A new feature consistent with the comprehensive release of copilot is copilot explain, which can translate code into natural language description. This is described as a research project designed to help novice developers or those unfamiliar with the code base.
"Earlier this year, we launched copilot labs, a separate extension of copilot, as a testing ground for machine learning experimental applications to improve the developer experience," Ryan J. Salva, vice president of products at GitHub, said in an email interview. "As part of the copilot lab, we have introduced 'explain this code' and 'translate this code'. This work is in line with the kind of experimental ability we are testing, which allows you to glimpse various possibilities and let us explore use cases. Perhaps by 'explain this code', a developer is weighing an unfamiliar code base and wants to quickly understand what is happening. This function allows you to highlight a generation Code block and ask copilot to explain it in simple language. Similarly, the purpose of the pilot laboratory is experimental, so problems may arise. Laboratory experiments may or may not develop into the permanent function of copilot. "
However, due to the complex nature of artificial intelligence model, copilot is still an imperfect system. GitHub warns that it may generate unsafe programming patterns, errors and references to outdated APIs, or reflect the teaching of imperfect code in its training data. The code suggested by copilot may not always compile, run, or even make no sense at all, because the code from AI does not actually test these suggestions. In addition, in rare cases, copilot's recommendations may include personal data, such as names and e-mails, from its training set word for word - or worse, "biased, discriminatory, abusive or offensive" text.
GitHub says it has implemented filters that can block it when displayed in standard formats and offensive words, and it is building a filter to help detect and suppress duplicate code from the public repository. The code should be carefully reviewed and recommended by copilot, but it should be better as the code is reviewed and recommended by copilot on other websites
Although copilot may have improved since the launch of the technical preview last year, the extent of the improvement is unclear. The capabilities of the underlying model Codex -- a descendant of openai's gpt-3 -- have been matched (or even surpassed) by systems such as deepmind's alphacode and open source polycoder.
"We've seen copilot make progress in generating better code... We're taking advantage of our [other ] Experience in tools to improve the quality of copilot recommendations - for example, by giving additional weight to the training data scanned by codeql, or analyzing recommendations at runtime, we are committed to helping developers improve productivity, code quality and security. In the long run, we believe that the code written by copilot will be safer than the average programmer, "Salva said.
The lack of transparency does not seem to curb the enthusiasm for copilot. Microsoft said today that about 35% of the code generated by developers in languages such as Java and python in the technology preview was prompted by copilot. The company said tens of thousands of people regularly used the tool throughout the preview process.