开发者用生成式人工智能魔法将数年工作缩减至数天_AI阅读总结 — 包阅AI

Every day now, we’re hearing new stories of developers using generative AI (GenAI) tools to slash the time and expense it takes to write new or refactor old code.

Just the other day, Kautuk Pandey, a director of data platforms at Visa, shared on LinkedIn that he used ChatGPT and GitHub Copilot to reverse engineer and document an 8-year-old Java codebase in one day.

In his very popular LinkedIn post, Pandey explained that “This was a code base written by people who have all left the company, the oldest person in my team is about 2 years into the system. So, I had absolutely zero context of what this code does and why/how it does that. Bring ChatGPT and CoPilot into the mix and I was kind of able to reverse ‘pair-program’ the whole code base into simple design docs in one day.”

GenAI Pair Programmer

Pandey acknowledged that he effectively used the GenAI tools as a pair programmer to save time and avoid grunt work.

“I don’t think I would ever have been this productive in understanding and breaking down legacy codebases,” Pandey wrote. “With the GenAI solutions available today, tedious and uninteresting tasks become a lot easier and somewhat more bearable (or fun).”

The post generated significant discussion — more than 120 comments and over 2,500 reactions, with many IT professionals sharing similar experiences of increased productivity using AI tools.

For instance, Ajai Govind Govindan, a chief data and analytics officer, said he did something similar recently with some Salesforce code. “But what you’ve mentioned in this comment is also very important in the process — a basic level of understanding of how code is written in any language (like class-object, folder structures etc.), but more important than that, an inquisitive and curious mind,” he wrote in response to Pandey. “Without that for someone who thinks ‘this is not my job,’ any number of GenAI solutions or advancements are not going to be helpful.”

Indeed, GenAI code generation can make for an excellent pair programmer, not just in developing new solutions but also in maintaining well-established code bases, said Brad Shimmin, an analyst at Omdia.

Eliminating Technical Debt

One of the biggest challenges faced by enterprise IT is dealing with the technical debt of past investments. Between staff turnover and shifting IT priorities — such as shifting from JavaScript to React for all of our UX work — knowledge gets lost very easily, Shimmin said. And this makes it very difficult for companies to both maintain and evolve their codebase over time.

“As luck would have it, an area where GenAI truly excels is in consuming and understanding tons of sequential information like a codebase,” Shimmin told The New Stack. “With this foundation, an LLM can readily and rapidly help to build an understanding of that codebase, even going so far as to create/improve documentation for functionality that may make no sense to a first-year developer, for example.”

That can help companies to better preserve their institutional knowledge and lessen their overall technical debt — a happy situation that just might make enough room for the company to think about potentially refactoring or rebasing legacy code.

“The trick of course is picking an LLM that a) is good at code generation, documentation, etc.; b) been trained on the entire code base either through fine-tuning or via in-prompt learning via RAG [Retrieval Augmented Generation] and a really large context window; and c) does all of this without exposing the company to any kind of IP, security, or privacy risk,” Shimmin said.

“Legacy code is nice, but everything we have now and develop should continuously be going through the AI analysis,” said Johnny Crupi, CTO for GenAI at Ryght.

Success Stories

Tarun Gandotra, a data science manager at Carelon, said he was able to complete four days of work in two hours using ChatGPT. That included parsing the content, creating objects from content and then creating a base machine learning model out of that, Gandotra wrote in response to Pandey.

Shirish Bhatt, co-founder of Coditas, a GenAI consulting firm, wrote in response to Pandey that his firm has been helping a lot of enterprise customers to convert legacy code to the modern tech stack “with the help of our GenAI-powered platform therix.ai. We also generate business requirements documents along with code. This has resulted in savings of approx 40% of time and money.”

And Abhinav Girdhar, founder of Appy Pie, wrote: “Leveraging GenAI tools like ChatGPT and CoPilot to reverse engineer and document a legacy codebase so quickly is a game-changer. It’s fascinating to see how these technologies can transform tedious tasks and make complex processes more manageable. Embracing GenAI is definitely the way forward in optimizing productivity and efficiency in our work.”

GitHub AI Survey

Meanwhile, GitHub this week released the results of its 2024 survey on the impact of AI coding tools in software development. The survey included 2,000 respondents from the US, Brazil, Germany, and India, including software engineers, developers, programmers, data scientists, and software designers.

Overall, the survey highlighted the growing importance of AI in software development and the need for organizations to strategically integrate these tools to maximize their benefits.

In terms of AI tool usage, more than 97% of respondents said they have used AI coding tools at some point. And a majority of respondents say their companies encourage or allow AI tool use. The U.S. leads with 88% of respondents indicating at least some company support for AI use, while Germany is lowest at 59%, wrote Kyle Daigle, chief operating officer at GitHub, in a blog post.

Moreover, the survey showed that software development teams are gaining more benefits with AI coding tools than previously reported. Some of these include building more secure software, improved code quality, better test case generation, and faster programming language adoption. This ultimately translated to time savings that they could use for more strategic tasks, Daigle wrote.

Most respondents in the U.S. (90%) and India (81%), along with more than half in Brazil (61%) and Germany (60%), reported a perceived increase in code quality when using AI coding tools, Daigle said in the post.

Stack Overflow Research

Research from Stack Overflow’s recently released 2024 developer survey indicated that in the next year, most developers say that AI tools will be more integrated mostly in the ways they document code (81%), test code (80%), and write code (76%).

Stack Overflow showed that 76% of all respondents using or planning to use AI tools in their development process this year is an increase from last year (70%). Many more developers are currently using AI tools this year, too (62% vs. 44% last year).

ChatGPT is used by twice as many developers as its next closest alternative, GitHub Copilot — 82.1% versus 41.2%, followed by Google Gemini and Microsoft Bing AI.

Making Sense of Legacy Code With GenAI

Making sense of legacy code is in fact one of the primary use cases of GenAI.

BMC Software’s BMC AMI DevX Code Insights tool, for example, can reveal the structure and dependencies within legacy COBOL code.

“Now that BMC has added GenAI, the tool can also make sense of the business logic behind the code and explain it in natural language,” Jason Bloomberg, an analyst at Intellyx, told The New Stack. “This capability both helps developers as they modernize legacy applications and also writes understandable documentation, both for legacy code as well as new code as the developers write it. What BMC can do for COBOL, other vendors can do with other languages.”

AWS Moves From Java 8 to Java 17

Meanwhile, in a major proof point of its own, on August 1st, Amazon CEO Andy Jassy AWS presented data showing that using its own Amazon Q Developeragent for code transformation capability, the company migrated over 30,000 Java JDK applications in a few months — saving over 4,500 years of development work for over a thousand developers (when compared to manual upgrades) and performance improvements worth $260 million dollars in annual cost savings.

Amazon Q Developeris a GenAI-powered assistant for software development. Doug Seven, AWS’ director of AI Developer Experiences, told The New Stack that the tool generates highly accurate code, can have a conversation about that code, tests, debugs, troubleshoots, performs security scans and fixes, filters out code suggestions that may be considered biased or unfair, and has multistep planning and reasoning capabilities that can transform (e.g., perform java version upgrades) and implement new code generated from developer requests.

A team of five Amazon developers usedAmazon Q Code Transformation to upgrade 1,000 production applications from Java 8 to Java 17 in just two days.The average time per application was less than 10 minutes. For comparison: It used to take two days just to upgrade one application. AWS ran a productivity challenge, and developers who usedAmazon Q Developer were 27% more likely to complete tasks successfully.

To determine the true business impact of Q Developer-assisted app upgrades, AWS estimated the time saved by looking at the number of Java dependencies they migrated, Seven said. Typically, it can take a day or more of a developer’s time to migrate just one dependency, and many applications have dozens of dependencies that need migrating. With the agent for code transformation, many of these dependencies can be migrated in minutes, resulting in significant time savings.

To estimate cost savings, AWS looked at the number of hosts they were able to remove from the applications due to the performance improvements achieved by upgrading to Java 17. Both of these estimates are conservative, and the actual cost and time saved is likely much greater.

Seven said in addition to moving code from one version of Java to another, the company has received customer requests to move code from .NET to .NET Core.

AI Brings a Fundamental Change to Software Development

“This is about making every developer ‘a team of developers’ through AI assistance,” Seven told The New Stack.

However, he emphasized that AI tools like AWS Q are not about replacing developers but transforming their roles and increasing productivity and business value creation speed.

“I don’t think anybody should be scared of AI,” Seven said. “I think what’s really fascinating about this is, as a lifelong developer, there are so many things that we spend our days doing that are the less enjoyable parts of what we do, and the idea that I now have this set of AI agents that I can leverage to do some of this more mundane work for me, leaves me with the creative problem solving, which is the most exciting part thing for me.”

With its agents, AWS is opening the door to a future where AI becomes a very serious collaborator in the software development process.

“That is going to fundamentally change how software developers work and how businesses are able to achieve value through software in the future,” Seven said.

To be sure, as Pandey said in his LinkedIn post: “Whether we like it or not, GenAI is going to stay and fundamentally alter the way we work. The sooner we make peace with that, the earlier we will be able to use this superpower in our daily work.”