Chatting Code Evolution: How AI is Automating Library Upgrades and Saving Developers’ Sanity

In the lightning-fast world of software development, API changes can feel like that one friend who always insists on changing plans last minute. You’ve just gotten cozy with the current version when a massive update comes along, bringing new features and breaking your perfectly polished code. Enter Large Language Models (LLMs) like ChatGPT, which are not just here for witty banter and poem writing but also stepping into the realm of automating code migrations. Let’s dive into how a group of clever researchers are harnessing AI’s prowess to carry out library migrations, starting with the popular Python ORM, SQLAlchemy.

The Genesis of the Code Migration Problem

Imagine being a developer managing an application that prominently uses an API, a software middleman that lets different programs chat. When the API gets a new version — usually stuffed with tempting new features — you’re tasked with updating your app. Often, these updates aren’t straightforward because they might introduce breaking changes (like when you suddenly find your morning coffee replaced with green tea).

The traditional approach to this daunting task? Roll up your sleeves and manually tweak the code. This not only costs time but also introduces potential for errors. Despite the progress AI has made in transforming other areas of coding—like generating code or catching bugs—using AI for code migrations felt like a glimmering opportunity not yet explored. This is where Aylton Almeida, Laerte Xavier, and Marco Tulio Valente swoop in with a novel idea: letting LLMs handle these tedious API migrations.

Research Breakdown: Let the AI Do the Heavy Lifting

The researchers chose ChatGPT (specifically, the GPT-4 model) to embark on this library migration journey, focusing on migrating from version 1 to 2 of SQLAlchemy in a Python app. SQLAlchemy 2 takes advantage of modern Python features, including static typing and improved asynchronous capabilities. This is like giving your app a turbocharged engine while ensuring it doesn’t fly apart on the highway.

But how does one make an AI understand their unique migration needs? Through something called prompts—basically, structured commands that tell AI what you want. The team tested three types of prompts:

Zero-Shot, One-Shot, and Chain of Thoughts: Different Paths to the Same Goal

1. Zero-Shot Prompting: This is akin to throwing a newbie into the deep end without giving them any instructions. You simply state what you need, and the AI takes a whack at it without guidance.

2. One-Shot Prompting: This approach gives AI a lifeline by providing an example of a successful migration. It’s like showing someone how to ride a bike before letting them try—much more effective and less likely to end in scraped knees.

3. Chain of Thoughts Prompting: Here, the prompt includes a step-by-step guide to gently nudge the AI toward the right solution. Think of it as giving your friend not just directions but a detailed travel itinerary.

Trial by Fire: Evaluating AI’s Code Migration Magic

The real test was to see how well ChatGPT could migrate code using these different approaches. The results were fascinating:

• Zero-Shot: Like a catapulting ball, it missed the target. The migrated application didn’t run; it flunked on connecting to the database and ignored Python’s typing features. Oopsie!

• One-Shot and Chain of Thoughts: These methods had more success. With examples and instructions, ChatGPT managed to migrate the application and even preserved its functionality. However, minor niggles like unused imports were still spotted—akin to forgetting to tidy up after redecorating.

In all cases, the quality of the migration was measured through various yardsticks like passing test cases and code quality checks. The One-Shot approach, with its inclusive example, turned out to be the MVP—most viable prompt—generating a working version of the application with fewer hitches than its counterparts.

Real-World Implications: How Developers Might Use This Magic Wand

You’re probably wondering, “How could this actually help me as a developer?” Picture this: Instead of spending your precious hours combing through documentation and manually updating each nook and cranny of your code, an AI assistant could handle the grunt work. This leaves you free to focus on the more creative aspects of software development or frankly, getting that much-needed coffee break.

Moreover, this isn’t just snake oil; it’s a peek into the future of software engineering where AI takes charge of routine tasks. This technology could potentially extend beyond just Python and SQLAlchemy to other languages and libraries, maybe even making the leap from your desktop to a broader cloud-based solution that global teams can access.

Key Takeaways

• AI’s Role Expands: Beyond generating code snippets or fixing bugs, ChatGPT is proving to be a valuable ally in automating tedious library migrations.

• Prompt Design Matters: How you talk to AI matters—giving it clear examples (One-Shot) or step-by-step instructions (Chain of Thoughts) leads to better outcomes than leaving it entirely to AI intuition (Zero-Shot).

• Real-World Potential: By handing over the migration burden to AI, developers can save time and reduce errors, focusing instead on innovation and creativity.

• Looking Ahead: While these first results are promising, there’s room for growth. Future explorations might include trying this approach on other programming ecosystems, prompting styles, and even integrating different AI models.

In conclusion, these advancements mark a thrilling entry into an era where AI dials down the manual labor in software maintenance, unlocking a productivity boost for developers. Just like how self-driving cars aim to redefine our commutes, LLMs are on the path to revolutionizing how we migrate and maintain code, one line at a time.

If you are looking to improve your prompting skills and haven’t already, check out our free Advanced Prompt Engineering course.

This blog post is based on the research article “Automatic Library Migration Using Large Language Models: First Results” by Authors: Aylton Almeida, Laerte Xavier, Marco Tulio Valente. You can find the original article here.