The future of software development
collaborating with AI

During our monthly workshops we exchange knowledge and try out new technologies in practice. This time Artificial Intelligence took centre stage, with one burning question: how can AI support us as developers, and where do we hit the limits?

It had been expected for a while. Last Friday at 10KB our first workshop on artificial intelligence took place. We were not there to talk about conquering the world with clever robots. We mainly wanted to find out how far you can get if you let AI help generate content, and where the limits suddenly appear.

Our practice project

A stripped-down version of Squaredle, a kind of word search. We dive in with the Cursor Editor, a fork of Visual Studio Code with three advanced language models built in: GPT-3.5, GPT-4o, and Claude 3.5 Sonnet. Three AIs in a single editor, loaded with solid features such as:

• AI-driven Autocompletion

  Cursor predicts not only words, but entire lines of code.

• Code Generation

  It can generate code fragments based on prompts, which is especially useful for quickly implementing repetitive or routine tasks.

• AI Debugging

  Cursor helps detect and quickly fix bugs by analysing the code and suggesting improvements and solutions.

• Codebase-Aware Chat

  The built-in AI assistant can search the entire codebase, making it possible to ask questions about specific functions, files or bugs. You can also highlight code blocks for more targeted answers.

That sounds as if with all those extras you can build software faster, smoother and more simply. Time to find out whether it works for us!

AI and coding: how well does that actually work?

It quickly becomes clear how powerful those AI tools really are. With Cursor's Composer feature you summon code at a crazy pace. "Move this variable for me", it just happens immediately. "Fix this little error", solved. It almost seems as if you can blindly build a working app without even glancing at the code. And oddly enough that works surprisingly well. AI does its thing, provides templates and examples that immediately serve as the basis of your project, and chronically politely throws a "Certainly!" on the end.

Too good to be true?

This is where it gets funny. Or frustrating, depending how you look at it: AI is too helpful and does not grasp its own limits. So you ask for a small change and suddenly it changes just a bit too much. Your app still works, sure, but behind the scenes all sorts of things are different, and that is not always an improvement. AI wants so badly to help that it calmly adjusts things you would rather have left alone.

Speaking of limits, AI will never admit that it does not understand something. Sentences like "I don't know what you mean" or "I'm not capable of doing this" are not part of this language model's vocabulary.

That simply means you always get an answer, whether it is correct or not. Clean code? That term means nothing to it. It hammers out code and blindly trusts that you will lintGlossary · In brieflintingLinting automatically checks source code for potential errors and violations of agreed coding rules, without running the application. The tool that performs these checks is called a linter.Read more⁠ it. Handy, until you have forgotten to turn that linter on of course.

AI is your colleague, not your boss

This workshop made one thing very clear: AI is not going to take our jobs. Junior-level tasks? No problem. We are happy to unload that repetitive work onto it. But you still have to keep a close eye on it. Anyone with real experience can smell when a tool is bluffing past its capacity limit, and that is exactly the expertise that sets a good developer apart. AI gives good tips, but you still have to understand what you are doing.

What does this mean for the future?

What AI brings about in a tool like Cursor Editor goes beyond convenience. Of course AI is strong at the basic work, but you can never miss a human eye to assess and refine the output. What you feed the tool, those instructions, equally asks for a clear goal in mind.

Potentially positive development:

Because tickets can be finished at a higher pace, you get breathing room. The gain sits mainly in the time freed up for testing and thoughtfully improving the stability of your codebaseGlossary · In briefcodebaseA codebase is the collection of source code used to build and maintain a software product or component.Read more⁠. In fact, the bar can go up.

Potentially negative development:

Exactly those entry-level tasks where AI is so strong can take away the chance for junior developers to learn and make mistakes. Companies push that type of job to automated systems faster. That squeezes starters, who now get less real code under their fingers in the early career years, so broader experience only becomes usable much later. More still, a company that later looks for a senior can really get stuck.

Want to know more about AI?

Ewout made a mini series of short videos about AI in software development, including tips and tricks. What works and what does not, and you will find everything on his LinkedIn.

Watch the mini series on AI on LinkedIn