Lately,

I’ve been paying more attention to what’s happening in the AI space.

Maybe because of all the hype that surrounds it.

Maybe because of the anxiety of whether AI will take my software development job.

That's why I’ve decided to spend some spare time during the Xmas holiday to explore the state of AI software development tools.

This email combines:

• My 1+ years of experience using a paid version of GitHub Copilot.
• 5+ hours of YouTube videos I’ve consumed in the last 2 weeks.

So, let’s dive in with the overview of the tools.

For the purposes of this email, I’ll split the AI tools into 3 categories:

1. AI IDE Extensions - extensions for your current IDE that predict and autocomplete the code you’ll write next.
2. AI Code Editors - they can create and modify multiple files in one go.
3. AI Agentic Tools (MVP producers) - they can create a complete first version of your application.

​

1. AI IDE Extensions

In the first category are AI tools that serve as an advanced autocompletion tool.

Based on the surrounding code and context it gets, it predicts and suggests the code you’ll write next.

My primary experience with these tools is through using GitHub Copilot.

The good:

• Increases productivity.
• Suggestions are good when it's clear what code comes next.
• Good for boilerplate code.
• Good for generating starting code.

The bad:

• Can struggle to come up with innovative code that requires more knowledge of the domain or codebase.
• There is also a GitHub Copilot chat extension. But the results are average. It needs a lot of context, referenced files, explanation, and a perfect prompt to output the code I want. At that point, it’s easier for me to write the code myself. It’s similar to ChatGPT. But ChatGPT does a much better job.

Some alternatives to GitHub Copilot are:

• ​Codeium (extension for Visual Studio, VS Code, and Rider)
• ​Supermaven (extension for VS Code and Rider)
• ​Cody (extension for Visual Studio, VS Code, and Rider)

​

2. AI Code Editors

In the second category are tools that can create and edit multiple files at once, and have more knowledge about your codebase.

The most popular at this stage are:

• ​Cursor - Fork of a VS Code that offers advanced AI capabilities.
• ​Windsurf - Fork of a VS Code that offers advanced AI capabilities.
• ​Cline - technically an extension for VS Code, but it’s capable of creating/editing files, executing commands, using the browser, and more with your permission every step of the way.

Those tools look impressive. Here’s the Windsurf promo video.

The good:

• The price for Cursor is $20/month for the Pro plan, which is acceptable given the productivity they claim you’ll have.
• The price for Windsurf is $15/month for the Pro plan, although there are some limitations around Cascade credits.
• Based on demos, it can speed up coding tasks a lot.

The bad:

• The price for Cline varies since you can use any LLM API. Based on the demo and discussion I saw online, those API calls add up quickly. So the daily cost can be easily $20+.
• I’m somewhat skeptical about giving AI access to the file system, web browsing, and access to all of my code.
• You lose a lot of features that Visual Studio and Rider have.
• Due to licensing rights, support for debugging of C# code is limited inside Cursor and Windsurf.

My goal for 2025 is to try those tools. And then I can conclude better on how good they are.

​

3. AI agentic tools (MVP producers)

The final category represents the AI tools that are more autonomous in producing applications:

1. You give them prompt and instructions.
2. The tool creates a bunch of files and MVP of an app.
3. You then iterate on further features and fix bugs.

The most popular tools right now are:

• ​Devin​
• ​v0​
• ​Bolt​
• ​Replit​
• ​Lovable.dev​

Of those tools, the one you’ve probably heard of is Devin.

Announced early in 2024, it received a lot of media attention. However, the Internet of Bugs YouTube channel exposed their demo as fake.

Here were my thoughts at the time when Devin's hype was at its peak:

“This is an AI agent that works like this:

1. You give him a command.
2. He goes on to solve it alone.
3. It has its own built-in browser to get the documentation, terminal, and code editor.

The only problem? There is no release date for it yet.

And uses print statements for debugging.

So I will skip analyzing it.

For now.”

Well, it was made available last month. For $500/month.

I haven’t tried it since it’s too expensive.

But Primeagen used it on his stream to create a web version of a ping-pong game.

​And Devin failed miserably.​

After prompting Devin for 2 hours to create a basic version of the game, Devin struggled to make any meaningful progress.

The good:

• You can create the first version of the app quickly.
• Useful for showing prototypes to customers and getting feedback on the ideas.

The bad:

• Even though you can create the first version fast, the tools soon get stuck and start hallucinating responses when fixing bugs and creating additional features.
• AI struggles to overcome those challenges without a human developer in the loop who knows the codebase, context, and requirements.
• Looks like most of those tools are frontend-oriented. From lovable.dev: “Lovable builds front-end using React, Tailwind & Vite, and can connect to OpenAPI backends. Support for data persistence and authentication is in alpha, using Supabase.” Or from Bolt: “Bolt does not support languages like Python, C#, C++, or any other non-web languages.” But if you are a .NET developer, this is actually not a bad thing?

​

My predictions for 2025

• Most software jobs are safe.
• But, AI is improving at lightning speed, which is both scary and exciting.
• Windsurf and Cursor will gain more popularity.
• Code writing will become a commodity skill.
• The AI tools will raise the bar of what it means to be a software developer. Blindly producing code without other skills soon won’t be enough.
• Due to current limitations, good software developers can’t be replaced with AI.
• But this means that every developer should start upskilling or risk losing his value on the market.
• To stay relevant, education is more important than ever. Mastering software engineering skills such as software architecture, security, CI/CD, and automated testing is a must.
• My goal for 2025 is to double down on learning more about software architecture, security, and AI tools.
  ​

What are your goals and predictions for 2025? Reply and let me know.

Have a nice weekend.

Kristijan

​

​