Dialing Back LLMs: A Candid Developer's Journey

Host

Welcome back to our podcast! Today, we're diving into a fascinating and timely topic: the reality of using large language models or LLMs in software development. We're thrilled to have Alberto Fortin, a seasoned software engineer with over 15 years of experience, sharing his insights.

Expert

Thanks for having me! I'm excited to discuss my experiences with LLMs and why I've decided to dial back their usage.

Host

Absolutely! You initially embraced LLMs with a lot of enthusiasm, thinking they would revolutionize your workflow. Can you tell us what led to your shift in perspective?

Expert

Sure! As I started integrating LLMs into my production code, I ran into a lot of unexpected challenges. I was shocked by the poor quality of some outputs. It wasn't just about bugs; I realized that the code wasn't as maintainable as I wanted it to be.

Host

That sounds frustrating. So, you're saying that even though LLMs can generate code, it's not always clean or easy to work with?

Expert

Exactly! For instance, I often felt like I was on the verge of fixing an issue, only for a new error to surface. It turned into a cycle of fixing one thing and breaking another.

Host

I can see how that would lead to a lot of wasted time! You mentioned something about the 'productivity illusion.' Can you explain what you mean by that?

Expert

Definitely. At first, the excitement was palpable. When you get a feature to autocomplete or suggest code, it feels like the LLM is reading your mind. But over time, I realized that we were expecting too much from these tools.

Host

It’s like being seduced by the initial magic of technology, right? But you found that it's not always as effective as it seems.

Expert

Exactly! I had to make a mental shift. I started viewing the LLM as an assistant rather than the main developer. I’m the one in control, making the plans and decisions.

Host

That's a powerful perspective! So, you decided to take a step back and handle things manually?

Expert

Yes, I lost trust in the LLMs for big features. Now, I only use them for smaller tasks like refactoring. It’s much quicker and more reliable when I really understand my codebase.

Host

That makes total sense. So, for listeners who might be relying heavily on LLMs, what advice would you give?

Expert

If you're a confident developer and the LLMs aren't working for you, don’t feel bad about reverting to your previous methods. Use AI to augment your skills, but remember, you still need to make crucial architectural decisions.

Host

Such valuable insights! So, striking a balance is key here?

Expert

Absolutely! We should embrace AI as a revolutionary technology, but we have to be realistic about its current limitations.

Host

Thank you, Alberto! This discussion has been enlightening and will surely help many developers navigate their use of LLMs.

Expert

Thank you for having me! It’s been great to share my journey.