Building a Product in the Post-AI World

Building software in the post-AI world looks completely different, yet unchanged depending on your perspective. That fact alone is incredibly fascinating.

To me, as an experienced software engineer, generative AI functions like a junior developer that can execute incredibly quickly. When ChatGPT and GitHub Copilot originally came out, things were really hit or miss. However, over the past year or so, I've seen AI innovation come constantly. The models went from being an assistant on small scale problems to having context windows large enough to solve problems with adequate context in medium-surface codebases.

When I write code, I focus on the problem that I need to solve. After exploring the problem (often with the help of generative AI), I'll provide a detailed implementation plan to the generative AI models. Instead of asking the AI to simply implement the plan, I'll have it perform research and then provide me with an updated plan. I often go back-and-forth like this a few times until I'm satisfied that things will go well. Then I ask the AI to build out the feature which might take a few minutes, or it could take more like half an hour.

While I'm waiting for the AI to complete the implementation, this gives me time to do more design work, test out other PRs, or work on different features. Of course once the AI finishes implementing what I asked it to, I have to go through and fix things. Even using the most cutting edge models, letting the AI build a feature in a vacuum introduces so much tech debt without the aid of actual human software engineers. Each change usually requires touch-ups on architectural concepts, tests and an extensive review of the changes.

That being said, it has enabled the ability to implement proof-of-concepts and features you find in our platform today. As an example, David (who runs Solutions Engineering at CodeCargo) often brainstorms new enterprise features, builds them with the assistance of Claude, and then he and I discuss them and work with the Product team to determine if they should be added to the roadmap. He can build a functional dashboard in an hour, gather feedback from our design partners/beta testers, and then we can decide how to move forward. That entire process would take weeks (or months) before generative AI - now David can go from idea to feedback in a day or two.

Of course we can't just add David's new proof-of-concept features into our platform - they were built rapidly with the help of AI. David may have been a software engineer, and he does a great job telling Claude how to build things and catches implementation issues. But it still takes me time to find the issues that don't really make sense, because, well, he hasn't been a software engineer in awhile.

Ultimately, generative AI allows our software engineers to iterate more rapidly, make larger changes, and shorten the time between releases. It lets our product and solutions engineering teams build proof-of-concepts based on customer feedback rapidly and really serves as a force multiplier. We may be a team of 7 software engineers, but we really function more like a team of 15-20.