The Reality of AI Coding

There is something that's been bugging me recently about all the online discourse surrounding AI coding: as time goes on, it feels like less and less of it applies to me.

Let me offer an example:

The newest industry craze seems to be commanding a bunch of parallel AI coding agents. We now have Jules, Codex, Devin, Copilot, and soon Cursor offering the ability to spin up dozens of super-intelligent autonomous agents across your codebase to complete the work of 10 engineers in 1/10th of the time (their words, not mine).

Large codebases is the key, I think. For weekend projects where I couldn't care less about long-term maintainability, I really can make great progress simply by prompting to a working solution. I usually don't even look at the code. But, for my day job, I've converged on a clear set of task-types where I leverage AI, and pretty much avoid it for the rest. It's just not worth the cortisol spike to spend 5 minutes crafting a detailed prompt and attaching all the right context, only to get back a useless result 90% of the time. The gap in AI capabilities between green-field work and mature codebases is astounding.

To be clear, I would actually consider myself a pretty strong AI optimist. It is excellent for greenfield work, for creating a dozen similar API endpoints in minutes, for writing docstrings, for going back-and-forth on tech specs, and for lots of other things too. I think any developer not using AI today is missing out big time. Most importantly, I believe that AI models will continue to improve substantially from where they are today.

But, to date, I have not met a single developer working on a mature codebase who has not had to coddle and meticulously review AI output for anything beyond mostly trivial or greenfield tasks (and I've talked to quite a few!). Many tell me that the time spent prompting and reviewing combined with the high risk of failure and/or subtle bugs has made the cost-benefit calculation top-of-mind when they consider using AI for their work.

My point with all this is I believe we're headed for some sort of reckoning. With the entire industry hype-cycle going at full tilt, there is huge pressure to adopt AI coding everywhere at breakneck speed, fueled by immense FOMO that has gripped board rooms and c-suites who believe this will moonshot their business.

I think we are about to start seeing a lot more software that is less secure, less maintainable, less reliable, and less enjoyable to use. Worse, I believe we will see a cultural shift in programming that will deem all this as ok. Outweighed by the productivity gains. This cultural shift will be driven by the same hype-machine that's pushing autonomous superintelligent AI coding agents so hard today.

There will be more Shopifys. One thing I think the crazy AI optimists are correct about: we're still in the early innings. AI coding is here. It is useful. But the reality is not what you see online.