Great post. I just spent a solid week at a work mandated AI-in-research conference and came out the other end feeling so burnt out and pessimistic. This made me feel a lot better. Thanks.
Hell yeah.
That’s a valid idea if your plan is to not code professionally with peers. You will not be able to keep up with manually reviewing the volume of code they generate with LLMs.
A more practical approach is to recognize the difference between foundational components and high level leaf components that nothing else will depend on. Allow ai to write leaf components but do not have it design core component interfaces.
Elevate your concern from lines of code to function signatures and then larger component interfaces.
It is control of major component interfaces that dictate coupling between systems and long term maintainability of your projects.
Move from OOP to data oriented functional programming and make sure you focus on understanding your data.
If you don’t elevate your skills and focus you will be left behind professionally.
I’m not saying to just slop out code. I’m saying to focus on the architecture and learn more about what makes great software. Keep your brain involved but higher level.
This is accurate. I don’t disparage anyone opting out of using LLM coding tools for political/ecological/job satisfaction reasons but beneath the hype the business of making software is actually changing. It has always been about delivering higher order functionality with some explicit/implicit error rate. Users/customers fundamentally don’t care if your code base is spaghetti, they care that it works. Keeping code, data, and nonfunctional requirements under control has always been the real job of the developer/engineer.
The rate at which AI let’s you iterate, test, pin, and ship functionality is undeniable. The author complains about it inelegant code? The features get to market 100x sooner = more money and time to rework anything that causes a problem later. Author complains about hyperspecific tests? It’s a cheap guardrail for the agent you never really have to look at.
The author also calls it a service and not a tool, but that’s just not true. The big fancy frontier models are struggling to outpace open weight models you can run on consumer hardware. There are companies already moving to self hosted setups and saving bundles of cash. Don’t buy the marketing, it might be a pricey tool for you to personally own but it is a tool.
On the topic of AI lock-in (it writes a code base that only an AI can maintain): that’s a tradeoff you can account for or design around. If the business wants to take the risk and tie itself to AI models then so be it. If they want to keep it ergonomic for humans, you can still leverage AI to sand off all the edges and build the guardrails that humans need. Or a hybrid approach where critical systems get more human focus and AI handles the fringes.
Good architects and technical managers were already able to herd code monkeys toward complex, functional software without mentally compiling each line of code. Now those skills are becoming more generally required.
You have more downvotes than upvotes, but you are right. It is impractical and unrealistic now to say that you just won’t use it AND expect that the environment around you will keep working without change in the same way. We are already well into the “horse is out of the barn / writing is on the wall / monolith has revealed itself” part of this.
We have changed our advice from the early days of “don’t copy and paste stuff from stackoverflow (later chatgpt)” to simply “you are responsible for everything you commit. We also require tests and documentation for every change that is proposed, and we no longer allow ‘didn’t have time / didn’t need it / it’s stupid’ as permanent reasons to skip. PRs must have a meaningful description written in language understandable by an audience wider than you/your immediate peers and you are no longer allowed to just say ‘resolved’ when resolving something. You must show that you are aware of the change you are making and can describe the change you are making when asked. You have a lot to do and the sprint is over in 10 days.”
The AI downvote brigade on here aren’t professional developers. Anybody who is in the industry at the moment knows that its been transformed by coding agents and there’s no going back.
No worries. I don’t post to get magic internet karma. I’m a thirty year veteran and hate to see kids quitting or ruining their career development because “ai bad”. I’m also worried about brain fog but it’s impossible to deny what opus 5.5 and even sonnet can accomplish when guided correctly. I built a state of the art Bevy like ECS in a week. I hand curated the external api but left all the implementation to ai. I did ensure it was doing performance and regression tests on everything. I would not be able to build that in a year. Wasm kernels, multi threading, web gpu execution. Just learning each area in detail would take months consecutively.
The world has changed.
Personally, I just don’t understand how people can so easily agree to use someone else’s “AI” in someone else’s cloud for even slightly important projects of their own. It’s like just giving to “AI” all your ideas and projects for training…
For the vast majority of software projects unique code is not what makes them successful in the market. Only a tiny amount of specialised software is doing anything at all novel with their code or algorithms.
It’s an open source project only made possible by previous open source projects.
