I finally got fed up with antigrav a couple of weeks ago - the rate limit issues were intolerable. Popped back to ask if things have improved?
What I’ve been doing in the meantime.
This is not a promo of any service, just a little feedback. And my pov on this is, I used to be a coder, but haven’t got the mental capacity for it anymore, so writing an app to sort out family photos for my father-in-law for fun, relying totally on ai to do the grunt work.
I’ve got a chatGPT account and they’ve released a product called codex which uses my existing, underutilised credit. It was really easy to point it at my antigrav folder and pickup the project. But I found it was quite slow doing fairly simple changes. So I got it to scan over the whole project that antigrav had produced, and look for way to improve the speed.
The results were not great.
It was at that point I realised that neither antigrav or codex produce code out of the box that has any measure of quality. It’s quite astonishing that, we now have the potential to write really great quality code everything, everywhere and we’re choosing to shrug shoulder and not. And don’t think that adding ‘always write good quality code’ to your custom instructions will fix it. The intelligence might be artificial but it’s also pretty dumb.
I then spent a couple of days getting it to rework the project to be more ‘ai friendly’.
I then thought about how to stop things getting in such a mess again. Installed the sonar extension and everything lit up like a christmas tree. So I delved into linting.
Found a couple of Rust linters that are much faster than ESlint - great - but also not as good coverage - boo. But found config for ESLint (and other js linters) - hooray! - but that made the js Linting sloooow - boo!
One of the main tools for making things better for ai seems to be putting limits on file/function size and complexity. That seems to be helpful because the ai can be more targeted making changes and, how shall I put this, not have a fit and randomly change stuff that’s nothing to do with the current change.
I diligently applied the linting across the project, and after another couple of days of codex finding and fixing, I had the 'perfect project’.
Then I started adding functionality. That’s when things got nasty. At the heart of my app is a workflow system. It’s there so I can have very slow running jobs like face detecting running in the background, but keep me up to date with what’s happening. As I asked Codex to produce a nice visual interface for the workflow, it became apparent that what antigrav had previously cooked up was, to be tactful, sub par.
Another couple of days of building new workflow system that didn’t totally suck. Another day getting it to remove the old workflow code that still wasn’t removed. I finally get back to a full, working workflow and press GO. Doesn’t work. It failed to migrate the functionality from the old workflow to the new, but it is able to go back through the repo pull out the deleted code and resurrect it. Press GO - doesn’t work. But I’ve now got visability of the workflow? Yes? Nope.
So I’ve replaced one truly exasperating experience with another. After yet another two weeks and consuming the water and power of a small town, it’s no further forwards. I mean, it is, but the goal I originally had is still not done. I discovered SKILLS in the meantime - they’re fun!
Maybe I’ve come to this all wrong. I was offered the ability to build apps without getting my hands dirty and yet the last few months have been 10% fast, exciting, build and 90% tedious unpicking of what the AIs have done. Am I doing something wrong, or does nobody else care about code quality any more?
More than that, when you look at the way that the ai’s build stuff, it’s just so dumb. It’s painful to watch. They generate code, they lint it, they rework it, they lint it, rework, lint…. We shouldn’t be accepting that. These things should be able to produce amazing code quality out of the box, without linting, or special instructions. They should be better than the best of us, not worse than the worst of us*. (*that’s me. If an AI is a worse code than me, that’s a big red flag). To be useful, efficient, an AI should understand our codebases, like really understand them. It doesn’t need every line of code to do that, just an awareness of the structure, the conventions, architecture, design, intention and vocab. It shouldn’t be throwing stuff at a wall to see what sticks.
It scares me the companies are using this stuff, and sacking coders as a result. Code quality was never great, but what’s the landscape going to look like in a couple of years time, when this drivel is everywhere.
I’m not writing this to have a moan, but out of genuine concern. We have a moment to raise the bar on coding, forever. But we need to choose that, demand that. I don’t care if it’s Google, OpenAI or somebody else who
long haul trucker, they can grind days - but no warnings! if they max out on context, the whole line is gone. and its common too where - they may not even pop up a download button so git connection is almost necessary. its more like “the slow cooker”. good spot where if you want a build cleaned up, refined, evaluated, or just need to back off a project because you’re going crazy - good way to pause and let it do its thing.
