An Indie Developer's Case Against Even a Little AI in Your Code
Is there a 'right amount' of AI?
A skeptical developer wrestles with whether there's an acceptable minimum of LLM-assisted coding. When Big Walk's House House admitted using LLMs for under a percent of a percent of its code, the author initially shrugged—then reconsidered. LLMs have no thoughts, so they can't craft anything thoughtfully, and the more you delegate to a coding agent, the less you understand your own code. The least LLM code you can use is always zero.
It is very possible to do things without using an LLM, even things that are hard; and if anything, one of the harms of the "AI" political project is the propaganda machine dedicated to making you feel stupid and helpless.
- adamddev1
> LLMs, by definition, can’t make anything thoughtfully, because LLMs have no thoughts. ... apps that feel crafted, that show refinement, that feel intuitive—requires a human developer not only with taste, but with understanding.
Well said, thank you.
It saddens me that people can't understand this. Already there is the fallacy of "I can make apps with AI, therefore no one needs to code anymore," as if all coding was was making apps. But even the apps that people think are passable feel so fake, clunky and sludgy.
- eviks
> I admit I have trouble reconciling “I don’t like using libraries I haven’t written” with “I love having robots write code for me”.
(speculating) The easier way out is to understand that the first approach didn't make much sense either (even at face value "wanted to understand his code inside and out" can be easily solved by reading/testing the library code), so it's easy to get hyped up by another fad, there isn't much inconsistency between those at the deeper level of personal fascination
- WillAdams
For awards, the use of A.I. should be baked into the category descriptions/rules of eligibility up-front, and the organizers should be free to have a contest which is AI-free (in terms of contestants, I'd like to hope bots would always be banned from voting), or a contest which allows it, or separate contest categories for use and non-use.
For my part, I'm late to the game --- I used Google's LLM-enabled notebook to generate some sample DXF files early on, but I mostly hack away and copy-paste from Stackexchange and so forth --- flet.dev hitting 1.0 and offering an online studio did have me buying tokens and ultimately a subscription to make:
https://studio.flet.dev/apps/bicQWc3Aul
but for a long while it wasn't running and I was out of tokens, so I coded up:
https://studio.flet.dev/apps/46tp8GPq9g
while waiting for my next month of tokens to show up by copy-pasting and in a couple of instances asking Copilot about errors.
Currently trying to work out how to get the previews in the first app to look right, maybe adding an optional toolpath preview to the latter, and how to get a larger and more complex version up-and-running with an architecture I can wrap my head around using a toolchain I feel comfortable with (rather than a browser app, my typical development chain is local driven by a .text file: https://github.com/WillAdams/gcodepreview using a Literate Programming system).
- ionetan
I think different parts of systems carry different levels of risk. Some time ago I proposed a model for assessing the risk in a granular way. Putting a value (CCD - Cost of Cognitive Debt) on individual parts may foster healthy discussions within teams. Elaborating on it here:
- vips7L
Yes, 0.