ai agents / llm / software engineering
AI Is Still Bad at Creating Software
I've used AI to write over 3 million lines of code. It handles a scoped ticket end to end. When the work is novel, it adds repeated code and new patterns that break the ones you set, and everything still passes. I ran a grep on my own server code: 53 files each define their own sleep().
Transcript
0:00As someone who has used AI to write over 3 million lines of code, I have come to the conclusion that AI is still extremely bad at creating software.
0:10Now, AI is extremely, extremely good at scoped specific features.
0:15If you have a Linear ticket and it says do X, Y, and Z and this is what it should look like, it can do that end to end.
0:22Honestly, you don't even have to read the code.
0:24When it comes to something novel, when it comes to taking a step back and thinking about what we should build, AI is outright awful at it.
0:35I mean it is genuinely atrocious.
0:37It will introduce so much repetitive code that you have throughout your codebase.
0:41It will introduce new patterns that directly violate the patterns that you've established.
0:47And if you're not paying attention these can seep through secretly because everything still works.
0:54The tests still pass.
0:56The screenshots look good.
0:58Everything still works.
1:00It has this hidden like disgusting slop that seeps through your architecture.
1:06And you have to be careful.
1:07You have to be careful.
1:08You have to make sure you understand what you should be building before the AI builds it.
1:13For small features, no problem.
1:16For moderately large but well scoped features, no problem.
1:21But for large features and you don't already have the design in mind, be careful.
1:27Opus 5.5 will destroy your codebase.
1:316.1 Sol will slopify your codebase.
Join the conversation
Loading conversation…