Paul Graham posted that if he were 17, he would learn how to build LLMs from scratch. Lars Faye published an article arguing that coding expertise is going to collapse from AI reliance. The two threads collected 603 and 447 comments respectively.

The disagreement is whether skill loss matters if the code ships. Graham's position implies understanding the tools themselves is the valuable skill. Faye's article frames AI reliance as preventing the development of expertise.

The top concern in both threads is whether developers who rely on AI tools can debug code they did not write themselves. Comments argue that tools let builders ship faster but may create a generation unable to fix what breaks when the model fails.

The stakes for small software builders are which skills to invest in. If AI tools are temporary scaffolding, time spent learning fundamentals pays off when the tools fail. If they are permanent infrastructure, time spent building them or understanding their internals is the better bet.

What remains unresolved is whether this is a new problem or the same one every generation faces with new abstractions. Nobody in either thread has measured whether AI-assisted developers actually perform worse at debugging, or whether they simply debug differently.