Do you still read the code?

The piece interrogates whether reading source code remains necessary now that AI can generate much of the implementation. It distinguishes two engineering mindsets: accelerators use AI to speed translation from intent to code while retaining deep understanding, reading generated code, explaining design decisions, and bearing responsibility for maintenance; vibecoders delegate implementation entirely to AI, investing instead in precise specifications, regeneration, and evaluation while treating code as disposable. Each choice carries costs: accelerators risk cognitive debt when generation outpaces comprehension, vibecoders risk intent debt and dependence on opaque models, and both face concrete failure modes from agent errors and context drift.

Grounded in the idea that programming is theory building, the piece emphasizes that the real product is the model or understanding that lets people explain, adapt, and extend software; code is a visible but often misleading artifact. Implementation work refines that theory, so abandoning reading can erase necessary knowledge. The practical consequence is organizational: teams must declare expectations and boundaries about who reads, owns, and maintains generated code. Without explicit agreements, unfamiliar maintenance burdens and mismatched assumptions will cause fragile systems; deliberately reading generated lines is presented as a reliable guard against such drift.