Critics demanding vibecoded Photoshops are themselves making unverified claims; AI lowered the cost of Level 1 typing, not Level 2 verification or Level 3 architecture.
Key Takeaways
The author’s SoulPlayer ships only after four bit-identical reference implementations agree, backed by a 90-test Python harness emitting C64 assembly and binary.
Three distinct levels exist: Level 1 typing, Level 2 test/verification, Level 3 architectural decisions. AI only reduced cost at Level 1.
No vibecoded Photoshop, Excel, Maya, Blender, compiler, or database exists because those artifacts require Levels 2 and 3 to hold together.
The “vibecoded” accusation is itself unverified output: no definition, no falsification, no evidence – the accuser performs the same pattern they attack.
The shame economy suppresses transparency; people fear sharing AI-assisted work not because it is wrong but because the accusation is cheap and the defense is costly.
Hacker News Comment Review
Commenters broadly accepted the Level 1/2/3 framing: AI kills typing toil but does not excuse you from architecture, testing discipline, or product judgment.
Disagreement centered on the claim’s scope: vibecoded apps clearly exist (wellness trackers, storefronts, todo tools) but nothing requiring deep architectural integrity has emerged from pure prompting.
The testing bottleneck drew practical focus – LLMs underperform on test generation without significant human scaffolding and tooling machinery around them.
Notable Comments
@hosaka: Built Hosaka Studio with Claude Code across six Linux distros, X11, and Wayland – three months of hard professional work, explicitly not accessible to nonprogrammers.
@fugaziboutit: “you can gauge AI progress by the velocity at which people move the goalposts.”
@themgt: The linked article returned a 503 during peak HN traffic – noted as accidental performance art.
