Infrastructure-as-code went through this exact cycle. Declarative specs were supposed to replace manual config, but Terraform still needs state files because specs drift from reality. Prompts have it worse since you can't even diff what changed between two generation runs.

The 8 years part is the real story. A lot of founders we talk to have the same pattern — technically capable, clear idea, but never shipped. The common explanation is "I was waiting until I had more time" but when AI removed the time constraint, the same projects still didn't ship. What actually changed here wasn't just velocity, it was that vibe coding lowered the psychological cost of starting on something uncertain. The spaghetti codebase problem is real but secondary — you can refactor bad code. You can't refactor years of not starting.

npm could solve half of this by letting packages opt into OIDC-only publishing at the registry level. v1 already had provenance attestations but the registry happily accepted the malicious publish without them.

The token efficiency argument only holds for teams paying per API call. Cursor, Copilot, and most tools developers actually use are flat subscription. On a flat monthly fee, there's no economic pressure toward brevity in generated code.

Apparently there’s no economic pressure toward brevity in LLM generated slop comments either.

And no, they’re not flat subscriptions. Use more tokens and your quota is gone faster.

Not true with github copilot. Cost is per prompt no matter how many tokens the prompt uses. Which can vary by 2 or 3 orders of magnitude...

This is what happens to every software valuation when production cost drops to near zero. The moat was never the code. Companies still pricing themselves on engineering effort are going to have a rough conversation with investors once any competitor can replicate the core product in a weekend. The only things that still hold are proprietary data and distribution lock-in.

Your plumber story is exactly what trips up most vertical AI pitches I see. The founder assumes every missed call is lost revenue, but for capacity-constrained shops, a missed call is just triage they did not have to do. Curious if anyone has seen an AI receptionist actually grow a shop's revenue vs just adding a layer on top of an already-full queue.

Satellite imagery companies spent a decade learning that raw data is a commodity and the margin lives in the analytics layer. If you're explicitly ceding that layer to partners, what stops a well-funded competitor from replicating the pad network and undercutting you on collection price?