AI doesn't generate working products, that's still your job

That landed because many people described the same failure mode from real projects. LLM output often looks fine change by change, yet the whole codebase slowly drifts into something subtly wrong. The common pattern was not obvious syntax bugs. It was incoherence. Requirements get “optimized away,” abstractions don’t line up with the actual problem, dead ends get pursued indefinitely, and a pile of individually plausible edits turns into a system that is harder to reason about than if a human had built it more slowly. Several commenters said this gets worse in long-lived or legacy codebases, where the model takes local shortcuts that pass superficial review but accumulate real maintenance and production risk. The strongest practical consensus was narrower and more useful than either the hype or the backlash. LLMs are very good at constrained work with clear rails. That includes prototyping, internal tools, simple CRUD apps, UI scaffolding, test generation, bug hunting, performance tweaks in isolated areas, and transformations like porting code or implementing well-specified tickets inside a documented architecture. They are much weaker at creating the architecture itself, preserving intent across months of work, and making the judgment calls that keep a product coherent as it evolves. The best reports came from people using models as aggressive assistants rather than autonomous engineers. They keep the design in their own head, force a planning step, review the implementation, and rely on tests and other non-AI checks. A second thread pushed back on the idea that there has been no real impact. Some teams said LLMs already let them ship more polished software, replace expensive internal tools, or clean up codebases they previously lacked time to improve. Others pointed to security work as a better fit than product creation, especially automated vulnerability discovery and mechanically scoped rewrites. Even there, the more credible version of the claim was not “AI builds products on its own.” It was that AI is useful when the problem is narrow, the success criteria are concrete, and humans still own the bar. The mood was skeptical of autonomous “vibe coding” and especially of the analogy that LLMs are just the next abstraction layer like compilers or high-level languages. People kept stressing the opposite. Compilers are deterministic. LLMs are probabilistic and will happily generate convincing nonsense. That difference changes how you manage risk. It also explains why a lot of people now see AI-generated output as raw material, not finished work. The sharpest takeaway was that prototypes are cheaper than ever, but the distance from prototype to product has not collapsed. If anything, sloppy AI use can make that distance longer by creating systems that look done before the real engineering has happened.