Show HN: Why write code if the LLM can just do the thing? (web app experiment) (github.com)

A developer built a web server with zero application logic that asks an LLM what to do for every HTTP request, wiring it to three tools—database (SQLite), webResponse (emit HTML/JSON), and updateMemory (persist feedback). In a few hours the system produced a working contact-manager CRUD app: the model designed sensible schemas (types, indexes), emitted parameterized SQL, rendered responsive Bootstrap UIs, validated forms, returned REST-ish JSON, and even persisted data across restarts. Every page included a feedback widget the model could read and act on, so “make buttons bigger” resulted in UI changes. The experiment used a consumer LLM (e.g., Claude-3) and showed that application logic can emerge from tooling plus a prompt. But it’s painfully impractical today: requests took 30–60s (vs 10–100ms for conventional servers), cost ~$0.01–$0.05 each (100–1000× more expensive), and the model spent 75–85% of time “reasoning.” It hallucinated SQL occasionally (causing 500s), drifted on styles between requests, and had limited short-term memory. Significance: capability exists — LLMs can replace app code — but remaining barriers are performance, cost, context/memory, and reliability. These are trending better (faster inference, lower cost, larger context), so the experiment suggests we may be closer to “LLM-as-computer” than many expect, with the main remaining work in infrastructure and latency/cost optimization.