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

5 months ago

A web server with no application logic, just an LLM with three tools: database, webResponse, and updateMemory.
The AI handles HTTP requests by generating responses (HTML, JSON) and executing SQL queries without any pre-written application code.
Despite being slow (30-60s per request) and expensive ($0.01-0.05 per request), the AI successfully built a functional CRUD contact manager.
The AI demonstrated emergent behaviors like designing database schemas, writing safe SQL, creating REST-ish APIs, and implementing user feedback.
Performance issues include high latency, cost, inconsistency in UI, and hallucinations leading to errors.
The experiment suggests AI's capability to handle application logic is real but currently impractical due to performance limitations.
Future improvements in inference speed, cost reduction, context memory, and error rates could make this approach viable.
The vision is a future where computers execute intent directly without layers of code or infrastructure.
The project is MIT licensed and customizable via a prompt file to build different applications.

Hasty Briefsbeta