The 3-Month Experiment: A Practical Approach to Unscalable Development

When it comes to launching a startup, many of us encounter wisdom from entrepreneurs like Paul Graham, who famously advises, “Do things that don’t scale.” But translating this advice into actionable steps, especially in coding, often feels ambiguous.

After eight months of developing my AI podcast platform, I’ve established a straightforward methodology: every quick, unscalable hack gets a trial period of three months. At the end of this period, each solution either proves its worth and is refined, or it faces the chopping block.

As engineers, we are conditioned to prioritize scalable solutions from the outset. We create sophisticated architecture encompassing design patterns, microservices, and distributed systems—tools meant for managing a vast number of users. However, this perspective can hinder the agility vital for startups. In fact, focusing on scalability can lead to costly delays, as we often strive to optimize for an audience that may not yet exist, addressing challenges that could be mere hypotheticals. My three-month rule compels me to produce straightforward, albeit less-than-perfect code that ultimately delivers insights about user needs.

Current Hacks That Enhance Learning, Not Just Functionality

1. Single-VM Setup for Everything

I’m running my database, web server, and background processes—all on a single $40-a-month virtual machine (VM). While this approach lacks redundancy and relies on manual backups, it has offered unparalleled insights. Within two months, I discovered firsthand my actual resource needs, revealing that my AI-centric platform only requires 4GB of RAM during peak usage. The complex Kubernetes setup I nearly constructed would have simply managed idle containers.

When the system crashes (which has happened twice), I gather real-time data about the failures. Surprisingly, the issues rarely align with my initial expectations.

2. Hardcoded Constants for Configuration

Instead of using environment variables or config files, I have hardcoded constants throughout my code:

python
PRICE_TIER_1 = 9.99
PRICE_TIER_2 = 19.99
MAX_USERS = 100
AI_MODEL = "gpt-4"

This might seem primitive, but the benefit of this approach is significant. I can search my entire codebase for any variable within seconds. Each change to a configuration is registered in git history, ensuring that every adjustment undergoes a review (albeit by me).

The time invested in creating a dynamic configuration service would