Embracing Imperfection: The 3-Month Rule for Learning in Tech Development

In the world of tech startups, there’s a well-known mantra from Paul Graham: “Do things that don’t scale.” This advice often sparks debate about its implementation, especially among developers. After spending the last eight months building my AI podcast platform, I’ve created a straightforward framework that embraces this principle. I call it the 3-Month Rule: for any non-scalable approach I adopt, it gets a three-month trial period. If it demonstrates its worth, it gets refined; if not, it gets the boot.

Challenging the Ideal of Scalability

As engineers, we are conditioned to design for scalability from the get-go. Our minds are trained to think in terms of elegant architectures, microservices, and distributed systems capable of supporting millions of users. However, this mentality can be counterproductive in startup environments. In many cases, the drive for scalability amounts to little more than an expensive form of procrastination, where we optimize for nonexistent users and imaginary problems.

My 3-Month Rule encourages me to create straightforward, even rudimentary, solutions that allow me to release features quickly. This approach fosters real feedback from users and a clearer understanding of their needs.

Current Hacks to Accelerate Learning

Here are some of the unconventional methods I’m employing and the insights they’re generating:

1. Single VM Setup

To keep costs low, I run my entire stack—database, web server, background jobs, and caching—on a single $40/month virtual machine (VM). While this lacks redundancy and relies on manual backups, it has provided invaluable learning. Within just two months, I grasped my actual resource usage better than any planning document could. My AI platform utilizes a mere 4GB of RAM at peak times. The complex Kubernetes set-up I nearly launched would have been a mission to maintain without addressing any real needs.

When the system crashes (yes, it has happened a couple of times), it gives me authentic insights into what fails—and often, it’s not what I anticipated.

2. Hardcoded Configurations

My code features numerous hardcoded constants, like:

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

While this might seem inefficient, it allows for rapid configurations changes and quick