Developer Traces a Year of ML Projects Back to a Single Evolving Algorithm

A developer reflecting on the past year noticed that multiple personal projects — from a NEAT-trained Pong paddle to a Rust-based creature genetics simulator and an autonomous Asteroids autopilot — all converged around the same core idea of letting algorithms find solutions independently. The pattern became clear during a CartPole reinforcement learning project last weekend, where a PPO agent achieved a perfect 500/500 reward score in just 39 seconds. Rather than collecting unrelated ML experiments, the developer realized each project was a new context for handing the same algorithmic approach a different problem. The current goal is EIC Auto, a project aimed at training a single RL agent to play the game 'Everything Is Crab' by learning its mechanics through experience rather than explicit instruction. The developer describes the recurring motivation as chasing the brief but compelling moment when a model figures out something it was never directly taught.