Stop Overfitting Your Life

In cybersecurity, we aim to reduce uncertainty: fewer attack paths, fewer breaches, fewer 3 a.m. pages. I’ve realized the same instinct, applied to my own life, was limiting my perspective.

Left alone, I drift into routine. I’ll visit the same restaurants and bars, spend time with the same people, and talk about the same things day after day. My girlfriend is not wired that way. Since we started dating, I’ve tried many new foods, some of which I despised, and others I cannot imagine going without. She somehow convinces me to go ice skating every winter and paddleboarding every summer. Every anime I’ve seen, I watched with her. Most years, we go to the Poteet Strawberry Festival and the San Antonio Rodeo. Without her, almost none of those experiences would be in my rotation. Not everything stuck, but enough did to redefine “normal”.

Travel amplified that. First around North America, then Europe, and even the Middle East—each place surprised me. The things people value, family dynamics, how strangers interact, even social cues and gestures. Those differences changed my perspective, inspired side projects, nudged career goals, and subtly reshaped my values. While there are still many places I wish to visit, I feel very blessed with how much I have been able to see in such a short time.

The pattern showed up in my technical work, too, though it took me longer to notice. I’ve worked in security for years, but never cared much for machine learning. Even when large language models arrived, my interest was strictly practical: tighter prompts, RAG systems, a few agent harnesses. Then curiosity took over. While I’m no expert, I’ve learned much more than I ever planned about transformer architecture, unsupervised and supervised learning, and, most recently, reinforcement learning.

A key concept in reinforcement learning is the balance between exploration and exploitation. Rewarding the model too much for early success will not lead to strong long-term strategies. The model does well initially, but never grows beyond that level of capability. Similarly, allowing too much exploration leads to very creative outputs that never achieve much. The model must have short-term goals and incentives while pursuing long-term rewards.

I wasn’t always like this. In college, I treated non-technical classes as an obstacle. I skipped class to grind Hack The Box. I rolled my eyes at people who studied abroad—why pay to learn the same thing in a different place? In other words, I rewarded short-term exploitation over long-term exploration. While I was fortunate to find career opportunities with this skill set, I may have struggled if offensive security ever dried up.

“Entropy” is often described as a measure of randomness. In LLMs, entropy measures how spread out the model’s next-token probabilities are. Low entropy means confidence; high entropy means several tokens look plausible. Sampling tricks (like higher temperature or top-p) boost output diversity, which is why they often feel more “creative.” Consistently low-entropy, high-confidence outputs across varied prompts can signal overfitting. The LLM has memorized enough of the dataset to perform well on a specific task, but it cannot generalize to new challenges or offer different perspectives on existing ones.

Kids are a great case study in entropy. They ask wild questions, pick up languages fast, and combine ideas in ways that jolt adults. They haven’t collapsed to a single, tight hypothesis of the world. How do you fix an overfit model? You train it on diverse, noisy data. Humans have versions of this as well. Novel experiences push us to widen our state space and develop more general patterns to understand the world. Even while sleeping, dreams remix our experiences into unexpected combinations.

I have seen quantifiable improvements in my technical skill set over the last few years, and I attribute them, in part, to increased exploration. Security may depend on certainty, but people thrive on creativity. So go with her to the weird restaurant. Study a random new topic. Say hi to a stranger. This mindset is the only way to keep from overfitting.