I've been spending a lot of time this summer working through data science concepts: statistics, probability, Python libraries, and increasingly, machine learning. I wanted to write down some of what I've noticed about how I learn, because I think the process is just as interesting as the content.
The biggest thing I've had to unlearn is the instinct to jump straight to a model. Early on, I'd get excited about a dataset and immediately ask: can I train something on this? The better question turned out to be: what does this data actually represent, and what are its limits? A dataset of emotional language, for example, isn't a clean sample of all human emotion. It's a slice, shaped by who wrote the text, when, in what context, and on what platform. Ignoring that produces results that feel rigorous but aren't.
That came up directly in my research on AI and emotional distress recognition. I was looking at how AI systems identify subtle emotional signals in text compared to how humans do it. The AI tends to perform well on explicit language, like "I feel anxious" or "I'm overwhelmed," but struggles when distress is implicit: coded in dry humor, understatement, or cultural idiom. Training data doesn't capture what's left unsaid, and that turns out to matter a lot when the stakes are someone's mental health.
Another thing I've found is that the best way to actually understand an algorithm is to implement it from scratch before touching a library. I spent a week at UT Austin writing simple decision trees, perceptrons, KNN, and reinforcement learning algorithms, which gave me a great foundation in ML. Switching to libraries increased both the speed and the accuracy of my results, but now I understood why specific things, like the learning rate, mattered. Just calling libraries never let me see that. There's something about building the tool yourself that makes the abstraction click.
There's still a lot to learn. But I think the deeper skill, the one that makes data science actually useful rather than just impressive-looking, is skepticism. Not about the methods, but about which questions are worth asking in the first place, and what the answers can and can't tell you.