I Stopped Trying to Learn Everything About AI

AI moves fast enough that it’s genuinely uncomfortable. New models, new frameworks, new papers, new agents, new benchmarks, new tools, new tutorials, every week. Try to learn all of it and you’ll run into an uncomfortable truth pretty quickly: you can’t.

I learned this the hard way while studying computer science and building ML projects on the side. At first I just collected topics: Python, ML, deep learning, NLP, computer vision, generative AI, LLMs, RAG, agents, MLOps, cloud, and probably ten more things I’ve since forgotten. The result was knowing a little about a lot without any real sense of how the pieces connected.

Things started making sense once I stopped asking “what should I learn next?” and started asking “what am I trying to build?” While working on a legal AI application, I naturally ran into OCR, NLP, embeddings, transformers, semantic similarity, summarization, retrieval, LLMs, APIs, and deployment, not because I was studying a curriculum, but because I needed each of them to get the thing working. That made them stick.

Real projects expose what tutorials tend to hide. A tutorial can make anything look easy: run three lines of code, get a prediction, done. Real projects are messier. The dataset is a mess. The package won’t install. The model behaves strangely. The output isn’t what you expected. The API fails somewhere in the middle. The deployment breaks for no obvious reason. That’s usually the moment you find out what you actually don’t understand yet, which is…