Do We Still Need Deep Learning?

Do We Still Need Deep Learning?

As a software engineer with years behind me, looking at AI feels strange.

My engineering instinct always tells me to start from the ground up. If you want to understand how things work, you study the foundations. For AI, that means deep learning: neural networks, backpropagation, matrix math, PyTorch. After all, Large Language Models are built on top of deep learning.

Yet every modern roadmap tells you to skip it.

Job posts don't ask for model training or gradient descent anymore. They want RAG. They want context engineering, agentic workflows, prompt evaluation, and vector search. For a while, this made me feel uneasy. It felt like learning web development by only tweaking CSS, without knowing how TCP/IP or databases work.

Then I remembered how we build software today.

Most of us never write custom database engines or balance B-trees in C. We use PostgreSQL. We learn indexing, transactions, and how to write good queries. The foundation model is simply the new engine. It was built by deep learning researchers, but our job as engineers is to build reliable systems around it.

RAG and context engineering are not just shallow wrappers. They are actually distributed systems and information retrieval problems dressed in new clothes. Latency, token budget, chunking strategies, and deterministic outputs—these are classic engineering challenges.

Deep learning is still good to know, at least conceptually. It helps you understand why an LLM hallucinates or why long context loses focus. But obsessing over the math before building anything is a trap.

Build something first. See where it breaks. The math can wait until you actually hit the wall.