Fundamentals still matter, but the center of gravity has shifted
It would be dishonest to say that understanding how models actually work no longer matters. Strong Python, real comprehension of how a transformer processes a sequence, how training and fine-tuning work, and what makes a model's outputs behave the way they do, still separates people who can debug a genuinely broken system from people who can only call an API and hope. When something goes wrong in production, when a model's outputs degrade, when a fine-tuning run isn't converging, when retrieval is returning irrelevant context, the people who can diagnose the actual cause are the ones with real fundamentals, not just familiarity with a library's documentation.
But it would be equally dishonest to pretend the field has not shifted meaningfully. A few years ago, a strong AI portfolio meant demonstrating you could train a model from scratch on a nontrivial dataset. Today, a large and growing share of real, well-paid AI work involves building on top of existing foundation models rather than training new ones: effective prompting and context design, retrieval-augmented generation, building and evaluating agents, and rigorous evaluation of model outputs against real failure modes. This is not a lesser skill set than training models, it is a different one, with its own genuine depth, and it is the one most entry-level and even mid-level roles now actually test for.