Most explainers hand you the finished Transformer like a museum placard: here is the self-attention, here are the queries, keys, and values, and here is why each part matters. "Before Q, K, and V: Reconstructing the Transformer" takes a different route. It asks the uncomfortable question that most technical writing skips: why does this architecture look the way it does? That is a more useful starting point, and it deserves more attention.
The gap between using a tool and understanding its design is where most of us get stuck. You can read about Unlock LLM Training: A Practical Guide to Distributed Algorithms and follow along with the mechanics, but if you never ask why the pieces were arranged in a particular order, you are memorizing a map instead of learning the terrain. The same applies here. Reconstructing the Transformer from first principles is not an academic exercise. It is the difference between being a passenger and being the person who can explain why the engine is in the front.
What makes this approach valuable is that it treats the architecture as a set of design decisions rather than a fixed truth. The authors are not just describing what Q, K, and V do; they are inviting you to imagine a world where those choices were made differently. That is a more honest and more useful way to learn. It also connects to something we have explored before in Exploring Paragraph Structure: How LLMs Navigate Token Space, where the focus shifts from what a model outputs to how it moves through the spaces it has learned. Both articles share a core belief: understanding the underlying structure matters more than memorizing the surface details.
For readers who are trying to move beyond copy-pasting code or tweaking hyperparameters without understanding why, this reconstruction approach is the antidote. It forces you to confront the assumptions baked into the architecture. Why do we need separate projections for queries and keys? Why is the feed-forward network placed where it is? These questions do not have obvious answers, and that is the point. The moment you start asking them, you move from passive consumer to active participant.
Our take is simple: if you want to work with these models rather than just use them, stop starting with the finished architecture. Start with the problem the designers were trying to solve. The specific takeaway here is that the Transformer is not a collection of clever tricks; it is a series of reasoned compromises. And the only way to see those compromises clearly is to rebuild them yourself, step by step. That is not just a better way to learn. It is the only way to genuinely own the knowledge.