My Workflow for Understanding LLM Architectures

There's a quiet irony buried in this piece: the people building the most talked-about AI models are writing thinner and thinner papers about them. The author, who's spent months sketching architecture diagrams for talks and his LLM gallery, says he used to start with technical reports. Not anymore. Industry labs releasing open-weight models increasingly ship terse documentation, sometimes little more than a blog post, leaving the actual mechanics of the model unclear.

So he's built a workaround. If a model's weights land on Hugging Face and the architecture is supported in the transformers library, you can skip the marketing copy entirely and go straight to the config file and the reference implementation. Code doesn't have an incentive to oversell itself. It either runs or it doesn't, and every layer, every attention variant, every normalization trick is sitting right there in the source, whether or not anyone bothered to write a sentence about it.

This only works for a specific slice of the AI world, though. Closed shops like OpenAI, Anthropic, and Google keep both their weights and their internals locked away, so there's nothing to inspect. The whole method depends on the open-weight ecosystem continuing to exist and continuing to publish real, runnable implementations rather than just numbers on a leaderboard.

What's notable is that the author isn't pitching this as something to automate away. He explicitly frames the manual slog through config files and code as the point, not a chore to eliminate. If you want to actually understand how a mixture-of-experts routing scheme or a new attention mechanism works, tracing it by hand through source code apparently teaches you more than skimming a diagram someone else already drew. It's a small, almost old-fashioned argument in a field obsessed with abstraction layers and one-click tooling.