Draft — this one is still being written. The outline below is where it stands.
The autocomplete notes elsewhere on this site go deep on each architecture. This article is the decision layer above them: how the families compared head to head, and how I landed on a direction for the first prototype.
What this will cover
- The baseline I set with n-grams, and why a floor mattered
- Where LSTMs held up and where they fell behind
- What autoencoders told me about how strings cluster
- Why transformers led on quality, and the compute cost of that
Full comparison coming — treat the list above as the table of contents.