firehose> #llmops

HTML as Native Output

The output format a model is fluent in is the one its training data is saturated with — and for laying out text and images beautifully, that is HTML. The framing: if Markdown is the model's native language on the input side, HTML might be its native language on the output side. The stated mechanism is training-data composition — the web is full of HTML pages, so the model is excellent at the question "how do I take a pile of text and images and lay it out into something clear and beautiful," and emitting HTML lets it use its full front-end design muscle rather than being boxed into a format (like a .pptx object model) it has seen far less of.

The payoff has a second half that makes HTML special among native formats: it is friendly to both the model and the human. Markdown is friendly to the model but not to us (nobody enjoys reading a wall of text); a rendered HTML page is pleasant for both. So choosing HTML as the output isn't just "pick what the model writes well" — it is the rare format where the model's fluency and the reader's experience point the same way.

Claims


Linked from