The sinopia—the reddish preparatory drawing sketched onto the rough plaster (arriccio) before the final layer is applied—offers a surprisingly rich analogy for the early stages of training a large language model (LLM).

Just as the sinopia is not the finished fresco but a hidden guide, the initial phase of LLM training is not about producing polished outputs. Instead, it lays down rough structures: patterns, relationships, and broad outlines learned from vast amounts of data. These early representations are often coarse and imperfect, much like the quick, confident strokes of iron-oxide red that map figures and composition without detail.

The arriccio, the textured substrate beneath the final plaster, resembles the raw training corpus. It is uneven, noisy, and unrefined, yet essential. The sinopia emerges from this surface as a first attempt to impose order—just as the model begins to form statistical regularities from unstructured text.

When the artist later applies the smooth intonaco and paints the visible fresco, the sinopia is partially or entirely obscured, yet it remains foundational. Similarly, in later training stages—fine-tuning, alignment, and optimization—the early learned representations are not directly visible in outputs, but they continue to shape every response. The polished fluency of the model rests on those initial, hidden approximations.

There is also an element of commitment: once the intonaco is laid, the painter must work quickly, following the guidance of the underlying sinopia. In LLMs, early training choices—data selection, architecture, objective functions—constrain and guide everything that follows. Adjustments can be made, but the foundational sketch exerts a lasting influence.

In both cases, what is ultimately seen—the vivid fresco or the coherent generated text—depends on a prior layer of invisible planning. The sinopia and the early training phase share a common role: they are acts of structuring possibility before refinement, where rough vision precedes finished form.