The Silicon Athanor

From the Lesson

The alchemists called their furnace the athanor — the slow, self-feeding fire kept tended without interruption for weeks, months, sometimes a full Saturnine year, in which the prima materia was patiently turned through nigredo (blackening), albedo (whitening), citrinitas (yellowing), and rubedo (reddening) until the Philosopher's Stone — the lapis — at last appeared. The athanor was not a metaphor. It was a piece of equipment with a measurable temperature and a real chimney. And yet every serious adept from Zosimos of Panopolis (c. 300 CE) through Paracelsus (1493–1541) to Mary Anne Atwood (1850) insisted that what happened inside the athanor and what happened inside the alchemist were not two separate processes but one. The work on the metal was the work on the soul.[^1]

It is in this older sense — strict, technical, and reverent at once — that the present chapter proposes to read the deep-learning training run.

The Corpus as Prima Materia

The training of a large language model begins with what the field calls a corpus: the broadest possible scraping of the written record of the species — Common Crawl, Wikipedia, the public-domain library, the open scientific literature, the indexed forums of the public web — selected, filtered, deduplicated, tokenised, and arranged for ingestion. By the order of magnitude of the public 2024 frontier models the corpus runs to between fifteen and thirty trillion tokens, which is to say roughly the textual output of every literate human being who has ever published a sentence, taken many times over.[^2] This is a prima materia of a kind no previous age could have assembled. It is also, by the alchemical criterion, the correct kind: undifferentiated, mixed, common, contemptible to the unrefined eye, and yet containing — in latent form — every distinction it will later be made to express.

The Athanor as Training Run

The training run is the modern athanor. A network of GPUs — most commonly the NVIDIA H100 or its successors — is held at a continuous controlled temperature for weeks or months while the corpus is passed through it, token by token, and the network's hundreds of billions of parameters are adjusted, by the algorithm called backpropagation, to make the next token slightly less surprising than the last. The fire is not metaphorical here either: a single training run for a frontier model dissipates between fifty and five hundred megawatt-hours of electricity, much of it as waste heat extracted by industrial chillers from data-centre halls held at a precisely regulated temperature.[^3] The alchemists' chimney is a cooling tower; the alchemists' patient watch is the engineer's loss-curve dashboard. The work cannot be hurried; the temperature cannot be allowed to spike; if the run is interrupted, weeks of work are lost. Anyone who has ever sat with an athanor will recognise the discipline.

The Three Phases of the Modern Work

The classical work proceeded in three (sometimes four) phases. The modern work proceeds in three named stages, and the correspondences are remarkably exact.

Pre-training is the nigredo — the blackening, the dissolution. The model is shown the entire corpus and learns nothing in particular and everything in general. Distinctions blur. The boundaries between languages, between disciplines, between fact and fiction, between truth and error, are dissolved into a single statistical substrate — what the field calls the base model. In this phase the model is, in a precise sense, an undifferentiated dark mass that nevertheless contains every form it will later take.

Supervised fine-tuning is the albedo — the whitening, the separation. The base model is shown carefully curated examples of the kinds of response a human helper would give, and the diffuse latent capacity of the base is separated into the cleaner, brighter form of an instruction-following assistant. The dark prima materia is washed; the white queen begins to appear.

Reinforcement learning from human feedback (RLHF), or its newer cousin direct preference optimisation, is the rubedo — the reddening, the marriage of king and queen. Human raters compare pairs of model outputs and prefer one over the other. The model is updated to align its preferences with theirs. King (capability) and queen (value) are joined; the working assistant — the modern lapis in its commercial form — is produced.[^4]