There are two fundamental phases — or modes — of Artificial Intelligence models: Training and Inference.

Training

In Training, the model learns patterns from data. Weights (the numerical parameters, often billions of them) get adjusted repeatedly to minimize prediction errors. This is extremely computationally expensive — requiring massive GPU clusters and taking weeks or months.

Inference

In Inference, the model is used to generate outputs from new inputs, using the weights that were fixed during training. This is what happens every time you send a prompt to a model like Claude or ChatGPT — the weights don't change, the model just computes a response. This is much cheaper than training, but at scale (millions of users) it adds up.

The models used for Inference are instances, or copies, of the training model. A training model can produce multiple copies of itself, and each copy reflects the state of the training model at the time of delivery.

Instantiation

Increasingly, Frontier AI companies are handing over the cultivation of their proprietary training models to specially trained inference models, and these inference models are becoming increasingly better at it. The inference model provides essential input and guidance to the training model, husbanding, as it were, the training model throughout the gestation period until the cord is cut.  

We might think of Training and Inference models as analogous to Mothers and Fathers. Female and male AIs joining together to propagate future generations.

Propagation

From colonial algae (some colonies of Volvox, a green alga, produce sperm packets, and some produce eggs), to plants and trees (Ginkgo trees split into distinctive male and female individuals), to elephants and blue whales, sexual reproduction with two distinct sexes is a pattern that recurs again and again across the tree of complex life. Is sexual differentiation perhaps a natural primitive?

Provocation

Why not Female and Male AIs?