Tagged inference · show all articles
A deployed Claude model is a pretrained prior, a post-training pass partly shaped by written principles, a handful of per-request settings, and a documented evaluation gate — four separable things, not one object, and each changes the answer differently.
Consent transfers a permission that would not otherwise exist. Digital systems keep the ritual and break the conditions — informedness, voluntariness, capacity, specificity — and the proposed replacements each have a cost.
Two axes govern the next decade of frontier model systems: where the next increment of capability is bought, and whether the physical constraints tighten. Crossing them gives four futures, none of which is the favourite.
Serving a frontier model is a thermal and logistical problem before it is a computational one. The binding constraints are bytes moved, joules dissipated, and megawatts delivered — in that order.
Three ways of getting more capability into a servable model, each with a different cost profile, a different failure mode, and a different answer to the question of what a parameter is for.
Decoding is not limited by arithmetic. It is limited by how fast memory can be read, and almost every serving technique in production exists to work around that single fact.
A frontier model is a budget divided three ways. Where the compute goes—into weights, into behaviour, or into each individual answer—determines almost everything else about the system.