LLM — Principia

Learn · LLM

Decision models: what fits in one pass?

The new System 1 models promise fast, cheap decisions without a chain of thought. A small arithmetic test shows why that can be useful, where it can fail, and why the confidence score deserves a closer look.

V. Levy dit Vehel, PhD · 21 Oct · 8 min

Learn · Agents

In search for lost context

We increasingly treat agents as collaborators and expect them to remember our work together. Yet the models at their centre do not acquire lasting memories from ordinary interactions. Continuity has to be built: by deciding what information survives, how it returns, and who gets to change it.

V. Levy dit Vehel, PhD · 5 Sept · 13 min

Learn · Agents

Stop hard-coding the model: routing for the age of agents

A model router lets an application ask for an outcome instead of naming a model. That can cut cost and increase efficiency—but in long-running agents, a switch can negatively interact with the prompt-cache economics that made the run affordable.

V. Levy dit Vehel, PhD · 3 Sept · 9 min