Infrastructure

Phi-3 Mini runs on a laptop and beats models twice its size on Microsoft's own table. Why model size stopped being the first question.

The OWASP list for LLM applications, read from a finance desk: prompt injection, training-data leakage, and a supply chain of someone else's code.

GPT-4-class capability fell from 20 dollars per million tokens to about 40 cents, and invoices went up anyway. Inference is the part that runs daily.