Compile Time AI Concept

rewrite this to make sense: Love your show, I have this idea I am calling Compile Time AI which is taking the use cases RAG, graphRAG and such that cost tokens per use at runtime and doing a large part of the compute first once and then provide a low latency intermediate representation such as a statically deployed next.js app. By doing the reasoning over the data first, you can solve most of the common questions and anticipate what kind of questions would be asked about the material. I do this through compiler passes of intermediate representations, first by extracting entities and such then extracting claims, resolving contradictions, deduplication, until you reach higher level JSON graphs you create a next.js app serving the API on Vercel in this example. I illustrate this all through the following post, GitHub repo, NotebookLM, and live demo in greater detail.

:::writing{variant="standard" id="58391"} Love your show. I have been developing an idea I call **Compile Time AI**: the concept of taking the capabilities of systems like RAG and GraphRAG — which typically perform expensive retrieval and reasoning at runtime — and moving much of that computation into a one-time compilation process. Instead of paying the token and latency costs every time a user asks a question, the system first analyzes a knowledge domain, performs reasoning over the data, and generates a low-latency intermediate representation. That representation could then become something like a statically deployed Next.js application that serves precomputed knowledge structures and APIs. The key idea is that many questions users ask about a domain are predictable. By reasoning over the source material ahead of time, you can extract entities, relationships, claims, concepts, and likely question paths before deployment. The process works like a compiler pipeline: each pass transforms the knowledge into a richer intermediate representation — from entity extraction, to claim extraction, contradiction resolution, deduplication, semantic clustering, and finally higher-level JSON knowledge graphs. The result is a compiled knowledge artifact rather than a raw document collection. In the example I built, these compiled representations are used to generate a Next.js application that exposes the resulting knowledge graph and APIs through a low-latency deployment on Vercel. I explain the architecture in more detail through the following blog post, GitHub repository, NotebookLM exploration, and live demo. :::