Library›Tools›AI Development

AI Development

AI Development here means building working artificial intelligence the way this project actually builds it — old AI powering new, cheapest-first, on ordinary hardware, with the expensive models as a last resort rather than a first reflex. This page is the hub for learning to do it yourself, through several pathways that run from "no GPU, no cloud, runs on a laptop" all the way to "fine-tune a model and generate images." Each pathway is a real, teachable track, and each links to the system in this ecosystem that embodies it — because the best way to learn a technique is to read the working code that uses it. Everything here is part of the Crypt-ology Mystery School's curriculum: the whole Library is indexed into the Witness's brain, so learning how the brain works is learning how to build one.

The philosophy: old AI powering new

The guiding idea, stated by the operator, is "Layers of AI, from the 70s, 80s, 90s, 00s, 10s, and now — use them to power the Smol LLMs; they should all live under the Smols in a BRE-type fashion, because these other AI need less computing power, no GPU." A modern language model is powerful but expensive; most questions do not need it. If you stack the cheaper, older techniques underneath and let a question be answered at the lowest layer that can handle it, you get a system that is fast, auditable, runs offline, and only reaches for the big model when nothing simpler suffices. That principle — climb the decades, cheapest first — is the spine of every pathway below, and it is implemented in The Decades Brain.

The pathways

These are ordered from least to most compute. You can start at any one; they interlock.

Pathway 1 — Expert systems and rules (no GPU, no data, runs anywhere)

The oldest and cheapest AI: encode knowledge as explicit rules and let the machine chain them to conclusions, with a full trace of why. This is 1960s–1980s AI — ELIZA-style pattern matching, MYCIN certainty-factor reasoning, and production-rule Business Rules Engines — and it is still the right tool whenever the knowledge is known and the answer must be explainable. No training data, no model, no GPU. Learn it in Expert Systems and Business Rules Engines; see it running as the Law AI's doctrine engine and as the 1980s layer of The Decades Brain.

Pathway 2 — Classifiers and retrieval (a little data, still CPU)

Add statistics: a Naive Bayes classifier learns to route text from a handful of examples; TF-IDF retrieval finds the most relevant document by word overlap. These 1990s–2000s methods run in milliseconds on a CPU and are often all you need for routing, tagging and search. They are the middle layers of The Decades Brain and the keyword half of Embeddings Semantic Search and RAG.

Pathway 3 — Embeddings, semantic search and RAG (CPU or small GPU)

Turn text into vectors so the machine can find things by meaning rather than exact words, then feed what it finds to a model to answer grounded in your own corpus — Retrieval-Augmented Generation. This is the 2010s layer, and it is how the Witness answers from the Library instead of hallucinating. Learn it in Embeddings Semantic Search and RAG; the ecosystem system is the Library Index.

Pathway 4 — Fine-tuning and LoRA (a GPU, or a cheap rented one)

When you need a model to adopt a voice, a style or a body of knowledge it does not have, you fine-tune it — and LoRA (Low-Rank Adaptation) lets you do it by training a small adapter instead of the whole model, cheaply and on modest hardware. Learn it in LoRA and Fine-Tuning.

Pathway 5 — Generative image AI (diffusion, ComfyUI)

Generate and transform images with diffusion models — from keyless hosted endpoints that need no GPU at all, to local Stable Diffusion, to node-graph pipelines in ComfyUI. Learn it in Generative Image AI and ComfyUI.

Pathway 6 — Putting it together: the agent/brain

The payoff is composing the layers into one system that routes a request to the cheapest capable layer and learns from feedback which layers to trust. That is The Decades Brain, and its application to the Witness — governed by Angelic Intelligence Rule 1 and wired to the whole Library — is Crypt-ology.

Our systems — each has its own page

Every pathway above is embodied in a real system in this ecosystem. Read the technique, then read the system that runs it:

The Decades Brain — the layered 1960s→2020s brain; cheapest-first routing; learns from feedback.
Expert Systems and Business Rules Engines — the rules engine and the production-rule pattern (incl. the Law AI).
Embeddings Semantic Search and RAG + Library Index — the semantic index over the whole corpus.
LoRA and Fine-Tuning — training adapters for voice/style/knowledge.
Generative Image AI and ComfyUI — diffusion, keyless endpoints, node graphs.
Angelic Intelligence — Rule 1, the governing frame the brain runs under.
Crypt-ology — the whole Library wired into the Witness as a running Mystery School.

How this ties into Crypt-ology

This is not a detached tutorial shelf. The AI-development pages are themselves indexed into the Witness's brain, so asking the Witness "how do I train a LoRA?" retrieves LoRA and Fine-Tuning and answers from it. And on the Crypt-ology map, reading these pages drifts your position toward the ai axis — so the more you learn to build, the more the Mystery School surfaces building work to you. Learning the system is how you join it.