AI Compendium · Part V — Building AI in Practice

draft

From theory to the workshop. Parts I–IV explain what AI is; this part explains how a frontier AI is built today — the real levers, the data, the concrete hardware, the tools, the evaluation, and one exemplary domain (programming). It is the translation of conceptual knowledge into engineering decisions.

🎨 Figure F-V.0The AI workshop. Brief: isometric aerial view of a stylized "workshop": a data bench (a funnel filtering), a GPU rack, a tools/frameworks panel, a benchmarks gauge, and a conveyor connecting everything to a model at the center. Compendium palette.


The AI workshop

The guiding question: what level of ambition?

Before building, situate yourself. The effort and the cost change by orders of magnitude:

Level What it is Who reaches it
1 Using APIs + prompt engineering Today, anyone
2 Fine-tuning on a specific domain A small team
3 Training from scratch on a specific domain Startup (\(200k–\)10M)
4 General frontier Billion-dollar labs

Central thesis of this part: for almost everyone, the winning path is open models + fine-tuning + specific domain. Beating the frontier in your domain is worth infinitely more than being mediocre at everything. The billion-dollar labs handle the foundation; you master a niche.


Map of the part

Doc Theme
01-frontier-levers.kmd What actually moves the needle: data, post-training, RLVR, test-time compute, scale
02-data-and-datasets.kmd The biggest lever: data quality, datasets, synthetics, filtering
03-hardware.kmd GPUs, VRAM, builds with prices, Blackwell/Rubin, cloud vs on-prem
04-tools-and-frameworks.kmd The practical stack: training, inference, sandbox, eval, observability
05-benchmarks-and-leaderboards.kmd How to evaluate; benchmarks, leaderboards, and why private ones matter
06-ai-for-code.kmd Programming: the best domain for a small team; the 7 ingredients
07-kode-case-study.kmd Case study: applying all of this to build Kode

How this part talks to the rest of the Compendium

  • Specializes the lifecycle (Part I) with concrete decisions for each stage.
  • Grounds the types (Part IV) into "how to train/serve this type".
  • Anchors itself in the state of the art (Part III, doc 8) for the current numbers.

Timeless × dated: the principles here (data > architecture, post-training, private benchmarks) endure; the numbers (prices, VRAM, models) are from 2026 and should be reread alongside the snapshot in Part III.