Most "AI tools" lists are a wall of 500 links sorted by… nothing. A beginner and a distributed-systems engineer open the same undifferentiated pile and both bounce off it.
So I built a different kind of map: 149 GenAI tools sorted by concept depth — not by how hard a tool is, but by how much you need to understand to use it well.
The five levels
- Level 0 — Consume ready-made GenAI. No code, no theory. (ChatGPT, Claude, Copilot.)
- Level 1 — Run & build, no/low code. GUI model runners, drag-and-drop apps. (Ollama, Flowise.)
- Level 2 — Code with libraries. Python against SDKs and frameworks. (OpenAI SDK, LangChain.)
- Level 3 — Build, fine-tune & ship. RAG, agents, vector DBs, eval, serving. (LangGraph, Pinecone, PEFT.)
- Level 4 — Research & scale. Distributed training, inference engines, GPU kernels. (vLLM, DeepSpeed, FlashAttention.)
The trick: ease of use ≠ concept depth. ChatGPT is trivial to use, so it's Level 0. FlashAttention is Level 4 — not hard to install, but you need to understand attention and GPU memory to know why it matters.
Two links per tool
On the map, every tool has exactly two links — 📖 one thing to read (a hand-picked, approachable guide) and 🐙 the code (repo or homepage). No taxonomy, no noise. Each entry looks like this:
Ollama — run open models locally · 📖 a beginner's tutorial · 🐙 the repo
Sorting by depth turns a directory into a path: look at your rung, see 10–25 tools instead of 500, pick one — and the reading and the code are right there. It's a searchable single-page app: no signup, no tracking, open source (CC0).
👉 Browse the map: https://maneesh-kumar-thakur.github.io/self-serve-learnings-4-all/
⭐ Repo: https://github.com/maneesh-kumar-thakur/self-serve-learnings-4-all
Feel free to share feedback or if you are looking for help/content on specific topics
Top comments (0)