ai-engineering-from-scratch vs llms-from-scratch
Two of the top projects in Learning and cookbooks, side by side: who leads on each score part and who each one suits, then setup, license, activity and what each review found.
ai-engineering-from-scratch
Hands-on curriculum for building AI systems from math to agents
llms-from-scratch
Code for the book on building a GPT-style LLM in PyTorch
Verdict
ai-engineering-from-scratch ranks higher, 75 to 74. Of the 4 score parts we measure for both, ai-engineering-from-scratch leads on 2, llms-from-scratch on 1 and 1 is even.
- Adoptionllms-from-scratch leads by 49 points, 88 to 39: a wide gap.
- FreshnessEven at 100.
- Maintenanceai-engineering-from-scratch leads by 37 points, 95 to 58: a wide gap.
- Agent-readyai-engineering-from-scratch leads by 45 points, 45 to 0: a wide gap.
Pick ai-engineering-from-scratch if you are
Developers learning to build LLM, RAG and agent systems from first principles
- Lessons ship runnable code; the first ones need only python3 and no extra dependencies
- Three staged projects with local graders (project_test.py --strict) and reference solutions
Pick llms-from-scratch if you are
Python developers learning how LLMs work by coding one in PyTorch
- Main chapter code runs on a conventional laptop and uses a GPU automatically if present
- Plain PyTorch with no external LLM libraries, so every component is readable
| What we compare | ai-engineering-from-scratch | llms-from-scratch |
|---|---|---|
| Score parts, out of 100 | ||
| Adoption | 39, known | 88, widely used |
| Freshness | 100, active | 100, active |
| Maintenance | 95, healthy | 58, fair |
| Easy to run | Not scored for this kind | Not scored for this kind |
| Agent-ready | 45, minimal | 0, none |
| Facts from GitHub and the README | ||
| Stars | 68.4k | 106.5k |
| License | MIT (permissive) | custom license (read the license) |
| Last commit | Oct 2026 | Oct 2026 |
| Last release | Oct 2026 | None published |
| Language | Python | Jupyter Notebook |
| Docker | No | No |
| GPU | Not stated | Optional |
| arm64 or Apple Silicon | Not stated | Not stated |
ai-engineering-from-scratch
A course of 528 lessons across 20 phases, covering math foundations, ML fundamentals, LLM engineering, agents, tools and protocols, and production infrastructure. Lessons are runnable code in Python, TypeScript, Rust and Julia, with three graded projects: retrieval evaluation, agent trace debugging and a tool-call firewall. It can be followed on the website, from a local clone, or through tutor skills installed into a coding agent.
Who it is for: Developers learning to build LLM, RAG and agent systems from first principles
Strengths
- Lessons ship runnable code; the first ones need only python3 and no extra dependencies
- Three staged projects with local graders (project_test.py --strict) and reference solutions
- Tutor skills install via npx skills and save progress to LEARNING.md
- Code in Python, TypeScript, Rust and Julia; README translated into 30+ languages
Weaknesses
- 528 lessons is a large scope; total study time is not stated in the visible README
- Project starters fail by design until you implement the stages
- Tutor skills need Node.js, npx and a skill-capable coding agent host
- GPU and RAM needs for later phases are not stated in the visible README
- Needs Python 3.10+, Node.js 22.18+ (TypeScript project), Rust (Tool Call Firewall project), npx (optional, tutor skills)
llms-from-scratch
Official code for the Manning book "Build a Large Language Model (From Scratch)". Notebooks and scripts implement tokenization, attention, a GPT model, pretraining, classification finetuning, instruction finetuning and LoRA in plain PyTorch, without external LLM libraries. Bonus folders add from-scratch Llama 3.2, Qwen3, Gemma 3, Olmo 3, KV cache, MoE and DPO.
Who it is for: Python developers learning how LLMs work by coding one in PyTorch
Strengths
- Main chapter code runs on a conventional laptop and uses a GPU automatically if present
- Plain PyTorch with no external LLM libraries, so every component is readable
- Bonus material covers KV cache, GQA, MoE, DPO and several model architectures
- Exercise solutions, a companion video course and CI tests on Linux, Windows and macOS
Weaknesses
- Tied to a print book; contributions extending main chapter code are not accepted
- Produces a small educational model, not a production-grade LLM
- License is reported as NOASSERTION; the README states no license terms
- Video course and quiz PDF are hosted by the publisher, not in the repo
- GPU optional
- Needs PyTorch, Jupyter, Ollama (optional, ch07 evaluation)
- Models: GPT-2, Llama 3.2, Qwen3, Gemma 3, Olmo 3
More in Learning and cookbooks
Building with these?
We help teams pick an open-source AI stack and run it in production. Client work never changes a rank.