You don't learn AI by watching YouTube courses.
You learn it by building the thing yourself.
Here are 10 GitHub repos that make you build AI, not watch someone else do it:
- nanochat by Karpathy
https://github.com/karpathy/nanochat
Train your own ChatGPT-style model end to end for ~$100. Tokenizer, pretraining, fine-tuning, web UI. 43K stars.
- LLMs-from-scratch by Sebastian Raschka
https://github.com/rasbt/LLMs-from-scratch
Code a ChatGPT-like LLM in PyTorch, one step at a time. The most complete from-scratch path that exists. 97K stars.
- nanoGPT — Karpathy
https://github.com/karpathy/nanoGPT
600 lines of clean code that reproduce GPT-2. Hackable enough to read in one sitting.
- LLM101n by Karpathy
https://github.com/karpathy/LLM101n
Build a Storyteller LLM from basics to a working web app in Python, C, and CUDA.
- llm.c by Karpathy
https://github.com/karpathy/llm.c
Train an LLM in raw C/CUDA. No PyTorch, no Python bloat. Learn what the framework hides from you.
- beyond-nanogpt by Tanishq Kumar
https://github.com/tanishqkumar/beyond-nanogpt
Almost 100 modern deep learning techniques implemented by hand. KV caching, speculative decoding, vision transformers, PPO, AlphaZero.
- reasoning-from-scratch by Raschka
https://github.com/rasbt/reasoning-from-scratch
Take a pretrained model and build reasoning into it from scratch. Inference-time scaling, RL, distillation.
- llm-council by Karpathy
https://github.com/karpathy/llm-council
A local app that runs your query across multiple LLMs, has them rank each other, then a chairman model decides. Build it in a Saturday.
- machine-learning-book by Raschka
https://github.com/rasbt/machine-learning-book
The full code for Machine Learning with PyTorch and Scikit-Learn. The foundation before you touch transformers.
- Prompt-Engineering-Guide by DAIR AI
https://github.com/dair-ai/Prompt-Engineering-Guide
Not a from-scratch build, but the reference you keep open while you build everything else.
Stop collecting tutorials.
Pick one repo. Clone it. Break it. Rebuild it.