Video Resources
Curated videos matched to this repo’s modules. Watching one before working through the corresponding notebook gives you the intuition; the notebook then makes it concrete in code. All links verified.
DNN Refresher (notebooks/00_dnn_refresher/)
But what is a neural network? | Deep Learning Chapter 1 (3Blue1Brown) Neurons, layers, and weights as a function machine — the visual version of the forward pass. Watch before
00_neurons_and_mlps.ipynb. WatchGradient descent, how neural networks learn | Deep Learning Chapter 2 (3Blue1Brown) The loss landscape and downhill steps, animated. Watch before
02_training_loop.ipynb. WatchBackpropagation, intuitively | Deep Learning Chapter 3 (3Blue1Brown) What the backward pass computes and why, with no calculus notation. Watch before
01_backprop_micrograd.ipynb. WatchBackpropagation calculus | Deep Learning Chapter 4 (3Blue1Brown) The chain-rule bookkeeping behind chapter 3, including the cross-entropy derivative that
03_softmax_crossentropy.ipynbre-derives in code. WatchThe spelled-out intro to neural networks and backpropagation: building micrograd (Andrej Karpathy) Builds the exact engine of
01_backprop_micrograd.ipynbin 2.5 hours of live code —src/ai_playground/fundamentals/autograd.pyfollows its design. Watch
Transformer Internals (notebooks/01_transformer_internals/)
Transformers, the tech behind LLMs | Deep Learning Chapter 5 (3Blue1Brown) The best visual intuition for what a transformer actually computes: embeddings as directions in space, the residual stream, and next-token prediction. Watch before
00_transformer_overview.ipynb. WatchAttention in transformers, step-by-step | Deep Learning Chapter 6 (3Blue1Brown) Queries, keys, and values animated — why attention is a soft, differentiable lookup. Pairs directly with
01_attention_mechanisms.ipynbandsrc/ai_playground/models/attention.py. WatchHow might LLMs store facts | Deep Learning Chapter 7 (3Blue1Brown) What the FFN/MLP blocks are for — the part of the architecture attention videos usually skip. Relevant to the SwiGLU implementation in
layers.py. WatchLet’s build GPT: from scratch, in code, spelled out (Andrej Karpathy) Builds a GPT in ~2 hours of live coding, from bigram counts to multi-head attention. The single most effective transformer tutorial; our from-scratch model follows the same spirit with LLaMA-style components. Watch
Let’s build the GPT Tokenizer (Andrej Karpathy) BPE from scratch — merges, regex splitting, and the weird failure modes tokenization causes (spelling, arithmetic, trailing whitespace). Background for the
vocab_sizeconfig choice; see also PAPERS.md § Tokenization. WatchA Visual Guide to Mixture of Experts (MoE) in LLMs (Maarten Grootendorst) Dense vs sparse MoE, top-k routing, expert capacity, and load balancing — the concepts implemented in
02_mixture_of_experts.ipynb. Watch
Training Optimization (notebooks/02_training_optimization/)
Let’s reproduce GPT-2 (124M) (Andrej Karpathy) A ~4-hour end-to-end training run that hits every topic in module 02: mixed precision, gradient accumulation, fused AdamW, LR warmup + cosine decay, torch.compile, and DDP. The closest thing to this repo in video form. Watch
Neural Networks: Zero to Hero (Andrej Karpathy, playlist) The prerequisite series: backprop by hand, micrograd, makemore. If gradients feel shaky, start here before the optimizer notebooks (
04_adamw_from_scratch.ipynb). WatchDeep Dive into LLMs like ChatGPT (Andrej Karpathy) The full lifecycle — pretraining, SFT, RLHF — at a conceptual level. Good for placing this repo’s pretraining focus in the bigger picture. Watch
Inference Optimization (notebooks/04_inference_optimization/)
Fast LLM Serving with vLLM and PagedAttention (Anyscale) The vLLM authors present PagedAttention and continuous batching — the two ideas built from scratch in
00_mini_vllm.ipynb. See PAPERS.md for the SOSP 2023 paper. WatchThe KV Cache: Memory Usage in Transformers (Efficient NLP) The memory math that makes serving a cache-management problem — the right primer before the mini-vLLM chapter. Watch
Distributed Training (notebooks/03_distributed_training/)
Distributed Training with PyTorch: complete tutorial with cloud infrastructure and code (Umar Jamil) DDP from the ground up: gradient synchronization, all-reduce, torchrun, and a real multi-node cloud setup. Directly applicable to
src/ai_playground/training/distributed.py. WatchDistributed Data Parallel in PyTorch — official video series (PyTorch) Short official tutorials going from single-GPU to multi-node DDP, including a minGPT training example. Watch
GPU & NVIDIA Tools (notebooks/05_gpu_nvidia_tools/)
How CUDA Programming Works (Stephen Jones, GTC 2022) A CUDA architect explains the hardware truths behind the programming model: memory bandwidth as the real constraint, warps, occupancy. Pairs with
01_cuda_basics.ipynb, and is the mental model behind the SGEMM notebook’s roofline analysis. WatchFlash Attention derived and coded from first principles with Triton (Umar Jamil) Derives online softmax and tiling from scratch, then implements Flash Attention in Triton. Long but self-contained — no prior CUDA/Triton assumed. See PAPERS.md for the Flash Attention papers. Watch
GPU MODE lecture series (GPU MODE) Ongoing lecture series on CUDA, Triton, and GPU performance by practitioners. Lecture 50 recounts a CUDA → Triton → Flash Attention learning journey similar to this repo’s module 05. Channel