Blog

Every agent framework is a disagreement about how much to trust the model.

Your GPU is mostly idle during text generation. The entire inference stack exists to fix that.

Karpathy's nanoGPT, annotated line by line.

From raw internet text to a streaming response: pretraining, alignment, inference.

The residual connection solved deep learning in 2015. Ten years later, Moonshot AI noticed it also dilutes signal in very deep networks. The fix is 30 lines of PyTorch.

Building PyTorch from scratch in ~1300 lines - reverse-mode autograd, conv2d, optimizers. Trains MNIST. No C++, no CUDA, just closures and NumPy.

Five questions I actually ask ML candidates.

My blogs are too long so I built a RAG over them. Chunking, FAISS, reranking, streaming via Groq - and an eval pipeline because vibes aren't metrics.

The math behind deep learning, with code for every concept.

Ground-up NumPy and PyTorch - we build MNIST classifiers two ways to see how the pieces fit together.

Taking DeepSeekV3 apart: MLA, MoE routing, multi-token prediction. And how to read a 53-page paper.

Maybe intelligence is just a lot of compute applied to a lot of data having a lot of parameters. OpenAI wrote a paper on this.

The full YOLO story from v1 to v10 - grid system, loss function, NMS, mAP, and a PyTorch implementation.

Neural nets from scratch. NumPy only. Build a digit classifier.

Random Python tricks I've picked up.

Forking CPython, modifying the grammar, adding a 'nuke' statement, and other things I found digging through Python's C source.