Posts

Every agent framework is just a disagreement about how much to trust the model. Here's how to actually think about the landscape.

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

Karpathy's nanoGPT, annotated line by line. (I call my version microGPT.)

From raw internet text to a streaming response in your browser - pretraining, alignment, inference, and everything in between.

The residual connection solved deep learning in 2015. Ten years later, Moonshot AI noticed it's been sabotaging deep networks the whole time. 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'd rather ask than 'derive Kadane's algorithm in 30 minutes pretending you've never seen it.'

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.

Every derivative, gradient, and matrix operation you need to understand deep learning - with code alongside 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 I read a 53-page paper without losing my mind.

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.

I built an MCP server so Claude can control my Spotify. Play, pause, queue songs by mood - all through natural language.

Neural networks from scratch with actual math - then we build a digit classifier with no frameworks, just NumPy.

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

A curated list of things I like - books, music, tools, and more.