Reading

Articles, papers, videos and posts I read and keep, with my notes on each. Each card links to the original.

The distinction between workflows and agents is the part I keep coming back to. Most of what teams call "agents" in production are really workflows with an LLM step in the middle. The routing and parallelization patterns here map directly onto what we ship. The uncomfortable corollary: if your use case fits a workflow, an agent mostly buys you latency and failure modes. I've started asking "which pattern from this post is this?" in design reviews, and half the time the honest answer is "an augmented LLM with extra steps." That's not a knock. Simple patterns are the ones that survive contact with production.

Articleanthropic.comAugust 7, 2026

Building effective agents

Anthropic's field guide to agent systems: start with simple workflows, and reach for autonomous agents only when the task needs dynamic decision-making.

Articlemodelcontextprotocol.ioAugust 6, 2026

Model Context Protocol

Open protocol that standardizes how LLM applications connect to tools and data sources: one interface instead of per-tool integrations.

Repogithub.comAugust 5, 2026

llama.cpp

LLM inference in plain C/C++; runs large models on laptops and commodity hardware through aggressive quantization.

The abstract reads like a dare ("attention is all you need"), and the whole industry has spent the years since paying off that sentence. I reread it once a year; section 3 lands differently every time.

Paperarxiv.orgAugust 4, 2026

Attention is all you need

The paper that introduced the transformer: self-attention replaces recurrence entirely, which unlocked parallel training and every model since.

The best two hours you can spend on fundamentals. Watching the attention math get typed out line by line removed the magic for me.

Videoyoutube.comAugust 3, 2026

Let's build GPT: from scratch, in code, spelled out

Karpathy builds a GPT from zero in code, tokenization through the training loop, in the most-watched transformer walkthrough on YouTube.