What I'm Reading

An exploration of Josh Kaufman’s idea that focused practice can build useful competence in roughly 20 hours: define the skill, learn enough to self-correct, remove friction, and put in the time. Chen’s useful counterpoint is that figuring out the real goal and decomposing a skill are substantial parts of learning themselves.

An argument that the inference-engine moat is rapidly commoditizing as NVIDIA’s reference stack, open source, and coding agents shrink the time an optimization remains exclusive. Durable advantage is moving to reliable serving operations, GPU capacity, capital, and ultimately data centers and power.

A deep dive into the surprising complexities of how language models tokenize integers and why naive approaches fail.

A reflective piece on the evolution of LMCache from 2023 to 2026. Chronicles the project’s journey from early research into KV cache systems to an influential open-source project that gained significant industry recognition at NVIDIA’s GTC conference. Thoughtful reflection on timing, community building, and learning through execution.

Best intro I’ve found before jumping deep into the LLM inference optimization world. Explains compute, memory bandwidth, and overhead from first principles. Understanding which bottleneck you’re in matters more than blindly tweaking parameters.

Westenberg makes the argument I keep wanting to make but don’t: social platforms optimized for engagement have hollowed out the kind of slow, considered thinking that blogs made possible. The case for owning your words and building something that lasts.

The best kind of explainer. Starts from how transformers actually work, then reveals that prompt caching is just reusing the KV matrices from the attention mechanism. Makes you appreciate why the savings are so dramatic and why prefix structure matters.