Posts on Coding in Space

See how next-token completion becomes few-shot prompting, chat, tool use, and reasoning through interactive, token-level exhibits.

Open the KV cache, count what every remembered token costs, and explore how modern attention designs handle long context.

Connect arithmetic intensity, continuous batching, sparse experts, and speculative decoding to the latency each user feels.

Tab-completion LLMs (large language models) are models that provide real-time suggestions while a developer types code, often integrated into editors ranging from Vim to IntelliJ as plugins, but sometimes as an integral part of the editor.

I recently noticed a strange quirk of Phi-3.

Thirty times faster than Handlebars, half the features!

I recently gave a five minute “lightning talk” on what garbage collection is and how it works in modern languages.