I'm writing "Understanding Go AI Inference": how the run of LLMs locally, from Go, work underneath. The stack is llama.cpp (the C/C++ inference engine), Yzma (which calls it from Go, without cgo), and Kronk (an OpenAI-API-feeling SDK and model server on top). The first article demystifies inference itself. A simple mental model for this is: inference is a pure function. It takes your text and billions of frozen numbers (the weights), and returns one prediction: the next token. Something like nextToken(input, weights). No learning, nothing updated. And a whole answer is just that function in a for loop: predict a token, append it, run again, until the model emits its "I'm done" token. Also, in this article we explore one fundamental part of the LLM, the GGUF model file, which is fully self-describing. The weights, the architecture, even the entire tokenizer live inside that one file. Article drops Tomorrow! Want these articles directly in your inbox? Subscribe to the newsletter: internals-for-interns.com/ne… #Go #AI