Foundations · 3 min
LLMs without the incense
Tokens, attention, training, and the strangely convincing art of generating the next piece of language.
Read the storyTOPIC
One concept. Several ways in.
Tokens, attention, training, and the strangely convincing art of generating the next piece of language.
Read the storyRetrieval can supply evidence. The rest of the system still has to read it correctly.
Read the storyTool use, memory, permissions, and the unglamorous mechanics of dependable autonomy.
Read the storyA small, well-designed test set can save you from a very large, very eloquent mistake.
Read the storyGive the system a job, evidence, boundaries, and a definition of done. Adjectives can wait.
Read the storyEfficient systems start by matching capability to the task—and counting the whole cost.
Read the story