Redditor hand-draws how an LLM works, from token embeddings to prediction

Helpful-Series132 · reddit · 2026-09-13

A Reddit user shared a hand-drawn, simplified diagram of a language model's neural network, aiming to explain the fundamentals without cognitive metaphors.

Per the post: input tokens are converted into embeddings, compared to one another, and concatenated into context vectors forming a hidden state vector; the token with the highest cosine similarity to that vector is most likely to be predicted.

The author also notes that top-k adjusts the number of probable tokens, while temperature controls how sharp the token probability distribution is. (Note: this is a heavily simplified analogy that deviates from real Transformer attention/softmax mechanics — useful as intuition, not as an accurate description.)

Original post →

More from Research

Research channel →