LinkedIn publication text
Opening the Black Box of LLM Inference
How does a language model decide whether “banco” means a bank branch or a bench in “O banco estava lotado”? With no explicit rules. I built an infographic (50 slides, EN/PT) that walks through the Transformer, from text to output token, with concept, mechanism, numerical example and references at every step:
- 🔵 Tokenization: text becomes numbers (BPE, SentencePiece)
- 🟢 Embeddings + position: IDs become vectors with order
- 🟠 Self-attention: each token consults the others (Q, K, V)
- 🟣 Multi-head and masks: relations in parallel; a generative model cannot “see” the future
- 🔴 Add & Norm and FFN: stability and per-token transformation
- 🔵 Stacking and training: layer-by-layer refinement; weights store what was learned
- 🟢 Output: logits, softmax, temperature, top-k and top-p
It also covers a numerical case study, emergent properties (and the debate around them), limitations, a 73-term glossary and 64 references.
⚠️ Transparency: the numbers are didactic (tiny dimensions, hand-set or random weights, computed by code). Of the 64 references, 12 were checked at source and 14 links audited; the rest still need verification.
© Aridio Silva · @aridiosilva · October 2026. All rights reserved. Please credit the author when sharing.