Spaces:

---

karlexmarin

/

taf-agent

like 0

Running

🧠 How Transformers Attend — a free, in-depth field guide to what actually happens inside an LLM

Most explainers stop at "attention is all you need." ✋ This one keeps going — all the way down to the mechanism. 🔩

How Transformers Attend is a 43-chapter field guide that takes you from tokens and embeddings to mechanistic interpretability — and does something I haven't seen anywhere else: it measures the internals instead of hand-waving them. 📊

✨ What makes it different

• 📉 The attention decay law A(d) ∝ d^−γ, with a γ-atlas measured across 42 real models — data, not vibes
• 🔬 A physicist's lens: phase structure, a thermodynamic dictionary, fractional transport, grokking dynamics
• 🧩 Foundations done right: residual stream, multi-head attention, RoPE, KV-cache, long-context, quantization, RAG, agents, multimodal
• 🩺 Interpretability & honesty: verified vs folklore vs numerology, evaluating without fooling yourself, and "reading the model's whiteboard" (the global workspace)
• ▶️ Runnable demos + a browser diagnostic environment (TAF Agent) — see the ideas on real models, don't just read about them
• 🔄 Continuously updated, backed by 3 published papers (Zenodo)

🆓 Free · 🌍 bilingual EN/ES · 🚫 no signup

👉 https://karlesmarin.github.io/transformers-guide/en/index.html

If you want to understand how these models work on the inside — not just use them — start here. 🚀