The single biggest source of LLM-discourse confusion
Every LLM exists at the intersection of four layers — backbone, training, inference, product. Marketing language deliberately collapses them. "AI-native architecture" sounds like it means something architectural; almost always it means a thin product layer. "Reasoning engine" sounds like a new wiring; almost always it means RL post-training plus extended-thinking inference. Until you separate the layers in your head, every announcement reads like noise.
The four layers, explicitly
| Layer | What it is | What changes here actually means |
|---|---|---|
| Backbone | The wiring of the network — attention, FFN, embeddings, position encoding | A genuine architectural change. Rare and usually clearly disclosed. |
| Training | Pretraining + post-training (SFT, RLHF, GRPO, DPO, distillation) | Where most "breakthroughs" of 2024–2026 actually live. |
| Inference | How the model is invoked — standard, extended thinking, speculative decoding, beam search | Same weights, different runtime. Cheap to switch. |
| Product | Scaffolding — RAG, tools, agents, system prompts, guardrails, memory | Application code. Backbone unchanged. |
The reading test
For any claim about a model, ask: "Which layer does this change live on?" If the answer is "backbone", you should be able to find it in the architecture section of the technical report — different attention type, different FFN structure, different position encoding. If you can't, the claim is on a higher layer.
Why this lesson exists
The next several lessons pick apart specific common claims — RAG, tool use, agents, long context, speculative decoding, GQA/MQA/MLA — and show which layer each lives on. None of them are backbone changes, even when the marketing copy insists they are. Once you see the pattern in concrete examples, you stop being fooled by the abstract version.