#llm
4 posts, earliest first
-
Working With LLMs Effectively: Mental Models That Actually Match the Mechanism
Effective collaboration with a capable LLM is mutual deference based on honest capability assessment. Most prompt engineering is the human refusing to do the partition.
-
The Context Window Is the Workspace, Not a Buffer
The folk model treats the context window as storage you load before the model knows things. The mechanism is the opposite: the window is the workspace itself, and every token in it shapes every answer the model gives.
-
Attention Is Not Memory: Why Models Re-Read Everything Every Turn
The model doesn't remember earlier turns. It re-reads the full transcript from scratch every turn. What feels like memory is re-reading, and the failure modes follow directly from that.
-
Generation Is Not Retrieval: What 'Hallucination' Actually Is
Developers treat hallucination as a lookup that went wrong. The mechanism is that the model generates plausible continuations and there is no lookup. Verification isn't a bug to fix on the model side. It's a structural job on the user side.