topic

LLMs

5 pieces tagged LLMs — notes and essays by Harshal Gajjar on this topic, newest first.

  1. A model trained on everything ever written knows everything ever written. That is a different thing from imagining what hasn't been.

  2. Open-weight models on owned hardware cost nothing per call. Claude Opus is billed per token and is better at almost everything. So the point where the cloud…

  3. The chat-vs-UI divide isn't about interfaces. It's about who holds the model of the task.

  4. The distinction isn't human vs. machine. It's what gets inspected.

  5. We built them on the info-transfer view of language. The places they feel hollow tell us what that view missed.