michael-dean-k/

On data sovereignty: I think the age of AI will show us that we want to unlock our data from apps and own it ourselves. We want the ability to organize it and feed it into applications. Only when our data becomes tremendously valuable will we demand ownership.

Precision embedding: one of the reasons I assume AI apps are bad at embedding is because they’re working on massive datasets and have no granularity or algorithm on how to properly chunk data. For example, my logs won’t have useful embeddings unless they are chopped. If the embedding is scoped to span 3 logs in some, and 10 logs in another, then it’s creating chunks that span a wide array of things—random noise.

If you generally copy + paste a lot for AI writing or code, you should immediately stop and use Cursor as your main tool. Even though it’s a developer tool, it’s the best way I’ve seen to connect your personal data to an array of models. It’s far from an identity clone, but it’s remarkably easier to load context and personalize outputs. My notes so far:

1. Easier prompt storage: In ChatGPT, I either, a) lose my prompts, or b) store them in a separate doc. In Cursor, all my prompts live as named files in my directory.
2. All your data in one environment: In ChatGPT I’ll sometimes copy in 10k words of context and it looks ridiculous. In Cursor, you can have all your data organized in a file structure (principles, essays, notes, highlights, etc.), and simply reference the file in your prompt.
3. Rapidly load context: When you’re writing a prompt in Cursor, you can use key commands to load in all your context. Opt+Cmd+P > then type the prompt_name & context_name you want to attach. Not only do I not copy and paste anymore, I barely use a mouse.
4. Auto-scan for the right context: In some cases, you might not know the context you need, so Cursor lets you “Search codebase.” Anything in your workspace is given an “embedding” (LLM coordinates). If you search your codebase, it then turns your prompt into a coordinate too, and will pull semantically related things you forgot about. This means it can quickly scan 10 years of Evernote, and pull in the relevant bits as context.
5. Patch edits: If you set your files as markdown (.md), it works as a document editor. You can selectively highlight and run prompts just on your selections (ie: edit this one paragraph).
6. Redo bad prompts: The conversation chain matters, and in ChatGPT, if you put in a bad prompt, you’re sort of stuck with it. Cursor lets you click into a prompt, edit it, and re-process the result.
7. All the models in one place: OpenAI, Claude, Gemini, DeepSeek, Grok.