Andrej Karpathy, the researcher who coined 'vibe coding,' posted in April about building personal knowledge bases using LLMs. The method: drop source documents into a local folder, have an LLM extract and organize the contents into a Markdown wiki, update it as new material arrives. Within hours, GitHub repos and YouTube tutorials flooded the web. Casey Newton of Platformer spent the next several weeks building one, not knowing whether it would pay off.
It paid off. Newton calls the LLM wiki the single most useful productivity change he made this year, ahead of everything else he tested. He still runs Raycast, the Spotlight replacement that now supports Windows and lets him run GPT-5.5 Instant queries without opening a browser. He still uses Capacities for daily journaling and tagged link archives. He kept Recall specifically for its Chrome extension, which produces near-instant text summaries of YouTube videos. What he dropped: Notion's agentic search, which worked but required too much friction and couldn't look forward, only back.
The full piece is worth reading for the specifics Newton gives on why the LLM wiki beat every prior system he tried for research memory, including why Notion's agent failed to become a habit despite doing what he asked. He also details what comes next in his setup, the gaps the wiki still doesn't solve, and the maintenance cost that makes this tool closer to a commitment than a shortcut.
[READ ORIGINAL →]