Sohejl@sohejl_
490 open browser tabs. Most people just forget about them. I turned them into a searchable knowledge base in 20 minutes.
@itsolelehmann wrote about @karpathy‘s viral LLM knowledge base post today and ended it with "someone please build this."
Karpathy's idea was to take everything you're interested in (articles, papers, tweets, videos etc.) dump it into one folder, point your AI at it, and let it read, organize, and index everything. Then you can query your entire personal library with natural language. Ask it to connect ideas across sources, surface things you forgot you saved or summarize everything you got on a topic. And the AI maintains it all as you add new stuff.
I had 490 open Safari tabs. Basically an unorganized dump of links to read later. Which I never end up reading later anyway :)
So I sent Karpathy's post to my AI agent and after some back & forth, it actually built this in one short session
I pasted all 490 URLs. I asked Pulse (my AI agent) how to do this easily and it told me how to easily bulk copy links in Safari on your phone with 2 clicks.
Out of the 490, it filtered to 251 worth saving: 206 X posts, 21 articles, 15 tools, 6 GitHub repos, 3 research papers. The rest was skipped automatically. All categorized, frontmatter written, saved to the right folder. Done in under 20 minutes!
It pulled metadata on every X link. We ran into a few bugs which got fixed after some back and forth: most of the "tweets" weren't tweets. They were full X articles sitting inside empty post wrappers, invisible unless you know where to look in the API response. 158 of 206 posts turned out to be complete long-form articles that were missed. It went back and reextracted it all.
What the system does now (image attached):
I drop any URL into the chat. It gets filed, tagged, indexed. Zero effort on my end. The ongoing maintenance is basically free.
I query it naturally. "That David Deutsch article about raising children" -> finds it. "Connect new ideas and frequent tips regarding SEO across my knowledge base" -> synthesizes it across 30+ sources. "Something about subscription pricing in a newsletter I read a while back" -> semantic search gets it from a half-memory.
Now and then, it surfaces relevant material proactively without me asking. The attached image was also generated by the agent itself, showing how the full flow works. The rabbit hole honestly doesn't end. You can do so much now, simply with an idea + giving an AI agent the right tools + iterating on the system together. It's wild.
We‘ve never had such high leverage on human creativity ever. As usual, technology keeps shifting the arena from the perspiration phase to the inspiration phase. We don‘t have to do the arduous tasks. „1% inspiration and 99% perspiration“ is a misleading idea of how progress happens: the perspiration phase can be automated.
I think creativity & understanding will increasingly become differentiators. And it has never been easier to build both: by using AI.
(Sidenote: I already have a personal AI agent running 24/7 on a server, with access to obsidian notes and QMD for semantic search. So the plumbing for this use case was already in place. But this setup is too much to get into for this post…will get into it another time!)
