Sharepoint Became Outdated in One Tweet
The Companies That Never Wrote Anything Down
The companies that end up in this situation rarely planned it that way. They grew, acquired, merged — each new business unit bringing its own filing habits. Four companies under one umbrella might have four completely different approaches to where anything lives.
The Pattern
One keeps everything in a shared drive organized by year. One has a Teams site that nobody updates. One has a senior operations manager who knows everything and keeps it in his head and in a folder on his desktop called "stuff." The fourth has a SharePoint site, technically, but the last thing added to it was a PDF from 2021.
When someone new joins any of these companies, the onboarding experience is: talk to whoever's been there longest and hope they have time for you. When someone leaves, their knowledge walks out with them and you find out what you've lost six months later when you're trying to figure out why the northeast supplier contract has a weird carve-out and nobody can remember the conversation that produced it.
Why the Existing Tools Miss
Search tools find files but don't synthesize them. Confluence and Notion require someone to actually write the documentation — which, in companies like these, never happens consistently. RAG-based AI search retrieves fragments but can't tell you the whole story.
The Tweet That Reframed It
I came across Andrej Karpathy's tweet on a slow afternoon in April. He was describing a workflow he'd been using to build personal knowledge bases — indexing raw documents into a folder, then using an LLM to incrementally compile them into a structured Markdown wiki. No vector databases, no chunking, no embeddings. Just the LLM reading your files and writing a wiki that it then maintains.
The last line was the part I kept thinking about: "I think there is room here for an incredible new product instead of a hacky collection of scripts."
What he described is something different from anything I'd seen: a system that reads the documents that already exist and writes the wiki for you. Drop in the PDFs, the Word docs, the old emails you exported, the CAD drawings. The LLM reads all of it and produces a structured, cross-linked knowledge base that your whole team can then query in plain English.
A Proof of Concept
I built something called AutoAtlas to see if this actually worked. The compile quality is good. The query loop — where the LLM navigates its own wiki to answer a question — holds up better than I expected. It even runs locally on Ollama, which matters when the documents you're feeding it contain things your company would prefer not to send to a cloud API.
There's still work to do before this is something you could hand to a non-technical team and walk away. But the core thesis is sound, and Karpathy already told 20 million people he thought so too.