Knowledge Systems

08 Jun 2026

This weekend I was playing around with great teacher Karpathy’s LLM wiki concept and applying it to space based data centers. Seems crazy at first glance, but I wanted to see how real it was. The viability assement of space based data center is that you need fully reusable system like SpaceX Starship to bring down the launch cost and you need to deploy large surface area of radiators. There also isn’t that much more room left in LEO orbit, so you have a few GW limit.

I wanted to write down some thoughts I had about knowledge system and sources of knowledge.

Documentation at Google

Google has g3doc, where documentation is checked in as markdown files and it gets deployed to a website, similar to how github puts the README.md on the web.

I like that you have a git history of edits. Also the edits need to be reviewed before they are checked in. There can also be an owner and refresh requirements on the doc. There is a dashboard and there are metrics on how many visits a page gets. You know which pages are important to modify. You can quantify the impact of your documentation and get credit for it.

The problem with documentation is that there is little incentive for engineers to create it. Maybe that point is changed now since you need to provide context for the LLMs and you need a summary for yourself to figure out what the LLM built.

The documentation website can be separate from the rest of the production servers to isolate in case of cascading failures.

Confluence

Corporate wiki, page rot.

Google Drive

Searchable, but there is rarely any organization structure on the contents. It acts like a dumping ground that is often duplicated, incomplete, and fragmented. Great for permission / access control. Sometimes you accidently happen on things that should have been shared widely. Documents with a lot of visitors become view only. The mechanism for comments is good since it targeted on the text.

Substack / Newsletter / Blog

Push-based, one-way expert generated content behind paywall.

Podcast

Maybe like an audio blog, but usually focus on guests, since creator doesn’t have own content.

NotebookLM

Helps you learn a topic. Seems like a better fit for some people than LLM Wiki.

Wikipedia

The original. Relies on volunteers and highly motivated individuals. The question in the back of the mind is if this topic is important enough to warrant a pge in wikipdia. Editing is likley more tricky if multiple edits at once. Active discussions and lockdown for in progress events.

Just looks old now, but great. Needs donations to keep it going.

Groups / Mailing List

A lot of knowledge is shared in list, threads can go off topic. Much filtering is needed, but sometimes you can only find the answer you need in a reply.

Stack Overflow

Voting mechanism and exact solutions to questions. Sort of dead now due to LLMs. The LLM can fix the problem without you having to search Stack Overflow and adapting the answer to your usecase.

Reddit

Has a voting mechanism to ensure quality. Moderators keep the garden free of weeds. Often 1st hand experience you wouldn’t get otherwise.

Hacknews

Comments from people you would have difficulty accessing otherwise. Insightful comments. Also a voting / reputation mechanism.

ArXiv / Academic Journals

Free access to preprints. Althought it doesn’t seem like people send papers for peer review anymore. Heck, your peer reviewers are probably just going to through your pdf into the LLM anyway.

One way knowledge transfer, but should be credible.

X / Twitter

Good for realtime events like earthquakes for verification. Content is kept short, so it does not get into keep topics, but it does capture opinions and events over time.

The time / historical event based aspect of twitter is interesting.

Slack / Chat / IRC

Too much noise. Often references of things, but decisions can be captured here. Not really knowledge content.

Discourse / Forums

Similar to reddit, but more topic focused, gets highly technicals.

Notion

Easy for people to generate light content to assist with work. Scalability concerns and maybe not the best way to structure data. Takes opinion of being tree / hierarchical. Not sure if that is best approach. Likely works if you are opininated enough.

Obsidian / Notes / Google Keep

Obsidian is graph based instead of hierarchical like Notion.

Onotologies

Formal specification of things for knowledge. Too complex for people to use. People don’t have ideas of what relationships and attributes things should have.

Perplexity / AI Search

You get an answer, depending on the answer, you more or less believe it, maybe. One the level of other people’s bullshit.

LLM Wiki

Everything is LLM powered. Need self repair mechanisms, but normal docs are usually broken anyway. Unknown if LLM is worse, at least you design the system around that. It doesn’t go as deep into topics as I would like. These are things that I could try to fix in next project / iteration. Likely decay over time, but unknown if worst that any other system.

GitHub Issues / Discussions / PR

You learn about workarounds that haven’t been checked in to the code base. You saw that someone had the same issue, but they either gave up or nobody cared.

Youtube / Video

There’s a video for everything it’s useful. Also a lot of AI generated garbage, but if you know exactly what you are search for, great. You can learn anything on YouTube.

Conferences / Talks

You hear things from people that they wouldn’t put on paper or say to the general public.

Books / Library

The gold standard. Physical effort required. Time delay.

Comparison (by Claude)