I must admit that I fall fot the knowledge base scam: the idea that we can delegate to the AI the management of “knowledge”. But on the flip side, I think I am one of the first to notice that that is a scam. Let’s see why

what this NOT about

I am not saying LLM/agents are not useful, that we will never reach AGI (as Professory Barry Smith seems to say), that “good enough” is not good enough etc

And I am not saying that LLM/agents will solve everything either and remove all jobs and that is so powerful we need to be scared, like like Dario Amodei seems to say

the second brain idea

Since Evernote, or even org-mode!, we felt we need to manage knowledge (we just took notes before, no need to “manage” beside having a file cabinet). In principle it sounds right as you can’t remember anything and there has been an exponential (in quantiti, not necessarily in quality) amounte of things to know. the problem is when you think you can delegate things to the physical world. 1

You would never treat a series of physical notebooks and sheets of paper as “second brain”, and yet you think the digital version of that can be.

While an academic definition of knowledge is complex to give, you can have a pretty good actionable definition by saying:

knowledge = data + metadata + experience

Notes are cool if you are not good at retaining data, and writing is an important part for experience, but that’s it

A SeCoNd BrAiN ThAt WoRkS

pre LLMs, Second brains never worked because:

  1. people get tired of bad UX and abstract management (feeling the sheet of paper is less tiring, and if you don’t need it you just trash it)
  2. there really is no increase in knowledge by writing notes

a self fullfilling profecy

Enter the LLM based knowledge base.

What’s the knowledge base essentially: a directory with structured sub-directory + index and markdown files with notes. Notes are generated by the LLM for the LLM. One suggestion is to use Obsidian as the front end of the note to curate them.

The LLM/agent treats them as CRUD (create, read, update, delete) db of post-its as part of its operations.

Now you have LOTS of DATA and METADATA but you don’t have experience, and henge knowledge

QED

How will I proceed from there

Will just continue taking and tossing notes, printing papers and create proof of concepts. Will use the LLM/agent for the PoC, but only delegating those things I already know, and as better google. That’s it

  1. my idea is that we thought that, since it worked for the calculator, it should work for everything. Let me specify a little: the calculator works because arithmetic is compressible, ideas are not. so you can delegate arithmetic to a calculator, but you can not delegate “ideas” to a machine. Also, solving math problem works because you have defined rules (that you already know). Defining the rules is an entire different thing (unless there are rules for rules etc) ↩︎