Augmenting values in the process

To augment the value embedded in the process.

Erosion of value due to automation

When doing wiki gardening, would it be good if an LLM automatically inserted links to related articles? I think it has more disadvantages than advantages. For example, when writing an article in the wiki, the act of making an effort to recall existing related pages and manually adding links functions as a kind of active recall testing. If an LLM does this automatically, then in exchange for momentary convenience, I lose an opportunity to improve my memory.

Excessive automation can undermine the value embedded in the process.

To preserve the value embedded in the process, does that mean we should avoid automation? Even while automating, is it possible to preserve the value embedded in the process, or even increase it?

Augmentation of value through automation

At present (as of August 21, 2025), my quantitative goal for quality management of my wiki is to reduce the graph’s average shortest path length. The average shortest path length L of a graph G is the arithmetic mean of the shortest paths over all possible pairs of nodes. (n is the total number of nodes. For reference, the reason this metric can be used is that there are no isolated documents or remote islands in the wiki, so it is guaranteed that a path always exists between any two arbitrary nodes.)

L (G) = \frac{1}{n (n - 1)} \sum_{u, v \in V, u \neq v} d (u, v)

If I run a script I made for wiki management, it shows examples of the longest shortest paths as follows.

> uv run wiki.py avg-shortestpath --verbose

Approximating average shortest path length on the largest connected component (8042 nodes) using 500 samples...

Average shortest path length (estimated): 4.9780

Longer than average path 1 (length: 32):
  ## 1. 35th century BCE
  ↓
  ## 2. 34th century BCE
  ↓
  (중략)
  ↓
  ## 3. Lycurgus
  [[9th century BCE|기원전 9세기]] 경의 전설상의 인물. [[Sparta|스파르타]]의 법제를 만들었다고 함
  ↓
  ## 29. What is history
  Chapter 1. 역사가와 그의 사실
  ↓
  ## 30. Books Ive read
  내가 읽은 책들.
  ↓
  ## 31. The science of addiction
  [[Addiction|중독]]에 대한 [[Neuroscience|신경과학]]적 연구.
  ↓
  ## 32. A flexible hippocampal population code for experience relative to reward
  New findings suggest the same brain systems that guide animals to rewards may be weakened in [[Dementia|dementia]] and intensified in ...

Analyze the above report, then suggest three new or existing topics that could potentially shorten the longest paths by creating shortcuts between two nodes. For each topic, briefly explain your reasoning. The goal is to make the graph more dense and interconnected. Use your massive knowledge base from the training data to suggest interesting connections between nodes that do not seem related.

The last sentence (“Analyze …”) triggers the AI agent’s next action, and as a result I can obtain a hint like the following.

Suggestion: Self-Control and State Control

This topic explores the tension between individual self-control and the state’s role in enforcing it.

How it connects to Lycurgus: His entire legal system for Sparta was an extreme form of state-enforced discipline. It was designed to subordinate individual desires (for luxury, comfort, family) to the needs of the state, thereby engineering a society of ultimate self-control and austerity.

How it connects to Addiction Science: The science of addiction studies the biological failure of self-control. It examines how brain chemistry can override rational decision-making. This often leads to debates about the state’s role in managing addiction through laws, healthcare, and public policy.

The Bridge: The note would contrast the ancient method of shaping behavior through rigid social structure with the modern understanding of behavior as a function of neurochemistry.

By using an LLM as a reverse dictionary search algorithm to obtain hints, I can look for something among them worth studying, and then, by adding what I studied to the wiki, I can add a path that directly connects two documents that had been far apart from each other.

I think this kind of automation is beneficial in that it helps me discover what new things would be good to study in order to connect two concepts that I already know but that had been far apart from each other (if the distance is great in the wiki, then there is a high chance that the two concepts are not well connected in my head either), and in that it actually induces me to study and create the corresponding document.

Meanwhile, this approach can be seen as a kind of recursive neuro-symbolic inversion of control pattern, in the sense that the agent controls whether to call the tool, the called tool performs a type of task that the LLM is not good at, and then the tool instructs the agent to do again a type of task that the tool is not good at (reverse dictionary search), thereby triggering the agent’s work.