Topics and anchors are related but distinct concepts. They’re both used to organize information around non-claim focal points, but anchors clarify the parameters of debates while topics are used in the knowledge management system to aid in both a) fulfilling user queries and b) the ranking of importance/relevance of information (e.g., claims, anchors, or other topics) so that useful suggestions can be shown to the user.
Topics are a superset of anchors
All anchors are topics. Not all topics are anchors.
Topics
Topics should, for now, be implemented as a hierarchical graph of tags which can be managed and searched by LLMs. The miro map around homelessness is clarifying of the idea of a hierarchical graph; you want neighbour topics to be immediately related but you also want to maintain a short path of relations from a more “central” topic node such as “homelessness” to each other node. This means that you want to break down the topic space into categories, and perhaps break the categories into categories, so that an LLM starting at a central node can quickly traverse anywhere it needs to during a search operation.
Relatedness edge weights
Besides providing a foundation for search, topics could/should also store information about how related their neighbours are in the form of edge weights. This would allow the system to have some “understanding” of what is the most relevant to a given user at a given time. However, this is somewhat speculative as it assumes that a) topics will have so many relations that you’ll actually need to rank them, and b) that you’d actually need to store relevance at all; that an LLM couldn’t just look at the available relations and decide ad hoc which of them is relevant.
References between topics and claims
Topics are linked to many claims. Claims may be linked to many topics.
Topics vs claims on subject overviews
Anchors
Anchors are the explicit objectives around which a map (or a view of a map) is based. They may be described as to “topics” or “debates” in other software, though “topic” also carries a different definition in these notes. Anchors may be described using the word context, though this would overload the term from its other usage in which it defines certain details pertaining to individual nodes. Anchors, in contrast, pertain to the motivation for engaging in a debate. They are probably best specified in as much detail as possible or even with specific goals in mind since a poorly-scoped or poorly-framed debate can be doomed from the outset.
Purpose
The point of working with anchors is to make discussions more productive and focused.
Modelling anchor drift
The tricky part about modelling anchors is that, in real discussion, they often shift — especially in certain types of conversation such as the ideation phase of a creative project. Furthermore, we are not used to explicitly shifting our anchors — it often happens without us realizing. It would be unsettling to ask users to acknowledge “okay, we’re talking about something else now.”
- It might be possible for an LLM to propose other anchors as you’re writing.
- Maybe you could see some suggestions for other popular anchors for the nodes you’re viewing as a reader.
- When anchors drift during discussion, it might be that they only drift so deep — you could add anchors to the “call stack” under the assumption that the stack will never get very tall and that you would/could pop your way back up it at some point.