Sketch on Concepts of Structuring Knowledge - (Roam Internal Training Video)
What this covers
Recorded a 15 minute video to introduce folks on our team to some concepts that hopefully connect tasks their doing in training to the short and medium term vision, one of them told me "Feel like I get Roam now" and "You could just release this"
So... 🤷♂️ Thread on twitter here https://twitter.com/Conaw/status/1437009557956427785
Source description (no synthesized summary yet).
The speaker argues that knowledge is inherently a network structure, but our communication media flatten this into linear sequences, causing loss of information and inefficiency in knowledge transmission and collaborative problem-solving; Roam Research is designed to preserve structural knowledge and enable more efficient knowledge work.
- Human knowledge exists as interconnected networks with context-dependent relationships, but language forces linearization into sequential prose
- Linear representations create friction in knowledge transfer and collaborative work, requiring people to read irrelevant context to understand specialized parts
- Structured, templated extraction of knowledge into discrete entities (questions, answers, claims, definitions) enables parallelization and efficient attention allocation
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Historical thinkers like Leonardo da Vinci maintained a set of standing questions they were interested in, and when encountering new observations or mental models, they would apply these existing questions as lenses to see if the new tools yielded insights into their standing questions.
“you could think about the like finding historically like seven questions he was interested in any given time anytime he had a new observation or he had a new mental model or a new medical technique for a new idea for physics or something this one was learning he would um apply a lens over the new engine he would basically throw that new tool he had against his existing questions and see if they yielded any insights”
When knowledge is transmitted through linear media like video, text, or speech, the network structure of knowledge must be linearized into a sequential stream, which causes information loss and creates inefficiencies in knowledge transfer.
“but when people talk to talk in just like a stream of words to another person um and the way that we have representing knowledge often is like either like a long form video someone's reading that thing more um you'll have planes down here which are dependent now like the base assumptions”
The world is a messy, nebulous place of all sorts of things, and understanding it requires applying different conceptual lenses (different questions, mental models, techniques) that cause you to notice different patterns depending on perspective.
“so the first stage is just simple attraction the second stage is um we will give you various like templates of like so another another metaphor that i've written here is that the the world is like i don't know the world is like a messy nebulous place of all sorts of things and any time you're looking at the world um you are uh this is where that pastor looked very prepared for questions or you could think about the like finding historically like seven questions he was interested in any given time anytime he had a new observation or he had a new mental model or a new medical technique for a new idea for physics or something this one was learning he would um apply a lens over the new engine he would basically throw that new tool he had against his existing questions and see if they yielded any insights we think about this is like that there's various lenses you can have over the world that will cause you to notice different things depending on what perspective it is”
Human knowledge and concepts are organized in the mind as a network of interconnected, overlapping pieces rather than as linear sequences or hierarchies.
“when in your mind when you have like concepts or like any sort of knowledge you have in like a crazy structure but your beliefs are a network um of connective things and you know their concepts build up smaller pieces um there's like nebulous overlapping just like this but like these are the things that are inside human mind”
Einstein needed a mathematical language (non-Euclidean geometry) to describe gravity that didn't previously exist in common use; a friend helped him identify which specific mathematical concepts from an unfamiliar branch of mathematics were necessary, allowing him to learn minimally and create general relativity.
“that's useful particularly if like you know you want a mental model from some other foreign discipline and like that might be used for social problems there's a great example of like uh um albert einstein you know needed a mathematical language to describe like the idea of gravity and he didn't have it i had a friend who knew i've been in geometry which was a relatively branch of mathematics at the time and that friends would point to just the tiny set of matthew you need to learn you need to go 720 you're saying all of that you study this particular set of like descriptive equations and thereby create the you know like all like social relativity”
Mathematical writing historically underwent an innovation when proofs began to be written with explicit enumerated logical claims and instructions, allowing readers to track dependencies between claims and verify reasoning structure.
“and you know we've got like as an example for many many many many years right you know the way that mathematical knowledge was transferred right and um there's a big innovation in the writing of papers when they started writing proofs and sort of like number sets of claims you know and saying like you know like some sort of instruction on these things right um but even though proofs don't have sort of a ability to sort of like uh scan a logical structure and say like you know this claim is based on these two this one is a separate claim based on these two right and like these two combined plus this one over here will give you like this conclusion”
When two people appear to disagree in an argument, they may actually be using the same terms in different ways rather than disagreeing about objective factors, making definitional clarity essential to genuine understanding of disagreement.
“these two people are using the same term but in different ways and that's why they're agreeing they don't actually disagree about any objective factor they just read about definition of terms at least a word that's been using one of these like there are many other ways of passing stuff”
When a person takes notes only for themselves, they write single-word reminders or partial phrases because they already have the full context in their head; these notes are pointers into their own memory rather than self-contained information.
“when people the notes that they take um initially are basically just reminders for them but something that they've already they already have the sense of expected and also so their notes are gonna be like single word notes you know um or like you know um so how is this kind of wrong like um when people write just for themselves um you know there's this there's this uh they already have this whole context they built up as they learned something”
The astrolabe, Roam's logo, was one of the most ancient devices for understanding abstract relationships between humans and celestial bodies, enabling map creation and navigation as people explored new territory.
“you know that that is why our astrolave is our logo is because it is one of the most ancient devices used for like having a general abstract understanding of the relationship between humans and the stars so that you could create maps as you explore new territory and you could know generally where you are and relate yourself to mecca or to your home”
Philosophy departments have attempted to use argument mapping as a visualization technique to represent logical structure with pros and cons, but this approach is limited because it's a lossy data structure that loses important contextual information.
“and so far you know philosophy departments universities will try to do basic things like um argument mapping where they're sort of imagining things just having a structure of like pros and cons and uh you know the visualizer arguments again something um but that also is uh this this is one kind of compression so lll but it is limited in its application for paralyzing problem-solving and it is a poor description of reality so it is a um it is a velocity data structure it is one where you know velocity performance is basically where you lose information in the transformation process”
The James Baldwin quote 'the purpose of art is to identify the questions behind the answers' encodes the idea that beneath any definitive statement lies an implicit question, and understanding the questions is as important as understanding the answers.
“there's a quote i love from uh james baldwin it says the purpose of art is to uh identify right um if you deconstruct that you could say there's a question of that's why they're disagreeing they don't actually you think about how a question has right there's a quote i love from uh james baldwin says the purpose of art is to identify the questions behind the answer to different right”
One of the first aims in building Roam is to extract prose-based knowledge into data structures that allow more efficient allocation of attention for both knowledge transfer and parallelizing complex problems.
“so um the like one of the first things that we're trying to you know establish in rome is how do you take something that's written in prose and extract it out into a data structure that allows for more efficient allocation attention um both for transfer knowledge and for parallelizing complex problems require like intervention like you know interesting stuff um that's useful particularly if like you know you want a mental model from some other foreign discipline”
When someone expert in one narrow domain needs to provide feedback on a multi-dimensional problem, they must read far more context than necessary—reading sections irrelevant to their expertise—because the knowledge is linearized in a document format that forces sequential reading.
“the problem with this is that any person who wants to get upset on the thing even if this person you know basically they've got a particular expertise uh on like a multi-dimensional problem and you want them to give you feedback on like the one area where they actually teach the top level problem and that might be this thing right here right or like everything like let's say um um whatever that's better in this big document you know the thing that they are uh that their expertise is gonna be on is like that section but that section in order to like make any sense of this they have to read like a little bit of here and like a little bit right and so when you're transferring documents or with pros you know um it's possible to do it but often person has to read way more information all this other stuff you know um all this stuff here he's like irrelevant for them”
All web and mobile applications are fundamentally context-sensitive information graphics—different representations of underlying data that are dynamically generated based on user inputs and constraints.
“or i would say all of uh um uh one way to think about all the web applications or like phone applications right now because it is all just fancy screens but you every application facebook twitter all these things you use what they're called what they are actually really is context sensitive um information graphics um if you are filtering your your like you know shoe options from amazon you are basically creating a dynamic information graphic based on your shoe size and the price that you're interested in like the kind of shoe that you're interested in that kind of thing um so context and some information graphics like you know mapquest or google maps is a context-sensitive information graphic that is drawing you the map based on data that you've put into it right”
To even conceive of useful features for structured knowledge interaction, you must first think deeply about the structure of knowledge itself, so that you can then design experiences that allow people to interact with that knowledge effectively.
“but the ability to even think about what features are useful for this stuff requires thinking about the structure of your knowledge so that you can then design experiences that allow you to interact with these much you know these these things are all about or i would say all of uh um out making an authoring platform for these cloud making and authoring a higher dimensional base that allows them to like features are useful for this stuff requires thinking about the structure of your knowledge so that you can then design experiences that allow you to interact with these much”
We do not yet have a map of the skills required for knowledge structuring and mapping work, and we definitely do not have a map of the most effective exercises and training methods to develop and transfer those skills.
“so um so rome is like trying to help people create these maps and this is the skill we're going to be trying to instill in you and like you know and you're going to be part of as well everyone is we do not yet have a map of how like of all the kinds of skills here right and we definitely do not have a map of the most effective exercises that will hone those skills and transfer those skills and develop those skills and train people in the active map making and then the act of building apps for using maps right but that is um uh that's the project over um you know the next few weeks”
The second stage of Roam knowledge work is applying different templates or lenses to a document—for example, asking what statements within it constitute agreement, what questions the author is answering, how terms are being defined, what claims are being made as true.
“so the second stage of the room like processes we give you examples of these templates like i'm going to go through and i'm going to find uh you know the the lens that will apply is what statements in this are inside this agreement right or i'm going to go through this like i said this talk i'm gonna say what questions is this person trying to answer what terms are the first defining um like what statements of truth are they are like what are they claiming to be true right that that compression activity is saying someone gives someone an essay and you just say well what are they saying is true as fast right”
Roam is designed as an authoring platform for context-sensitive information graphics that represent conceptual space, similar to how Google Maps authors a platform for geographic information graphics that allow users to explore spatial relationships.
“so rome is about making an authoring platform for these contact sensitive information graphics that allows you to explore conceptual space um the way that mapquests or like you know the the way that um the way that google maps takes all the potential routes that exists in the world for um you know like going from point a to point b and it allows you to view just the information that's relevant to use in an actionable way so you can get driving directions um rome is helping you you know it's a different kind of tool because it is just as much for the creation of maps as you are exploring the space as it is for just helping somebody navigate um but uh you know that that is why our astrolave is our logo is because it is one of the most ancient devices used for like having a general abstract understanding of the relationship between humans and the stars so that you could create maps as you explore new territory and you could know generally where you are and relate yourself to mecca or to your home”
By restructuring knowledge as question-answer pairs rather than declarative statements, you enable easier comparison between different answers to the same question and can break down complex questions into sub-questions.
“if you start to structure it in a different way it's easier to start doing comparison and then you might even go out to this statement of messing with pope and that can be extracted into whatever characteristics of interpreter that would make them best right or like you know to like mechanical monk right um if you start to structure it in a different way it's easier to start doing comparison and then you might even go out to this statement of best in group and that can be extracted into whatever characteristics of interval that would make them best right depending on how you want to understand the question you may deconstruct these kinds of things into you know potentially a larger question like what makes the best and then based on traits that i might know and then i might say okay from here i'm going to ask what are the traits of jersey versus the woman you know and how the streets match up to my idea”
The first stage of Rome's onboarding process involves extracting interesting ideas from flat, serialized text (like transcripts) into discrete units through highlighting and segmentation.
“the first part of it is going to be like taking a bunch of to take information that is stored in a um like a flat serialized string data structure it's sorted basically as a continuous it starts at point and it adjusts those and there's no structure to it um one of the first things that we'll you know have you do is like a um uh it's like it's highlighting activity obviously he's pulling out interesting kids right um and this is starting to when you pull out those instruments into um new blocks right each one of these becomes discrete unit the first perspective is just extraction of three things”
Rome training involves a progression where initial stages teach template recognition (finding existing patterns), followed by middle stages where trainees apply templates to new material, and advanced stages where trainees create their own templates.
“the next one would be basically writing these things out um and then the two following stages from that which uh rome as a tool is particularly trying to do is to um in the on the more technical side like rome has properties that no other application has in terms of user extensibility that make it so that you can with a relatively small amount of technical knowledge create totally different experiences of interacting with that structured data”
Roam has technical properties around user extensibility that no other application possesses, allowing users with modest technical knowledge to create entirely different interaction experiences with structured data.
“which uh rome as a tool is particularly trying to do is to um in the on the more technical side like rome has properties that no other application has in terms of user extensibility that make it so that you can with a relatively small amount of technical knowledge create totally different experiences of interacting with that structured data of like you know allowing you to do things around like the simplest is using the tools we already have like querying um and uh you know like different view options that we have for like tables or or those kind of things”
By explicitly identifying the characteristics or properties that define what makes something 'best' in a category, you can extract this into a reusable framework that can be applied to compare different entities across multiple contexts.
“that can be extracted into whatever characteristics of interpreter that would make them best right or like you know to like mechanical monk right um if you start to structure it in a different way it's easier to start doing comparison and then you might even go out to this statement of messing with pope and that can be extracted into whatever characteristics of interval that would make them best right depending on how you want to understand the question”
The next stage of note-taking sophistication beyond sparse pointers is writing prose that lays out coherent reasoning where the author has mentally separated chunks of knowledge, carved out natural conceptual joints, and articulated the connections between pieces.
“so the next level that you see when people and this is something we obviously we're trying to find two writers is they're at least thinking about trying to like lay out reasoning in a sense of a way and they're writing prose that is like you know if you read the whole document like there's coherent arguments there's like you know they're they're separating out mentally the chunks and they're doing a good job of like carving values joints as possible and like laying out the conceptualized they're they're they're they're they're they're they're they're doing instances and they're interesting this is almost resource material they will get any blog posts or youtube talks or whatever”
When analyzing an essay to extract what the author claims to be true, you must often compress information across multiple paragraphs into new words that the author didn't explicitly use, representing the author's distributed claims in condensed form.
“that that compression activity is saying someone gives someone an essay and you just say well what are they saying is true as fast right they might not have you might have to write new words that they didn't write in order to say because they're saying it across the four five paragraphs but really you're compressing it into a different through different lens”