Too Much Knowledge. What Can We Do?

Publication date: 2026-01-08

Information abundance

The speed of new content creation

The pace of modern life is multiple times — or even orders of magnitude — faster than the pace of previous generations. News used to be passed by word of mouth; then came newspapers, radio, and TV; now it’s social networks.

In the past, the source of up-to-the-minute information was people around you (relatives, friends, random interlocutors, spies, ultimately); then it became specialists (journalists, reporters — and again, spies); now it’s everyone with a smartphone in their pocket.

People used to discuss their ideas in a narrow circle and decide which ones were worth spending paper on and sharing widely. Today anyone can post an article, a video, or even a tweet — and become publicly known if they get lucky (or unlucky).

Knowledge overload

Modern humans have to exist within staggering flows of information. From childhood we are told “knowledge is power”, but we aren’t taught how to distinguish useful knowledge from useless knowledge. As a result we try to become “even stronger”, to know as much as possible — and we can’t keep up.

“Now, here, you see, it takes all the running you can do, to keep in the same place. If you want to get somewhere else, you must run at least twice as fast!” (Alice in Wonderland)

We fear missing something important (FOMO) and overload ourselves. Plus everyone who wants to sell us something “useful” (as they see it) tries to get into our minds by any means. And then there’s advertising — of course.

Specialization is inevitable

Just 40 years ago, one person could assemble a computer’s hardware (ZX Spectrum) and write a BASIC program for it. Then came the split into hard & soft, then admins and developers, systems and applications, desktop and web, front and back, React and Angular.

This is how we humans try to adapt to explosive information growth and fit into its streams. Each of us chooses a direction we want — or have to — develop in (“run as fast as you can to stay where you are”).

What’s in our heads?

A personal worldview

Just as one LLM differs from another, every person differs from everyone else. But unlike an LLM, a person builds their own worldview (WW). From childhood, we interpret reality through the senses available to us (vision, hearing, smell, touch, etc.) and assemble a coherent (or not so coherent) but individual worldview.

And this worldview isn’t static — it changes over time. Something new appears, something old is forgotten. Age, injuries, diseases — all of this affects the size and integrity of a worldview.

Why other people’s thoughts don’t “transfer” as-is

We have no way to access another person’s thoughts directly. In that sense we are closed systems. Our personal experience is multimodal — a combination of signals from all sensory channels available to an individual. And these channels aren’t interchangeable: today you cannot transplant someone’s eyeball into another person and restore their vision.

The only way to learn someone else’s thoughts is if they communicate them. Knowledge is passed from one person to another via codes (language, writing, gestures) or by demonstration (“show me with your finger!”). In doing so, the “sender’s” knowledge is integrated into the “receiver’s” existing worldview and can end up very different from the original. For instance, our head nod for “yes” means “no” for Bulgarians.

Integrating new experience

Any new knowledge (or a set of it) is just a fragment relative to our worldview. It can integrate only if there are “hooks” to latch onto: the new fragment must contain familiar details through which we can locate its place inside our worldview — which is galactically large compared to the fragment being integrated.

Local African residents can explain why it’s better not to go outside after sunset, but if we don’t understand their language, that information won’t help us. Maybe they were telling a story about a successful hunt a couple of days ago? Who knows. Their “hooks” have nothing to latch onto. In our worldview, our “latches” are shaped differently than their “hooks”.

If we drop down to gestures, then yes — mutual understanding becomes more likely. It’s a fairly universal language for our species. It’s sparse, but expressive. For example, a quick, sharp, and dense contact of a clenched fist with the other person’s nose signals disagreement in all cultures.

Knowledge transfer

Encoding and decoding knowledge

In most cases, knowledge transfer (worldview fragments) between people happens through language — written or spoken. This allows transferring knowledge across distance (radio) or time (messages in bottles and cave paintings).

But in most cases, knowledge gets distorted in transmission, especially when it must be encoded/decoded through text. Try describing a sunset sky in words. Now try describing a sunset sky to a blind person.

It seems to me only mathematicians have solved this problem well: transferring knowledge through text without distortion. But even they aren’t omnipotent. No one has derived the formula for love, for example.

Why exchange knowledge at all?

Originally — for survival (these mushrooms are safe, those ones you can eat only once), and later — to improve quality of life (never eat yellow snow). An individual’s direct experience is limited by their capacities — physical, psychological, material — and knowledge exchange is great because knowledge doesn’t get alienated when it’s shared.

This way we can go beyond our own limits and build our worldview from fragments of other people’s experience, giving ours in return and enriching each other. As the saying goes: it’s better to live among the rich and healthy than among the poor and sick.

Why we read the same text differently

It depends on our preparation level and each reader’s worldview. The author encodes a fragment of their worldview into a text and introduces noise and deviations relative to the original. Philosophers of the past called “words” “shadows of ideas”. Words cannot convey an author’s ideas in full.

How easy it is for a reader to understand and accept a text depends both on how many “hooks” the author left in it — hooks the reader can use to pull the fragment into their worldview — and on the configuration of the reader’s worldview.

If the reader’s worldview already contains the idea being described (the author’s fragment), then one or two not very obvious “hooks” are enough for integration. In that case even the author’s clumsiness may not be a barrier to understanding.

If the author “chewed it up and put it in your mouth” but the reader lacks the foundation to latch on, then no integration into the reader’s worldview will happen at all.

Attention is a consumable resource

Why reading became harder

University professors often advised using older editions of textbooks — they were considered “more understandable”. A good example is Leonty Magnitsky’s “Arithmetic”. In a book of under 700 pages, the 18th century covered arithmetic, basic algebra, geometry, trigonometry, navigation, and astronomy — everything required for a mathematically educated person of that time.

Today, the amount of knowledge corresponding to a “mathematically educated person” has increased manyfold and is spread across countless textbooks and courses. Reading didn’t become harder — there is simply more to read. That is what increases the load and creates the feeling of difficulty.

We pay for knowledge with attention

People can’t pick up a book, flip through it, and then recite it word for word. At least I’ve never met anyone who could. Maybe some people have photographic memory, but that’s not about understanding.

When we read fiction, its plots integrate into our worldview via universal “hooks” — emotions. There aren’t many of them, and they are all well known. A writer’s art is to use those emotions to integrate a plot into the reader’s worldview automatically.

But when you need to integrate the multiplication table into your worldview, you have to work. You need to understand what the author meant, understand why you need it, and then think how to make it “stay with you” — if you do need it.

Long mental work, like physical work, exhausts a person. If physically we get weaker, then mentally we get duller. Working intellectually without rest, we become worse at seeing “hooks” in new fragments that must be integrated into our worldview. At the limit, a text turns into “white noise” — we read words but see no meaning behind them at all. To regain the ability to “see hooks”, we need to pause intellectual activity and do something mindless. That’s called a hobby. Sometimes — alcoholism.

Why measure attention spending?

For roughly the same reason you measure fuel consumption in a car. Without fuel the car doesn’t go; without attention a person stops functioning adequately in the world — they might try to cross the tracks in front of a speeding train.

We have no way to measure attention consumption precisely, but the fact that intense mental work tires us faster than light work suggests that attention (the capacity to focus and think effectively) is a consumable resource — and it can be spent at different rates.

If we want to be more effective in mental work, we must take this into account — even if we don’t yet know how to do it precisely.

Structured knowledge transfer

A book is not just text

A book is a classic way to transfer knowledge. Using it as an example, we can look at established structuring techniques:

Different entry levels

This “knowledge transfer” structure evolved to minimize attention spending.

I ordered these levels by increasing attention cost required to get acquainted with the transmitted knowledge.

Do you have to read everything?

No — absolutely not. Each of us builds our worldview from fragments we consider necessary. That’s an objective reality caused by information abundance, the need for specialization, and attention spending required to build a worldview.

Habr readers, for example, are all very different. No matter how good or bad an article I write, there will almost always be someone who downvotes it and someone who bookmarks it after finding something useful.

Everyone chooses what is important and needed for them. The author’s task is to enable as many different people as possible to synchronize with the described fragment of the author’s worldview (plus manipulation and entertainment — but that’s more common on platforms other than Habr).

Personalization through tracking attention

Classic personalization through declared interests

On Habr (and in many other places), to reduce cognitive load on the reader, users are asked to specify a list of interests in their profile, and a personalized feed is built from those interests.

Readers are also given the ability to rate publications (+/-) and comment. This is also a manifestation of interest and can serve as a marker to evaluate the quality and relevance of the published “worldview fragment”.

Modern personalization through behavior

Authors know what Habr statistics look like; for everyone else — here it is:

The latest article turned out better than the “hospital average”.

Reader behavior on the page makes it possible to estimate their personal “attention investments” into this publication. Aggregated across readers, it gives an author some feedback about how (un)successful their next piece was.

Behavior analysis also lets algorithms estimate a user’s real, current interests, which may differ significantly from the interests declared in their profile. It’s exactly this behavior analysis that personalizes the feed in TikTok, for example.

Publications and readers as intersecting trajectories

Just as readers want to find information relevant to them, authors want to find their relevant readers. Many readers now downvote articles without reading them just because “it’s neural slop”. Yet compressing meaning with LLMs can be quite effective — and I hope it is so for this text.

So if, in order to reach the right audience, I need to structure my publication as (title / blurb / table of contents / preface / main text) so that the platform’s algorithms can match me with the “right” readers — then I will do it.

Matching interests of readers and authors is a win-win for everyone: readers, authors, and the platform itself.

How this leads to “co-knowledge”

Why a shared canon no longer exists — and won’t

In the familiar education model there is a shared canon: a corpus of knowledge that “every educated person should have”. Well, there used to be, at least. Today it’s more like “every proper physicist”, “every proper biologist”, or “every proper frontend developer”. (I’m exaggerating about frontend developers, of course — there are no norms there at all.)

Our desire to “run as fast as we can to stay where we are” forces us to invest our finite (but replenishable) attention into knowledge that seems most promising to us (specialization). And in those narrow but promising areas (like AGI), we care about the opinions of people who are at least somewhat knowledgeable — not everyone.

Yes, there are still venerable Nature and Science, but there is also the more democratic ArXiv, and many ideas — like vibe coding — are launched through social networks altogether. Now it’s important not only to “ride the trend wave”, but to keep it going as long as possible — and ideally create trends yourself.

Basic knowledge should exist for every educated person, but today the “superstructure” (specialization) pulls harder than the “base”.

Partial overlaps instead of full agreement

When there is no canon and you have to create trends yourself, the ability for individuals to cooperate — intellectually as well — becomes important. You need to find co-thinkers and “dig the gold vein” together (or what seems like a vein to you). After all, exchanging ideas makes the whole community richer (if we ignore attention costs of learning them).

In such conditions, it becomes valuable for an individual to separate ideas/knowledge from their carriers. Two different people may share some ideas and develop them together, while other ideas might make them cut off any relationship to avoid potential recoding.

You can be Christian or Buddhist, support Spartak or Zenit, eat meat or be vegan — and still be interested, say, in spec-driven agentic development and dig into it together. You just need to be able to negotiate. Like the Rothschilds and the Rockefellers, for example.

Collective thinking as a side effect

At scale, analyzing such partial overlaps across fragments of personal worldviews yields development vectors of the community’s interests (specializations). We’re already moving in this direction: email newsletters, X/Twitter, Meta/Facebook, StackOverflow, Habr, reddit, Medium, WhatsApp, Telegram, … thousands of them.

We already group by interests on many platforms, share ideas, discuss them, and even rate them. Ideas live their own lives inside us. As a colleague @Kamil_GR said: “We are all just instruments of narratives fighting for the right to the substrate.” (c)

And narratives do fine in both human and artificial intelligence. The word “narrative”, for example, entered my everyday vocabulary via ChatGPT — I didn’t use it at all before.

Wrap-up: building castles in the air

Narrative as a “seed” of knowledge

A narrative is knowledge clothed in words (text or speech). It’s not even the knowledge itself, but the possibility of it emerging in another mind. Knowledge reproduces through narratives if they land in suitable conditions.

The same knowledge can be “packaged” in different ways — main text / preface / table of contents / blurb / title. When “unpacking” narratives of different density for the same knowledge, the probability of it “sprouting” in another mind is inversely proportional to the “packaging density”. It’s harder to extract knowledge from a blurb than from a preface, and if you can’t extract it from the main text — that’s a disaster. Either the packaging is bad, or the conditions are unsuitable.

Knowledge compression

It’s hard to lay out knowledge in full detail (for example, to write an essay like this). To compress a publication into a “preface” or a “blurb”, you can use an LLM.

In principle, any publication (narrative) on any platform (Habr, Medium, Nature, Science, ArXiv, …) can be compressed with an LLM down to a size comfortable for perception (a more compact narrative). A size that lets the reader get a sense of the original with a comfortable attention cost. That cost depends on the reader’s current interests and their worldview, which reacts to “hooks” in the narrative.

Embeddings as a fingerprint of the “seed”

Each packaging form of knowledge (narrative) can be converted into an embedding, and those embeddings will be very close to each other, though not identical due to differences in detail.

Embeddings of other, similar knowledge narratives will also be close. The denser the packaging, the more precisely embeddings indicate similarity/difference between narratives (and therefore knowledge). You can cluster narratives by embeddings across different packaging densities. In theory, these clusters should be very similar for each density — but, as they say, the devil is in the details.

Accounting for attention spending

Structuring exposition (narratives of different density) allows building a hierarchy of access to the most detailed narrative (the full publication) — from titles to blurbs, to prefaces, to main text. Moving from one level to the next increases attention cost and acts as a marker of the reader’s interest.

Using these markers, you can estimate reader interest in a publication similarly to time-on-page or scroll depth.

User interest profile

Attention markers of a user, in accordance with their worldview, can be obtained automatically while they read articles. Summing embeddings of the articles that interest a user across levels yields a vector of their current interests (separately for each level, of course).

And this vector works even better than a classifier or tags, because it requires no extra actions from the reader — it’s purely behavioral metrics.

One model for everyone

If you assume that the same model with the same settings is used to compute embeddings across different platforms, then you can compare embeddings to find publications similar in meaning.

Moreover, a user’s interest vector from one platform (say, Habr) can be used on another platform (ArXiv). And vice versa.

If these ideas sound interesting, you can discuss them with me privately. Thanks for reading.