add_post

Author: anmo-fish

config/_default/menus.yaml

@@ -8,7 +8,7 @@ main:
     url: /
     weight: 10
   - name: Papers
-    url: /#papers
+    url: /publications
     weight: 11
   - name: Blog
     url: /blog

content/_index.md

@@ -76,13 +76,13 @@ sections:
   - block: collection
     id: news
     content:
-      title: Recent News
+      title: Recent Posts
       subtitle: ''
       text: ''
       # Page type to display. E.g. post, talk, publication...
       page_type: blog
       # Choose how many pages you would like to display (0 = all pages)
-      count: 10
+      count: 5
       # Filter on criteria
       filters:
         author: ''
@@ -98,7 +98,7 @@ sections:
       order: desc
     design:
       # Choose a layout view
-      view: article-grid
+      view: card
       # Reduce spacing
       spacing:
         padding: [0, 0, 0, 0]

content/blog/20260224/featured.png

@@ -0,0 +1,117 @@
+---
+title: LLMs Will Fragment — Together with Humanity
+date: 2026-02-24
+authors:
+  - me
+
+image:
+  caption: ''
+  focal_point: 'Center'
+  preview_only: true
+
+tags:
+  - LLM
+  - AI
+  - Coarsening
+  - Impossibility
+  - Essay
+
+content_meta:
+  trending: true
+---
+
+When DeepSeek arrived, the Western AI world was shaken. I was not.
+
+To explain why requires a detour.
+
+---
+
+{{< toc mobile_only=true is_open=true >}}
+
+## On Coarsening
+
+We do not see the world as it is.
+
+The color spectrum is continuous, but we call things red, blue, green. Political positions form a continuum, but we split them into left and right. The actions of others carry endless context, yet we reduce them to good or evil.
+
+I call this operation *coarsening* — formally, a map ρ from an objective position space L to a label space Σ. Coarsening loses information. Two people at genuinely different positions get folded into the same label. Two people who are close get pulled apart by a boundary they never drew.
+
+The question I have been working on is this: can a given pattern of relations — who can connect with whom — be exactly reproduced using only the labels that coarsening produces?
+
+Sometimes it cannot. When the relational pattern is not constant within the fibers of ρ — within the sets of positions that share the same label — no label-level rule can match it without error. False positives and false negatives. Apparent cohesion where there is none; apparent fragmentation where things are actually close. This is not a design flaw. It is structural.
+
+---
+
+## There Seem to Be Two Kinds of Coarsening
+
+Attributes like coordinates or nationality, and evaluations like justice or evil, are both called coarsenings — but their nature is fundamentally different.
+
+In the first case, a shared underlying space exists. The problem is resolution, and refinement is possible — up to a point. That limit never fully disappears. Cognition has granularity. Observation has floors. But the space itself is common ground.
+
+In the second case, ρ itself differs between people. Person A and person B may use the same label, yet the structure of what each includes under that label is entirely different. This is no longer a question of resolution. They are measuring different things.
+
+These two kinds tend to arrive together. The fact that someone performed action X is a shared 0/1. The label that action X is *evil* is the output of a map — and each person runs their own. The act is observable. The evaluation is not.
+
+This raises a question that has stayed with me: when a single rule Rb is chosen, whose ρ is it built from?
+
+---
+
+## What LLMs Are Actually Choosing
+
+Large language models learn from vast human text, extract personas from it, and through post-training, refine those personas in a particular direction. Anthropic recently published a paper calling this the persona selection model.
+
+Translated into my framework: an LLM is selecting a single Rb. That Rb is optimized around a particular ρ. There is no escaping the question of whose feedback shapes it. The ρ of Western liberal democracy — and more specifically, the cultural ρ of Silicon Valley — is what currently anchors the label-level rules of the dominant models.
+
+I am not inclined to call this malicious. It is structural. Once you choose a single Rb, errors for someone are unavoidable. And that Rb being pulled toward some ρ rather than none is equally unavoidable. The reason "neutral AI" cannot exist is not ethical. It is mathematical.
+
+---
+
+## A Forecast
+
+From DeepSeek's emergence, I see a trajectory.
+
+China is building its own models and insisting that each nation should develop AI aligned to its own values. This is not primarily a technology competition. It is a disagreement about ρ. A model built to embody "socialist core values" and one built around "individual autonomy and free expression" will return different labels for the same observable act. My theory says nothing about which is correct. It says only that both are choosing a single Rb.
+
+This will accelerate, and not only in China. Governments, religious communities, political factions, subcultures — each will move toward models optimized around their own ρ. And within each of those, fragmentation will continue. The question of what counts as conservative, or feminist, or orthodox does not resolve once a label is shared. The fiber structure — what the label actually contains — remains perpetually contested.
+
+When a community moves toward the lower approximation — refining conditions, keeping only what its own ρ can assert with certainty — there is a kind of integrity to it. But the more precisely the conditions are stated, the more plainly it looks like a lie from any position that does not share them. To say anything is to divide the world. Move toward the upper approximation — try to hold every position — and the result says nothing at all. There is no exit in either direction.
+
+---
+
+## On Ambiguity
+
+What I keep thinking about is the region B.
+
+B sits between the lower and upper approximations — the zone where no label-level rule can render a definitive answer. The standard treatment is to regard this as a problem. Removing ambiguity is the goal.
+
+But B cannot be emptied. The structure will not allow it.
+
+Seen from the other side: B is the territory that remains undecided for everyone. Neither A's ρ nor B's ρ can force an answer here. That makes it the only genuinely common ground — the place where no one has yet staked a claim, the place where no ρ is fully confirmed.
+
+The history of legislation shows something about this. Legal boundaries always arrive late. An event happens, harm becomes visible, and only then does a line get drawn through part of B. This delay is widely criticized as a failure of the system. To me it looks like evidence that B is structurally persistent. Every time a boundary is drawn, a new B opens somewhere else. The errors do not disappear. They relocate.
+
+As long as ambiguity is treated as a defect to be eliminated, fragmentation will not stop. Every faction will spend its energy trying to collapse B. The time and attention of humanity will flow there.
+
+If B could be held as a place of negotiation rather than a problem to be solved — if the agreement that *this is not yet decided* carried as much weight as the agreement that *this is so* — something different might be possible.
+
+I do not believe this is easy. For finite beings whose survival depends on maintaining their own ρ, holding ambiguity open carries real cost. But whether it is feasible and whether it is the only non-trivial response are separate questions.
+
+---
+
+## This Text Also Divides the World
+
+Writing this, I notice something.
+
+The act of writing it is already a coarsening. The moment I frame the proposition that LLMs will fragment, I have taken a position. Some will agree. Others will not. That is not a failure of the argument. It is the structural consequence of saying anything at all.
+
+To speak is to divide. To take a position is to become an error for someone who does not share it. To stay silent is to say nothing. This is not a property of theories about language. It is a property of every act of expression.
+
+So what I am trying to say here is not that a theory is correct.
+
+It is a question: what does it mean to speak, knowing that speaking always draws a line? Is it possible to hold ambiguity — and still not be silent?
+
+I do not have an answer. Only this: no attempt to dissolve ambiguity has ever succeeded. And I do not expect the next one will either.
+
+---
+
+*Related papers: "The Impossibility of Cohesion Without Fragmentation" (2026), "Space Reduction and Representability of Relational Feasibility" (2026), Daisuke Hirota, arXiv / Zenodo*

content/blog/20260224/index.md

@@ -0,0 +1,143 @@
+---
+title: Why Communities Fragment — A Structural Account of Social Transition
+date: 2026-02-25
+authors:
+  - me
+image:
+  caption: ''
+  focal_point: 'Center'
+  preview_only: true
+tags:
+  - Social Structure
+  - Fragmentation
+  - Cohesion
+  - Community
+  - Essay
+
+content_meta:
+  trending: true
+---
+
+Scroll through a timeline long enough and a certain feeling sets in. People who once occupied the same space now seem to live in entirely different worlds. Nobody was expelled. There was no decisive falling-out. And yet the distance keeps widening.
+
+When we try to explain this, the usual suspects appear: radicalization, algorithms, political polarization, the quirks of human psychology. These are not wrong, exactly. But I have a prior question — can this transition be explained from structure alone, without assuming incentives or psychology?
+
+I think it can.
+
+---
+
+{{< toc mobile_only=true is_open=true >}}
+
+## It Begins with Loose Minimum Conditions
+
+Imagine a new platform. At the start, nobody knows anyone.
+
+The threshold for a relationship to hold is almost nothing. Being on the same platform, looking at the same timeline — that is enough. Agreement is not required. Shared values are not required. You can coexist with people you find uncomfortable. The community is loosely, widely, lightly connected.
+
+In my framework, I call these *minimum conditions* — the bundle of requirements whose absence would cause someone to withdraw from a relation. If someone cannot tolerate jokes about a certain topic, that intolerance is a minimum condition. The relation holds only as long as those conditions are met. In the early stages of a platform, these conditions are coarse. A low threshold for coexistence means that people whose actual positions differ widely are still counted as being in the same place.
+
+Formally, this is a coarse label map ρ over a position space L. When the granularity of labels is low, people at genuinely different positions get bundled into the same fiber — the same label. Apparent cohesion forms. Most of it is not real.
+
+---
+
+## Every Action Becomes a Disclosure
+
+Then conversation begins.
+
+What you say, what you react to, what you accept and what you refuse — these things involuntarily expose the outline of your minimum conditions. *I can go this far. Beyond that, I cannot.* On a platform, action is more visible than silence. So minimum conditions cannot stay hidden.
+
+This is position-fixing. Any act that signals which side you are on locks your position into the relational network. Once fixed, that position structurally determines with whom a relationship can hold.
+
+What matters here is the absence of intent. Nobody decided to fragment. Each person is simply acting in accordance with their own minimum conditions, and positions become visible as a consequence.
+
+---
+
+## Fragmentation Begins Not with Exclusion but with Exposed Incompatibility
+
+When minimum conditions become visible, what happens?
+
+Between compatible people, nothing in particular. Between incompatible ones, things simply fail to mesh. Nobody expels anyone. Nobody deliberately raises the bar. Edges that do not fit naturally fall away.
+
+This is the first fragmentation — and it appears not as attack but as distance.
+
+In the language of my theory, this is the structural consequence of a *bifurcation event*. Once positions are fixed, the survivability of each relation is structurally determined. Relations between compatible positions are maintained; relations between incompatible ones collapse. Nobody intended this. It is what the structure outputs.
+
+---
+
+## What Remains After the Peeling Becomes Stronger
+
+After the separation, those who remain share more similar minimum conditions and can talk with less friction. Formally: once bifurcation has filtered the network, internal compatibility density rises. Cohesion strengthens — structurally.
+
+This structural outcome then drives psychological and economic effects downstream. A sense of safety, solidarity, of being understood — these emerge as byproducts. And that reward promotes further action and disclosure. Structure comes first; incentives follow from it. The order matters. Cohesion strengthens structurally, and the reward arises from that — not the other way around.
+
+And that stronger cohesion promotes further disclosure of minimum conditions. People want to stay in a place that works, so they get clearer about what they will accept and what they will not.
+
+---
+
+## Refinement — When Labels Split
+
+Here a key mechanism operates. Minimum conditions do not get *raised*. They get *refined*.
+
+Among people who once gathered under the label "conservative," the difference between economic conservatism and cultural conservatism begins to surface. What had been bundled inside a coarse fiber turns out to contain incompatible elements. Interaction makes this visible. The fiber splits. ρ is subdivided.
+
+This can be understood in terms of representational error. When coarse labels are in use, the rules operating at the label level cannot always reproduce the actual pattern of relations. Differences within a fiber — people who share a label but occupy genuinely different positions — generate apparent cohesion, or pseudo-cohesion. As this error accumulates and interaction increases, it becomes visible. The label splits.
+
+Each split label then draws more compatible relations toward it. Fragmentation sharpens; cohesion strengthens. The process repeats.
+
+---
+
+## Event Density Determines the Speed
+
+Here, something specific to platforms comes into play.
+
+When high platform throughput accelerates fragmentation, two distinct variables are at work. One is *frequency* — the raw number of posts, reactions, and shares. The other is the *depth of disclosure per event* — how much a single action reveals about one's position.
+
+The two are independent. High frequency with shallow actions produces slow disclosure. Low frequency with position-fixing actions can produce rapid separation. Most platforms are designed to maximize both simultaneously. Posts are numerous, and each post is amplified through tags, quotes, and the echo of reactions, making positions more legible.
+
+The pace of fragmentation can be understood as the product of these two. Timeline velocity, visible quote-reply chains, recommendation algorithms that highlight positional signals — these are design features that push both frequency and depth upward at once. But the point worth holding onto is this: design accelerates the process, but the process happens without it. The problem with platform design is not that it favors certain content. It is that it compresses a transition that would otherwise unfold slowly and runs it fast.
+
+---
+
+## The Lightness of Coarse Labels, and the Weight of Their Consequences
+
+There is another mechanism. Coarse perception — crude labels like good/evil, ally/enemy, right/wrong — accelerates fragmentation on its own. But the reason is not simple.
+
+Attaching a coarse label is cognitively cheap. You can say "that is evil" in a second. But what that label bundles together does not match between the person who applies it and the person who receives it. The fiber of a label like "evil" — the set of positions it actually covers — differs from person to person. What A calls evil and what B calls evil are the same word pointing at different sets.
+
+This is fiber mismatch. When people using coarse labels talk to each other, they believe they share a vocabulary while actually passing each other by. That friction accumulates and eventually surfaces as disclosed minimum conditions.
+
+The problem is the asymmetry between the ease of the act and the weight of its consequences. A single coarse label can trigger multiple fiber mismatches simultaneously. The person who applies it cannot see what they have set in motion. The cost is local; the effect is broad. That asymmetry is what makes coarse perception not a form of tolerance but a catalyst for fragmentation.
+
+---
+
+## Is Structure Enough?
+
+Here is a question worth pausing on. The core of this account contains neither incentives nor psychological tendencies. Is that sufficient?
+
+Yes — with a division of roles.
+
+To explain *why* fragmentation is unavoidable, structure alone is enough. When positions are fixed and minimum conditions exist, cohesion and fragmentation emerge as the two necessary outputs of the same process. That is a structural fact, not a psychological one.
+
+To explain *what triggers* fragmentation and *at what speed* it proceeds, psychology and reward are needed. Describing the content of events — what produces fiber mismatch, what prompts the disclosure of minimum conditions — is where psychological tendencies are useful. Explaining frequency — why people perform so many actions — is where the structure of reward is useful. Psychology and reward are downstream of structure, but they are indispensable for describing the dynamics.
+
+This is what makes the account clean rather than thin. It does not eliminate explanatory variables. It assigns them to the right level. Structure determines *what* happens. Psychology and reward describe *when* and *how*.
+
+---
+
+## Treating Only Malicious Intent Is Not Enough
+
+Throughout this account, malicious intent did not appear. But it exists — which is where I need to pause before continuing. Malicious intent is difficult to define. The label "evil" carries different fiber contents for different people, each running their own recognition map ρ. What one person calls malicious, another may regard as legitimate.
+
+If a working definition is needed, I would put it this way: malicious intent is *using the structure to benefit oneself*. Deliberately accelerating the natural processes of fragmentation and cohesion, or manipulating them from outside for gain. That happens. It is one of the serious problems of contemporary society.
+
+But the difficulty is not only that intent is hard to detect. There is a more fundamental problem. Removing malicious intent does not solve what the structure produces on its own.
+
+When attention is directed at intent, the structure keeps running in its current shape behind it. Without any malicious intervention, minimum conditions still refine, fibers still split, separation still occurs, cohesion still tightens. Addressing malicious intent is necessary. It is not sufficient.
+
+When structure determines the shape of a problem, solutions need to operate at the structural level too. Describing precisely what is happening is the first step toward that. Understanding fragmentation not as someone's failure but as an unavoidable consequence of holding relations at all changes the direction of response — away from finding causes to treat, toward seeing the structure and questioning the design.
+
+To say something is to divide the world. To hold a relation is to draw a boundary. That is not a failure. It is what the structure outputs. The question is what we do with that.
+
+---
+
+*Related papers: "The Impossibility of Cohesion Without Fragmentation" (2026), "Space Reduction and Representability of Relational Feasibility" (2026), Daisuke Hirota, arXiv / Zenodo*