Author: Nebula Walker Date: 06AUG2026 MYTHOGEN ENGINE (mythogenengine.com)
One Like, One Point; One Flame War, Thirty Points
Late 2017. Facebook's internal message board.
An employee posed a question he himself labelled as "devil's advocate": would weighting emoji reactions at five times the value of a Like cause the News Feed to surface a higher proportion of controversial content?
He offered an example on the spot. Post "I bought a cup of coffee" and you'd probably get a few Likes. Post "Steve Bannon punched Hillary" and you'd get floods of angry reactions — earning five times the distribution.
A colleague replied: possibly, but the company is aware of this risk and is working to mitigate it.
Everything that happened in the years that followed was, in fact, already written in that single comment.
I. Only Creators Notice
Let me start with something it took me a long time to understand.
A person who uses Facebook daily for over a decade has no reason whatsoever to verify its distribution mechanism. They read news, check on friends, comment, Like, occasionally share. As long as the News Feed has something in it, the system looks normal to them.
They'd probably notice more ads, more recommendations, fewer posts from friends. These changes are all small — too small to merit questioning.
The people who actually start to suspect are always a different group: creators, community admins, ad buyers, social media managers. Because only this group watches the numbers — why did ten thousand views become one thousand? Why do the comments keep coming but reach is plummeting? Why doesn't search find my own article?
I didn't learn that Facebook was no longer the Facebook I'd known for over a decade until I started writing.
There's a structural problem embedded in that sentence: if a system can only be verified by a tiny minority with the motivation to do so, then society effectively doesn't know what that system has become. This isn't because people aren't smart. It's because most people simply have no reason to look at that layer.
The first two volumes addressed my own layer — why a post wasn't seen, why an article was judged as not good enough. This volume asks a bigger question: when the entire venue for public discourse changed owners, what exactly did we lose?
II. Four Major Forums: Three Folded, One Half-Dead
Let's start with what can be verified.
HKGolden (Hong Kong Golden Forum) began operations in 2000, initially as a spinoff of a computer information site. From 2006 onward, members' remixed creations and parodies began being cited by mainstream media including Apple Daily and Ming Pao. For a period, HKGolden was one of the most influential forums in all of Hong Kong.
Then came the decline. The Wikipedia entry lists reasons worth reading item by item: slow connection speeds, primitive mobile apps, floods of political shills, poor management quality, and administrators engaging in political censorship and frequently banning member accounts arbitrarily. The result: in 2016, LIHKG split off. HKGolden haemorrhaged users and page views plummeted. Before Alexa's service ended in 2022, HKGolden's ranking had dropped from the top twenty in Hong Kong at its peak to beyond 250th — that was the last snapshot this ranking service left behind. No comparable public data exists after that.
Note the last item on that list. Admin censorship, arbitrary banning — this is moderator-level power, occurring before any machine-learning ranking existed. The patent described in Volume I was filed in 2015; HKGolden's members had already left over the same issue before that.
The same mechanism, a more primitive version. This is also why I don't accept the framing of "this is an algorithm problem": who holds the distribution power and how they use it is a question far older than the technology.
Uwants' story is different, and more worth examining.
It didn't shut down. After Tencent — a Chinese-capital entity — became a major investor, the site continued operating. But Alexa data at the time showed that nearly half of visits came from Mainland China; over the course of 2018, its global ranking dropped by nearly one thousand positions, and SimilarWeb's independent data showed the same long-term declining trend. These are observations from around 2018, not current figures. What can be stated is the trend and the ownership — not the present state.
The signboard is still there. The URL is still there. The forum categories are still there. The people inside have changed, the ownership has changed, and the market it faces has changed.
III. So "Disappearance" Takes Four Forms
Lay the two cases above side by side and you'll see that the disappearance of public space actually takes several forms, each feeling completely different to users:
One: Closure. The rarest, and the easiest to notice. There's a date, an announcement, someone will remember.
Two: Migration. HKGolden to LIHKG is this type. The old place still exists, but the people left. What remains is a shell. Users know what happened because they were the ones who left.
Three: Same-name replacement. Uwants is this type. The name didn't change, the URL didn't change, the interface didn't change — so there is no single moment when you'd realise "this place is no longer the place it was."
The third is the hardest to detect, because it produces no event. And following the principle established in the first two volumes: a change that produces no event is a change that will not be recorded.
All three forms lead to the same result: that room is gone. The only difference is how many people know they once lost it.
But there's a fourth form — one that only emerged in recent years, and is most easily mistaken for revival.
Four: After migration, the form was preserved but the container vanished.
On Telegram, you can find a cluster of Hong Kong community groups, rebuilt by users after the forums scattered. They transplanted the forum's categories wholesale: Idle Chat, Gaming Console, Food & Travel, Hardware & Toys, Sports & Betting, Film & Anime — even the naming conventions are identical. Over two thousand members, with unread messages accumulating into the tens of thousands.
It looks like the forum simply moved house.
But one layer is missing.
Forums have a two-layer structure: boards, then threads. Telegram's topic sections have only one layer: boards, then a message stream.
What vanished is the "thread." And the thread is the container that actually does the work. It has a title, it stays there waiting for you, you can enter and leave, you can copy a link and share it, and even if it sinks it doesn't get in the way. Opening a thread in a forum costs nothing — one person starts it, three reply, nobody cares if it's ignored.
Chat doesn't have this. A reply doesn't generate a container; it simply draws a line within the same stream. So five simultaneous conversations in one section are physically stacked on top of each other. Want further subdivision? The only tool at hand is "create another section" — using the most expensive tool to do the cheapest job. And each new section needs independent critical mass to sustain itself; diluted, it simply dissolves.
So discussion is still happening, but it no longer accumulates. Every sentence has been spoken; none endures. A 2008 forum thread can still be found, cited, and linked to today; a discussion in some group chat last year is functionally non-existent.
Nothing was deleted, and nothing remained.
This is the most complete of the four forms. The first three at least leave an identifiable loss. This one doesn't even generate the record that something "once existed."
IV. The Platform That Took Over — And Its Scorecard
After the forums scattered, discussion concentrated onto a handful of large platforms. These platforms don't use moderators to decide who rises. They use scores.
And the scores have concrete numbers.
In 2018, Facebook adopted "Meaningful Social Interactions" (MSI) as its ranking objective. According to leaked internal documents, the scoring works like this: one Like, one point; one emoji reaction, one text-free share, or one event RSVP, five points; one substantive comment, private message, share, or attendance response, thirty points. Additional multipliers apply based on the relationship between the two interacting parties.
A year earlier, in 2017, Facebook set the weight of its newly launched emoji reactions at five times that of a Like. An internal employee immediately questioned this on the message board: would weighting reactions at five times cause the News Feed to surface more controversial content? His example was blunt — post "I bought a cup of coffee" and you'd get a few Likes; post a provocative political headline and you'd get floods of angry reactions, earning five times the distribution.
A colleague replied: possibly, but the company is aware of this risk and is working to mitigate it.
This scorecard is the core of this article.
An article that makes you nod is worth one point. An article that makes you want to argue is worth thirty.
And the definition of a moderate position happens to be exactly "won't make people want to argue." It asks you to acknowledge complexity, to maintain uncertainty, to verify before judging. None of these actions produce comments, much less flame wars.
So the moderate position isn't banned. It's permanently worth only one point.
Subsequent corrections are also on record: in 2018 the reaction weight dropped to four times; in 2019 a mechanism was built to demote content receiving disproportionate angry reactions; in 2020 all reactions dropped to 1.5 times; and eventually the angry-reaction weight was set to zero. And when anger was zeroed out, the company's own data scientists found: users saw less misinformation, less disturbing content, less graphic violence — and user engagement was unaffected.
In other words, the trade-off that had always been cited as justification simply didn't exist.
V. An Honest Admission: There Is No Academic Consensus
At this point I need to hit the brakes, because the next step could easily slide into a conclusion I can't prove.
The statement "algorithms cause polarisation" has no settled verdict in academia, and the evidence contradicts itself.
In July 2023, a study on the 2020 US election, conducted jointly by Meta and external researchers, was published simultaneously in Science and Nature. The experiment switched consenting users from algorithmic ranking to pure chronological order for three months. The result: their time and engagement on the platform dropped dramatically; exposure to political content and untrustworthy sources increased; incivil content decreased — but issue polarisation, affective polarisation, political knowledge, and other major attitudinal measures showed no significant change over the three months.
In September 2024, Science published a critical letter: during the study period, Facebook had implemented a series of emergency algorithmic changes to reduce misinformation spread, and the paper had not adequately flagged for readers the possible impact of these changes. The original authors rejected this criticism and stood by their conclusions; the journal's editor-in-chief said no correction would be required but would let readers know about these concerns.
On the other side, in November 2025, Science published a Stanford team's experiment: 1,256 X (formerly Twitter) users, over ten days, using a large language model to rerank their feeds in real time, increasing or decreasing posts expressing partisan hostility and anti-democratic attitudes. The result was bidirectional — reducing hostile content raised feeling thermometer scores toward the opposing camp; increasing it lowered them; the magnitude was over two points on a 100-point scale, with no detectable difference between parties. This is causal evidence.
In 2026, Nature published another X platform experiment: turning the algorithm on shifted political attitudes in a conservative direction, but turning it on or off produced no significant effect on affective polarisation or partisan identity.
Four studies, three directions. Anyone who tells you this matter is settled is not conveying research. They're conveying a position.
Including myself. I cannot claim algorithms are the primary cause of polarisation, which is why I haven't omitted a single one of the studies that run counter to my argument.
VI. But One Finding Was Overlooked
Yet within this pile of contradictory conclusions, one fact is undisputed across three of the studies — and it happens to be exactly the one this article needs.
Meta's own research records it clearly: when users were switched to chronological News Feeds, content from moderate friends increased, content from ideologically mixed-audience sources increased, and incivil content decreased.
Read that sentence in reverse: algorithmic ranking reduced moderate content and reduced mixed sources.
This statement holds without any conclusion about polarisation. It is an exposure-layer observation, not an attitude-layer inference. What researchers debate is "does this make people more extreme" — that's a question about effects, and a hard one. But "the algorithm shows you less moderate content" is a question about allocation, and it has already been measured.
What I'm arguing is the latter, not the former.
The moderate position was not banned. It was simply ranked lower.
And this is the same action appearing for the third time, following Volume I's demotion and Volume II's detector: no deletion, no announcement, no record — just an adjustment of order.
VII. The Hong Kong Version
For the Hong Kong section, I will use the most conservative possible framing, because this is where verifiable observations are most easily stated as unprovable accusations.
I cannot write "in 2019, all of Hong Kong's free-speech forums were attacked." That sentence is too absolute, and I cannot prove it.
What I can write is: during that period, numerous online communities simultaneously faced multiple pressures — mass reporting, cyberattacks, commercial and advertising pressure, legal risk, rising management costs, and user migration. Some platforms reduced services, some changed their management approach, some gradually lost influence. All of this is on public record.
Nor can I assert that any given platform "intentionally controlled speech." Individual incidents can be analysed, management policies can be examined, but a consistent purpose cannot be reverse-engineered from outcomes. This is the standard I applied to group administrators in Volume I, and it must apply equally here — otherwise I'm simply changing the target while doing the same dishonest thing.
But one thing doesn't require those accusations to state:
In an environment that only rewards high-engagement content, what disappears first is not either extreme. It's the people willing to say, "It's not that simple."
Because both extremes each have communities, shares, ready-made enemies, and thirty points on the scorecard. The middle has nothing. You can support, you can oppose, but you cannot be sober — sobriety doesn't generate comments, sobriety has no score.
This isn't unique to Hong Kong. The US, Taiwan, and the UK all have the same discussion, and the controversy is equally unresolved. Hong Kong is merely one case, and its distinctiveness is that the collapse of forums, the concentration of platforms, and a period of intense social upheaval happened almost simultaneously — so the three effects are stacked on top of each other and are very hard to disentangle after the fact.
VIII. Where Did Those People Go?
After the forums scattered, the most natural question is: where did the people go?
I asked this question for a long time. And the shape of the answer itself illustrates the problem.
First: the people didn't vanish. They stopped speaking publicly.
An Incogni survey from June 2026 showed that 55% of respondents post less than they did five years ago; 53% have become more selective about "who can see"; 47% have at some point deleted social or messaging apps due to stress or anxiety. More than half said maintaining an online presence feels like a job, rising to 60% among Gen Z. Instagram head Mosseri has also acknowledged that DMs are now the primary way people share on IG.
In other words: the broadcast layer is shrinking, while the sharing layer is actually larger than ever.
The difference is that the private layer cannot be reached unilaterally. You can't send an article into someone else's group chat — it can only be forwarded in by someone else.
Second: the public arena hasn't entirely withered, but the concentration is staggering.
Threads accounts for 21% of its global traffic from Taiwan — the highest in the world. Taiwanese users spend twenty-two times as long on it daily as American users. By July 2025, Taiwan's Threads posting volume had already surpassed Facebook.
One place growing against the trend while everywhere else contracts — this isn't revival. It's another round of concentration.
And Hong Kong has no such receiver. People dispersed to Threads, IG, Telegram channels, YouTube, and physical diaspora communities. The inability to find "the place where Hong Kong readers gather" isn't because the search wasn't thorough enough. It's because that place doesn't exist.
Third, and the most troublesome: departure is invisible.
No one posts "I'm quitting." Even if they did, the ranking system wouldn't surface it for you. And an account that has stopped posting is, in your News Feed, exactly the same as one that never existed.
This yields a rule: information about platform attrition can only propagate through non-platform channels. Any "is everyone still here?" survey conducted on the platform will systematically overestimate that platform, because those who've left aren't eligible to answer.
I personally learned about this not from any data, but from someone mentioning it in an offline classroom. That wasn't luck. It was the only possible pathway.
So the question "where did everyone go" can never be answered from inside the platform. And a society that cannot surface this answer also cannot know how much of its discussion it has lost.
IX. What Disappeared Was Not a Position — It Was the Space
So what was the real loss?
Not any particular opinion being silenced. When an opinion is silenced, people protest, records are made, it becomes news.
What truly disappeared is the space that could accommodate complexity.
A society doesn't need to shut down a single forum to lose public discourse. It only needs to make rational discussion increasingly hard to sustain: make the moderate position permanently worth only one point, ensure those willing to verify can't get distribution, turn "I'm not sure" into a statement that doesn't exist on the scorecard.
And the defining feature of complexity is precisely that it doesn't protest.
Nobody takes to the streets for "it's not that simple." Nobody holds a memorial for a position that no longer has a place. When it disappears, it makes no sound — because it was always the kind of thing that doesn't make sound.
This circles back to the thing all three volumes share:
Volume I: a person was demoted without knowing. Volume II: the criteria were replaced without knowing. Volume III: an entire society lost a category of discussion without knowing.
All three are the same structure — the loss left no trace, so the loss will not be counted.
A design feature written into a patent in 2015 ultimately scaled not to a single author unable to see their own post. It scaled to a society unable to see what it has lost.
And this argument rests on only three verifiable facts: the rise and fall of forums is documented, the scorecard has numbers, and the exposure-layer change has experimental data. Whether all of this adds up to making people more extreme — that question is left to those still doing the research.
I only want to point out something smaller and more certain: the place where you could speak slowly, where you could say "I'm not sure," where you could be corrected by others without it turning into a fight — that place is gone.
And when it disappeared, no one needed to be held responsible.
Sources
- Wikipedia: HKGolden (Hong Kong Golden Forum), Uwants Forum (operational timeline, schism events, ranking changes, ownership and investors)
- Hong Kong Internet Encyclopaedia: HKGolden history, commentary on Uwants
- The Washington Post (October 2021), based on internal documents submitted by Frances Haugen to the US Securities and Exchange Commission and Congress: emoji reaction 5x weighting, subsequent adjustments, and zeroing
- US House of Representatives public documents: Facebook MSI scoring weights (Like 1 / reaction and text-free share 5 / substantive comment 30)
- Guess et al., Science 381 (2023); Nyhan et al., Nature 620 (2023): Meta 2020 US Election Study
- Science (September 2024): critical letter and editor-in-chief's statement
- Piccardi et al., Science 390 (2025): field experiment on real-time reranking of partisan-hostile content
- Gauthier et al., Nature (2026): political effects of the X platform algorithm
- Incogni (June 2026): social media usage behaviour survey
- Adam Mosseri's public statement on Instagram sharing behaviour shifting to DMs
- Threads Taiwan usage data (2025–2026 compiled statistics)