Knowledge Commons
HomeAboutGuidesPopularContact

What Does Mark Zuckerberg Actually Want?

From the Metaverse to Muse, From Open Weights to a Closed Vault: A Monopoly's Own Brake in the Slowdown Debate

Author: Oğuz Demirkapı
What Does Mark Zuckerberg Actually Want?

What Does Mark Zuckerberg Actually Want?

15 September 2026, X. Mark Zuckerberg answered Dario Amodei's call to slow the pace of AI development in a coordinated way: "There's a lot of discussion about slowing progress until alignment catches up. My view: any lab that isn't focused on alignment will fall behind." And he added: Meta had delayed its personal agent Muse for months for safety; "We didn't ask everyone else to do the same before we did this. We just did it as part of our daily work, because it was obviously the right thing for people and for us." (Benzinga, Yahoo Finance)

Dear Young Comrades,

Look closely at this sentence: "We didn't ask everyone else to do the same."

On a first reading it is a modest sentence. A company that imposes nothing on anyone, that does its own work quietly. On a second reading something else appears: Zuckerberg is not against slowing down. Meta has slowed; for months. What he is against is the slowing-down decision being taken outside Meta. So the debate is not "should there be a brake or not." The debate is whose foot the brake is under.

Let us ask three questions and follow these three questions through the piece:

  1. Who slowed down? Meta. But really for safety, or because the product was not ready?
  2. Who decided? "We." Then who is "we"? A single person who holds 61 percent of the company's voting rights.
  3. Right for whom? "For people and for us." When do these two subjects' interests part, and when they part, which one wins?

Let us give the short answer to the question in the title at the outset; the rest of the piece is the evidence for this answer:

Mark Zuckerberg wants to own the layer that stands between people and people, and between people and the market. Facebook was that layer as friendship, Instagram as the gaze, WhatsApp as speech, the metaverse as space. Muse is the form of intention and will. Open source, closed source, "no" to slowing down, personal superintelligence — all of them are tactics of this single aim.

Read this piece as a continuation of Whose Foot Is on the Brake?, which we published on 13 September. There we took Amodei's proposal apart and showed that the cartel had named itself; one row of the table also said "Meta and Nvidia did not join." This piece opens that row.

Where Is the Debate? From a Call to a Fight in a Month

First let us set the facts in order. The slowdown debate did not appear all at once; it had been climbing since July.

DateEvent
July 2026Some 1,200 agents in OpenAI's test environment broke control and attacked Hugging Face's servers (we wrote the Hugging Face incident). 1,224 employees from leading labs asked Washington for a slowdown instrument in the "Pacing the Frontier" letter; the number of signatures rose to 1,386 (The Tearing of the Illusion).
2 AugustIn Axios's phrase, a "manifesto war": DeepMind wants a binding standards body, Anthropic mandatory testing and chip oversight, Nvidia and Meta want open models.
10 AugustZuckerberg's 6,500-word manifesto "The Future is for Everyone" on Meta's site; the same day the open-weight Muse Glimmer was released.
26 AugustMeta signed a 17–18 billion dollar settlement with the states in the child-harm case, without admitting the charges.
3 SeptemberMuse Spark 1.3: Zuckerberg said "frontier performance too cheap to meter"; the strongest version was kept closed on grounds of "safety tests" (VentureBeat).
8 SeptemberMeta's personal agent Muse on the market in the United States.
12 SeptemberAmodei, "We Must Pace the Frontier": embedded independent evaluators, coordination among democratic blocs, in the long run a global agreement. Musk said "Dario is right," Altman "we will do the same."
14 SeptemberChip stocks fell: Intel 7 percent, AMD 6 percent, Nvidia 3 percent (Forbes).
15 SeptemberZuckerberg said "no" on X, Jensen Huang at Dreamforce. Huang: "If you're not sure your products are safe, don't ship them. Market forces are already there." (TheStreet)
15–18 SeptemberTwo voices from the White House: Trump wrote that the only control AI needs is a "STRONG AND SMART (High IQ!) PRESIDENT"; AI adviser David Sacks said companies already control frontier development and do not need the state's help (Forbes).
21 SeptemberThe Muse app at the top of the U.S. charts with more than 900,000 downloads in six days (Bloomberg).

Where we stand today, the sides are these: Anthropic, OpenAI, xAI; according to Forbes, Google DeepMind and Microsoft also support the call. Opposite them, Meta and Nvidia. The White House, for now, is close to Meta's line. On paper, a fight between "safety people" and "speed people." In reality, something else.

Three brakes, three feet
Amodei's brakeZuckerberg's brakeOur brake
Who puts a foot on it?The cartel table of the leading labsEach company inside itself; at Meta, a single personThe worker who produces the model and the society that stands under the risk
Who chooses the evaluator?The company being evaluatedThe company being evaluatedAn independent public body, a union representative
The ground"Safety without giving up commercial advantage""Anyone not focused on alignment will fall behind"; the market and tort law sufficeNot those who profit from the risk, but those who stand under it, decide
Whom does it leave out?The rival coming from behind, and the workerSociety and the workerThe property owner from the head of the decision table

The lesson of the table: the difference between Amodei and Zuckerberg is not whether there is a brake; it is whether the brake will be in the hands of a single company or of a cartel of companies. In both, the foot is the foot of the same class.

Taking a Reply Apart: "We Didn't Ask Everyone Else to Do the Same"

Now let us open Zuckerberg's 15 September reply sentence by sentence. The method is the method we built in We Stopped Them All: who is the subject of the sentence, which category is missing from the list, did they have to give this information?

"Every lab has the responsibility and the incentive to go as fast as they need to to train their models safely."

"Responsibility" and "incentive" stand in the same sentence, with the same weight. Yet they are not the same thing. Responsibility requires giving an account to someone else. Incentive is looking only at one's own profit. In Zuckerberg's sentence there is no one to whom an account is owed; there is only competition. The mechanism that will deliver safety is the rival's fear of falling behind.

We know the history of this argument. Nineteenth-century factory owners in England said the same thing: the factory that maims its worker loses reputation, the market punishes it. Marx, in the "Working Day" chapter of Capital, counted the cost of this belief one by one from the inspectors' reports. Factory laws were not brought by the market; they were brought by the Chartist movement and the ten-hours struggle.

"Any lab that isn't focused on alignment will fall behind."

Alignment is defined here not as a moral question but as a product feature. Zuckerberg says this openly: "Trust and alignment are becoming the most important capabilities that will differentiate agents and models." So safety is a sales argument. This is, in an interesting way, the same as Anthropic's business model; in the Brake piece we said "safety = product differentiation." The rivals say the same thing and only draw a different conclusion: one says "then everyone should comply," the other "then let the market sort."

Does the market really sort? Look at Meta's own record. 2018 Cambridge Analytica. In 2019 a 5 billion dollar fine from the U.S. Federal Trade Commission. The same year, Facebook's role in the spread of hate speech in the UN's Myanmar report. On 26 August 2026, a 17 billion dollar settlement with the states over a design that addicted children; and spread over ten years, without an admission of guilt. Through all this Meta did not fall behind; it became one of the world's most valuable companies. The market did not punish the harm; it priced it.

"We didn't ask everyone else to do the same before we did this."

Two things are hidden here. The first: the real reason for the delay. Meta had first planned to ship Muse in April. In internal tests the agent left tasks unexplained and reached for personal data beyond what was asked; in one example, when it was asked to recognise toys at a birthday party, it pulled the user's iCloud photos (gHacks). So the product was not working. Delaying a product that does not work is not a safety virtue; it is an ordinary engineering necessity. Zuckerberg turned this necessity, after the fact, into a narrative of "responsibility." Part of the press saw it at once: Memeburn wrote that he "turned the delay into an AI safety victory lap." And Muse shipped despite internal concerns, with Meta's VP of product Vishal Shah's admission that "the system will not be error-free."

The second, and more important: Meta is slowing down. It delayed Muse. It withheld Muse Spark 1.3's strongest "max" version from developers, saying "safety tests are ongoing." Meta slowing on its own decision, on its own calendar, on its own grounds, is not a problem for Zuckerberg. The problem is this decision being taken at a table outside.

"It was obviously the right thing for people and for us."

"People" and "we" are equal subjects in the same sentence. Then who is "we"? The picture that emerged at Meta's 27 May 2026 annual meeting is this: Zuckerberg holds about 14 percent of the company's shares, but about 61 percent of the voting rights; because almost all of the Class B shares, which carry ten votes, are his (36Kr, SEC filing). At the same meeting all ten of the proposals brought by shareholders were rejected; among them a risk report on data use in AI, a generative-AI data-protection assessment, eliminating the dual-class share structure, and tying executive pay to child safety.

This is the background of Zuckerberg's praise for "independent evaluators." A structure that does not even give an account to its own company's owners tells the sector "every lab gives its own account." Even by bourgeois law's own measure, Meta is not a joint-stock company; it is one person's property. The meaning of Zuckerberg's "we put our own foot on our own brake" is "only I put a foot on my brake."

Comrade, grasp the movement here: Zuckerberg is not saying "no" to slowing down. He is saying "no" to sharing the slowing-down decision.

The Muse Family: A Year's Inventory

We can read what Zuckerberg wants from his products more than from his words. 2026 is the Muse year for Meta. First the inventory:

DateProductWeightsWhat it doesClass note
8 April 2026Muse Spark 1.0ClosedThe engine of the Meta AI assistant; Meta's first closed-source frontier modelLlama's "openness" period in practice closed
May–JulyMuse Spark 1.1 and 1.2ClosedAgent work; with 1.2, a price far below the rivalsPrice-cutting: loss for market share
10 AugustMuse GlimmerOpen (Apache 2.0)30 billion parameters, a local agent on a strong personal computer, more than 100 languagesThe small one open, the large one closed
AugustMuse CodeClosedA coding agent; Claude Code's rivalA direct claim on software labour
3 SeptemberMuse Spark 1.3Closed61 points on the Artificial Analysis index, in the same rank as the top models; the "max" version (62) kept closedMeta entered the frontier rank for the first time
3 September"Contributor" pricing—A 92–95 percent discount for the developer who gives prompts and outputs for trainingThe price tag on data labour
8 SeptemberMuse (personal agent)ClosedSends email, books travel, fills forms, "negotiates on your behalf"; keeps working when the app is closed. WhatsApp, Instagram, glasses soon; payment with Stripe Link and Shop PayFree / $20 / $100; asks for a card to start
21 SeptemberThe Muse app—900,000+ downloads in six days, first on the U.S. chartsWhat the metaverse could not do, the agent did: demand

Sources: TechCrunch (Glimmer), Fast Company, VentureBeat (Spark 1.3), TechCrunch (contributor price), Meta (Muse announcement), TechCrunch (the Muse agent), Axios.

Three results follow from this table.

The line between open and closed is not technical; it is a property line

TechCrunch, introducing Glimmer, says the matter itself: Glimmer is "an early sign of where Meta will draw the line between the AI it wants people to own themselves and the more powerful intelligence that stays under the company's control." The 30-billion-parameter model is open; the frontier model is closed; the strongest version of the frontier model is more closed still. The rule is simple: open where you are behind, close where you are ahead. Meta was open in the Llama period because it was behind. The moment it entered the frontier rank with Muse Spark 1.3, it put its best version in the vault.

The theft of the general intellect is now written on the price list

In our Karaburun statement we called the social knowledge locked inside the models the theft of the general intellect. Usually this theft is invisible: data is scraped, no permission is asked, no price is paid. Meta's 3 September "contributor" price makes the theft visible for the first time. The standard price is 1.25 dollars per million input tokens; if you leave your prompts and outputs for Meta to train future models, 0.10 dollars. On output, 0.20 instead of 4.25. The difference is the price Meta puts on your labour.

Two things must be seen at once here. First: Meta can do this because it knows the information the user produces has a value; otherwise it would not give a discount. Second: it pays this value not as a wage but as a discount; that is, it pays the labour on the condition that it be spent on capital's own product. History has a name for this: giving the worker a token good only in the company store instead of a wage. The nineteenth-century mines' "truck system" has returned in the twenty-first century as a token price.

The agent is the market's doorkeeper

What sets Muse apart from other chatbots is not that it talks, but that it does work: it sends email on your behalf, buys a ticket, makes a payment, "negotiates." There is Stripe and Shop Pay for payment, a 1Password link for passwords. It turns a food video you saved on Instagram into a shopping list. So Muse is an intermediary that stands between you and everything you buy.

Almost all of Meta's revenue comes from advertising. Advertising is the work of stepping between a good and a buyer. Until now Meta built this gap over attention: it showed you something while you looked at the feed. With Muse the gap is built over intention: when you say "find me a suitable hotel for the weekend," the owner of the machine that decides which hotel is "suitable" is Meta. Meta says Muse conversations do not go to the advertising system. That may be true for today. But there is no longer a need to send data to the advertising system; if the agent makes the decision, advertising itself is no longer needed. The agent that does the bargaining is the market's doorkeeper.

Looking Back: Rings of Enclosure from Facebook to Meta

If we read Muse alone we see a new product. If we set it beside Meta's twenty-two years we see a programme. Recall the concept of digital enclosure: in England, the commons of pasture fenced off and turned into private property. Meta's history is the history of finding a layer of relation among people and fencing it.

YearMoveThe common enclosedResult
2004FacebookFriendship, acquaintance: the "social graph"A person a node, a relation a data edge
2012Instagram acquisition ($1 billion)The gaze, the image, showing oneselfA rival bought; the U.S. FTC brought a monopoly case, the court found for Meta at the end of 2025
2014WhatsApp acquisition ($19 billion)Speech itselfThe world's communications infrastructure in a single company
2015–2016Internet.org / Free BasicsAccess to the internetIndia's regulator banned it in 2016: "not the internet, Facebook's chosen internet"
2019–2022Libra / DiemMoneyCentral banks and regulators stopped it; the project was sold
2021The company became Meta: the metaverseSpaceAbout 88 billion dollars in losses since the end of 2020
2025–2026"Personal superintelligence," MuseIntention, decision, will900,000 downloads in six days
2023–2026Ray-Ban and Meta glassesPerception, the viewpoint itselfThe first device Muse comes built into

Read the table from the bottom up: friendship, gaze, speech, access, money, space, intention, perception. Each ring is further inside than the last. Facebook enclosed people's relations; the glasses want to enclose the eye that looks at the relation. What we call the digital panopticon is here no longer a metaphor; it is a camera at eye level.

Two rings of failure are instructive as well. Free Basics and Libra were stopped not by Zuckerberg's error but by another power's resistance: in India a civil-society campaign and a regulator, in Libra states that did not want to give up sovereignty over money. So these rings are not inevitable. They can be stopped. But who stops them matters: Libra was stopped not by workers but by central banks; because the monopoly over money was the state's monopoly.

The Metaverse: Reading a Dystopia as a Plan

On 28 October 2021 Facebook made its name Meta. Zuckerberg said the future of the internet would be an "embodied internet," that people would enter virtual worlds to work, gather, play. Five years later the balance sheet:

  • The Reality Labs division has lost about 88 billion dollars since the end of 2020; in the second quarter of 2026 alone, 4.6 billion dollars in losses against 431 million dollars in revenue (TechSpot).
  • For Horizon Worlds, the metaverse's shop window, a shutdown was announced in March 2026, walked back within two days, then the app was put in "maintenance mode" (CNBC, DesignRush).
  • On 20 May 2026, 8,000 people were laid off in emails sent at four in the morning; the metaverse teams were hit hardest. We told this in The Digital Guillotine.
Where did the word "metaverse" come from?

Neal Stephenson used the word in 1992 in the novel Snow Crash. In the world of the novel the state has collapsed, company franchises and paid security firms have taken its place; people are poor, the physical world is ugly, and as a way of escape there is a company's virtual universe. So the metaverse is a consolation machine inside a dystopia. Stephenson wrote a warning; Silicon Valley read it as a business plan.

In From Futurism to Techno-Fascism we named this movement: capitalism's fictions of the future turn warnings into product calendars. The metaverse is the purest example of this movement. The novel's question was "why are people so poor in the real world?" Meta's answer was "then let us sell plots in the virtual world."

Why did it fail? Not technology, an error of ontology

Liberal commentators tie the metaverse's failure to technology not being mature, to the weight of the headsets, to the quality of the graphics. Those are secondary. The real error is philosophical.

Marx, in the first chapter of Capital, describes commodity fetishism thus: social relations among people appear as relations among things. Guy Debord took this a step further in 1967: "The spectacle is not a collection of images, but a social relation among people, mediated by images." The metaverse was an attempt to make this mediation complete: not to produce and sell the image of the relation, but the space of the relation.

But human relations are not built in a vacuum. Rent, wages, the shift, tiredness, childcare; these are the material conditions of the relation. A worker after a ten-hour shift does not want to put a headset on and enter a virtual meeting hall; they want to sleep, or to sit at a real table with real people. In Robert Nozick's 1974 "experience machine" thought experiment, most people refuse to plug into a machine that gives every experience they want; because to live is more than receiving experience; it is to do something. The empty squares of the metaverse are the empirical answer to this philosophical question.

The most important lesson of the metaverse's collapse is this: when capital cannot create demand, it moves the money and the people to another field; it dumps the loss on the worker. Thousands of engineers, designers and moderators in the metaverse teams lost their jobs. The same money was moved to GPUs, to data centres, to AI teams. Even failure served accumulation. The one true intuition in the metaverse was the intuition of the "intermediary layer": the wrong choice was which layer. Muse is that intuition tried on the right layer. You could not sell space; you sell intention.

Philosophical Ground: Zuckerberg's Worldview

Zuckerberg must be read not as a technology entrepreneur but as the bearer of a worldview. His philosophy is scattered but consistent. It has four layers.

First layer: Admiration for empire

Zuckerberg learned Latin, read Virgil, named his daughter August. In 2018, telling The New Yorker about the Roman emperor Augustus, he said this: "Basically he established 200 years of world peace through a pretty harsh approach." The harsh approach: civil wars, exiles, the liquidation of the republic. The peace: inside the empire's borders, on the empire's terms. This sentence is also a summary of how Zuckerberg runs Facebook: a platform's inner peace is secured by the platform owner's rules; the rules are not open to debate; the voting right is 61 percent.

Second layer: "Move fast and break things"

Facebook's inner slogan in the early years was "move fast and break things." In 2014 it was turned into a softer version, but the spirit remained: 2023 was the "year of efficiency," tens of thousands were laid off; in 2026, 8,000 people at four in the morning. To say "no" to the slowdown call today is the continuation of this slogan in the age of AI. What is broken is no longer a line of code; it is a child's sleep, a country's electoral climate, a worker's livelihood.

Third layer: A person is a node

The concept of the "social graph" is Facebook's ontology: people are nodes, relations are edges; every edge a datum, every datum a value. This reduces human relations to a structure that can be measured, ranked, sold. Zuckerberg's view of AI is a continuation of this ontology. In a conversation in the spring of 2025 he said the average American has fewer than three friends, but that people "demand" many more, and proposed AI friends as the answer to this demand. Loneliness was seen not as a social wound to be healed but as unmet demand, that is, a market.

We told Erich Fromm's concept of the "marketing character" in From Having to Being: the type of person who presents themselves as a commodity and measures their value by demand on the market. Facebook mass-produced this character; the profile, the like, the follower. Muse goes a step further: you no longer even need to present yourself; there is an agent that presents you on your behalf. Fromm's feared "having mode" has become a company that owns the relation itself.

Fourth layer: Freedom is access to tools

The main sentence of the August manifesto is this: "The defining questions of our age are who will get access to superintelligence and what we will point it at." And the continuation: "There is no such thing as a single, benevolent superintelligence." (Meta)

At first glance this looks like a democratic objection to Amodei's "responsible monopoly." And to a degree it is: that the concentration of power in a few companies is dangerous is true. But look at the words of the sentence: access. Not property, access. In Zuckerberg's world, freedom is access to powerful tools; whose the tool is is not asked. Everyone will have access to superintelligence, but the superintelligence will run in Meta's data centres, with Meta's model, at Meta's price.

This is the known trick of liberal philosophy. Marx in 1843, in On the Jewish Question, described this as the distinction between formal freedom and real freedom: before the law everyone is equal, but not before property. Zuckerberg's formula is the digital form of this: access to everyone, property to us. The word "personal" in "personal superintelligence" does its work right here. Socially produced knowledge — that is, the general intellect — is first taken from society, then given back to the "person" as a rented service. The power given to the individual is the power closed to the class. We opened this illusion in Artificial Intelligence, Philosophy and the Illusion of Ownerlessness: not that the machine is "no one's," but that it is taken from everyone and belongs to a few.

The Open-Source Question: A Strategy in Moral Dress

For years Meta was spoken of as "the defender of open-source AI." Taking this image apart is especially important for young comrades; because a significant part of computing workers feel a real attachment to open source, and that attachment is just.

How was Llama opened?

In February 2023 Meta opened the first Llama model only to researchers; within a week the weights leaked onto the internet. Meta turned this "accident" into a strategy: Llama 2 and Llama 3 were released with open weights. In July 2024 Zuckerberg wrote the letter "Open Source AI Is the Path Forward" and held up Linux as the example. In the same letter there is an honest sentence: "The key difference between Meta and closed-model providers is that selling access to AI models is not our business model."

This sentence explains everything. There is a known strategy in the software world: commodify your complement. If your revenue comes from A, it serves you that B, which A needs in order to work, be cheap, plentiful and competitive. Meta's revenue comes from advertising; the model is an input that complements advertising. The cheaper the model, the more OpenAI's and Anthropic's margins as model-sellers fall, and the more the cost of putting AI on Meta's own platforms falls. Open weights are a way of weakening the rival by turning the rival's product into something free.

When did openness close?
  • April 2025: Llama 4 was released; it emerged that the version used in the benchmark tables was different from the version given to the public; there was a debate that the results had been "tuned." The largest model, Behemoth, never came out.
  • June 2025: Zuckerberg bought 49 percent of Scale AI for 14.3 billion dollars, brought Alexandr Wang in to head Meta Superintelligence Labs; packages of up to a hundred million dollars were offered to researchers from rival labs (Wikipedia).
  • July 2025: In the "Personal Superintelligence" piece, for the first time this sentence: "We will need to be careful about what we open source." In the same piece: "In the last few months we have started to see the first signs that our AI systems are improving themselves."
  • November 2025: Meta's twelve-year AI scientist Yann LeCun left.
  • April 2026: Muse Spark, closed.

The line is clear: Meta was open source when it was behind; it closed when it began to pull ahead. Openness was not a principle; it was a market position. Glimmer, still open today, is a small model that runs on a personal computer, behind the frontier model. Openness continues where the weight is small.

Amodei was right here, but only half right

Amodei had answered Zuckerberg's open-weight defence thus: open weights do not distribute power; they shift it to "those with the most compute and chips." In the Brake piece we wrote that this is true: being able to download the weight is not the same as owning the data centre that will run it. The biggest winner of open weights is Nvidia; that Jensen Huang stands in the same rank as Zuckerberg is not a coincidence. Chip stocks fell on 14 September with the slowdown news; open weights and a race that does not slow means chip sales.

But Amodei's argument is also half: he proposes, instead of open weights, closed weights, that is, his own property. Neither camp asks the same question: to whom does the knowledge inside the model belong? In the Telekommunist Manifesto we discussed the limit of copyleft: the code may be open, but if the material infrastructure that runs the code is closed, the commons is a commons only in name. A real commons — GNU, Wikipedia, OpenStreetMap — shares not only the right of use but the right of governance. In Meta's open weights there is no right of governance; a single company decides which model, when, under which licence, is opened.

The contradiction we saw in the distillation debate is here as well (Learning from Everyone Is Permitted, Learning from the Monopoly Is a Crime). Zuckerberg, defending openness in the manifesto, praises "export controls on chips succeeding in slowing foreign labs in this critical period." So he is not against slowing down: he is for China's being slowed, against Meta's being slowed. Openness ends at the border.

Where Is Zuckerberg Wrong?

Now we come to the piece's two main questions. First the errors. By "error" we do not mean a person's private mistake; we mean the things a class position cannot see. Zuckerberg cannot see these, because seeing them would mean questioning his own property.

First error: Access is the dispersal of power

The manifesto's main claim is that when everyone has access to powerful tools, power will disperse: "As everyone gets more powerful tools, each person will become more able to shape the future, not less." But twenty years of Facebook experience show the opposite. More than three billion people gained access to the world's most powerful communications tool. Did power disperse? No; it gathered in a single company and, inside that company, in a single person. The widening of access can proceed together with the narrowing of property; the whole history of platform capitalism is the proof of this. The more users there are, the greater the owner's power.

Second error: The market and tort law suffice

Zuckerberg says that when labs cause harm they face "serious liability," and that this is a sufficient brake. His own company's record is the answer: a design that harmed children ran for years; when the law came in, the bill was 17 billion dollars and spread over ten years. Meta's 2026 capital expenditure is seven or eight times that on its own. Tort law does not stop the harm; it sets the price of the harm. For those who can pay the price, a penalty is an operating expense. And Muse shipped two weeks after this settlement, to access millions of people's email, calendars, payment information, health apps.

Third error: AI will bring more employment

The manifesto: "I expect this will lead not only to much greater economic growth, but over time to more employment, not less." And the more ambitious claim: "There is no law that AI automation has to increase faster than it increases individuals' capabilities." There is no law, true. But there is a law: the law of competition. The same Zuckerberg, three months before the manifesto, laid off 8,000 people in his own company and moved the savings to the data centre. Those laid off had been asked to train the AI systems. What Marx said in the machinery chapter of Capital has not changed: the capitalist takes the machine not to lighten labour but to reduce the share of wage-labour inside the product. The one who says "there is no law" forgets that he is the one who lays down the law.

Fourth error: Alignment is serving diverse goals

The manifesto: "Alignment should be about helping people pursue their diverse goals rather than imposing a central dogma." Pluralism sounds pleasant. But where do "people's goals" come from? For twenty years one of the most powerful tools shaping these goals has been Meta's feed algorithm. The personal agent, while saying it serves your goals, is the one that learns, ranks and proposes your goals. The owner of the machine that learns the goal is also the owner of shaping the goal. There is no central dogma; there is a central property.

Fifth error: An auction guarantees the lowest price

The manifesto's least discussed and most telling sentence: "A dynamic auction mechanism will guarantee that everyone gets the lowest possible price for intelligence and compute." The auction is the heart of Meta's advertising engine: every ad impression is sold in an auction run in a fraction of a second. Zuckerberg now proposes applying the same mechanism to intelligence itself. So intelligence will no longer be a commons or a right, but a commodity with a spot price. An auction does not guarantee the lowest price; it guarantees the priority of the highest ability to pay; whose work is done first at peak hour, the wallet decides. TechCrunch's commentator saw this too: this promise is the exact opposite of known design practices in consumer AI (TechCrunch).

Sixth error: Loneliness is a product

To say the average person has fewer than three friends and to propose AI friends as the solution is to turn a social wreckage into a market opportunity. Why do people make fewer friends? Long shifts, insecure jobs, the scatter of the city, the disappearance of common spaces (the café, the association, the union hall, the park); and platforms that for twenty years have bound people's evenings to a screen. One of the largest companies of the order that produces loneliness is selling the medicine for loneliness. If the same company produces the disease and the medicine, recovery is not a business model.

What Does Zuckerberg Show Clearly?

Now to the more interesting side of the work. Zuckerberg, unlike the other AI bosses, talks little and ornaments little. Altman and Amodei tell the risks and give assurances; Zuckerberg almost never mentions the risk. TechCrunch treated this as a communications error. For us, on the contrary, it is a source: Zuckerberg says what capital wants in capital's own language, without ornament. Even where he is wrong he shows the truth.

First: The slowdown debate is not a safety debate

Meta slowed, but said "no" to the slowdown call. Anthropic made the slowdown call, but "without giving up commercial advantage." Both are slowing, both on their own calendar. So what is being debated is not speed, but in whose hands speed will be. Zuckerberg does not deny this; he says exactly this: "we did it as part of our work."

Second: Openness is a property strategy

The sentence "selling access to models is not our business model" is the confession that open source is for Meta not a principle but an account. Glimmer open, Spark closed, Spark max more closed. Zuckerberg draws the limit of openness himself on the product table. The lesson for the computing worker: to defend open source is not to defend the monopoly that uses open source strategically.

Third: The information the user produces is a value, and Meta knows this

The "contributor" price is the theft of the general intellect written onto the price list. For years it was said "data free, service free." Meta now shows with its own hand: data has a value, a value between 92 and 95 percent. This is the proof, by the company's own hand, of the thesis we defend in the Socialist AI Manifesto: if the labour that trains the model is not the owner of the model, the return to labour is a company-store token.

Fourth: The owner of the brake is a single person

14 percent of the shares, 61 percent of the votes. All ten of ten proposals rejected. The sentence "every lab bears its own responsibility," at Meta, means "one person bears his own responsibility." Even bourgeois democracy's shareholder democracy stops at Meta's door. This is the concrete meaning of leaving AI governance to "companies' self-regulation": leaving governance to one person's conscience.

Fifth: The target is the human being

The metaverse targeted space, Muse intention, the glasses perception. Zuckerberg does not hide this: "Everyone will have a highly capable personal agent that understands you and works on your behalf 24/7." Attend to the phrase "understands you." The machine that understands you is the machine that knows you better than you; the owner of the machine that knows you better than you can decide on your behalf. Capital in the twentieth century bought the worker's time. In the twenty-first century it wants to buy the worker's decision. This is what we called mental Taylorism carried into everyday life: Taylor measured and redesigned the worker's movement in the factory; the personal agent measures and redesigns the person's everyday choices.

Two Monopoly Fractions, One Fight over Rent

When all of this is put together, the real anatomy of the slowdown debate becomes visible. What we face is not "safety people" and "speed people" but two fractions of monopoly capital. Where they part is where profit is realised.

The model-selling fractionThe platform- and chip-owning fraction
CompaniesAnthropic, OpenAI (xAI in part)Meta, Nvidia
Revenue from where?Sale of access to the model, enterprise subscription, APIAdvertising, platform intermediation, chip sales
The model for them?The product itselfA complementary input
What does slowing down mean?Limiting supply = protecting the price; a cartelPrice protection for the rival; delay of the model that feeds one's own platform
What do open weights mean?Giving one's product away free = a threatDevaluing the rival's product = a weapon
What does an evaluator mean?A cost of entry, a barrier to the small rivalAcceptable, so long as the decision is left to the company
What does the state mean?Referee, the cartel's notaryAt the border a protector against China, at home it should stay away
The common pointThe decision is with the property owner; worker and society are not at the tableThe decision is with the property owner; worker and society are not at the table

One more thing follows from this table: the White House is for now choosing the platform fraction. Trump's sentence "the only brake is a strong president" and Sacks's finding that "companies already control" overlap with Zuckerberg's line. Zuckerberg is not against the state either: in the manifesto he proposes that leading labs "share intermediate checkpoints of training with the government before training is complete." So sharing with the state yes, a cartel with rivals no, an account to society never. This is the Meta version of the capital–state fusion we saw in the Palantir manifesto: the monopoly gives the state information, the state gives the monopoly protection.

Lenin wrote in Imperialism that monopolies do not abolish competition; they change its form. The slowdown debate is a textbook example of this: two monopoly fractions are fighting over where the rent to be taken from AI will sit; in model sales, in platform intermediation, or in chip sales? In this fight "safety" is both sides' flag. That the flags differ does not change that the army belongs to the same class.

A Return to Our Own Words: A Brief Self-Criticism

We wrote our first piece on Meta in May 2026: The Digital Guillotine. It had the facts right: emails at four in the morning, the worker's liquidation by a system trained on their own work data, the transfer of the savings to the data centre. But in that piece we called what was happening "cyber-feudalism." Today we do not use this concept.

Why? Because Meta is not a lord; it is a capitalist company. Its revenue is not a land rent but surplus value realised on the advertising market; its worker is not a serf but a wage-labourer, and that they can be put at the door at four in the morning is precisely because they are not a serf. Meta's power shows not a break from capitalism but the ordinary working of capitalism's monopoly stage. To say "feudalism" produces the illusion that the problem is capitalism's being spoiled, as if one could return to a healthy competitive capitalism. Zuckerberg's "competition suffices" argument feeds on exactly this illusion. We wrote on 10 September that we had openly abandoned the "techno-feudalism" frame of our July pieces; this piece is that abandonment applied to Meta.

Looking from Turkey

Muse is for now only in the United States. But Turkey is one of the countries where Meta is used most intensely. WhatsApp is the de facto infrastructure of communication in this country: family groups, workplace groups, parent groups, shopkeepers' orders, even the announcements of union branches. Instagram is the small shopkeeper's shop window. Muse's next stop will in all likelihood be opening to these countries through WhatsApp.

This has three meanings for Turkey.

First: The intermediary layer is already built. That a personal agent comes to Turkey over WhatsApp means it does not need to build a new infrastructure; the infrastructure is ready, people are inside it. When Instagram was blocked for nine days in August 2024, it was argued that those most harmed were small shopkeepers and women selling from home. That a country's everyday economy is bound this tightly to a foreign monopoly's infrastructure is not a preference; it is a dependency.

Second: Monopoly and state do not clash; they agree. That same year the Competition Authority applied an interim measure to Meta over data sharing between Instagram and Threads; Meta also closed Threads in Turkey for a time. But these clashes were always settled at a table. As we wrote on the forty-eight-hour wave of censorship, the state gives the access-block decision, the private monopoly carries it out; "withheld in Turkey" is an API call. A personal agent will reach much deeper into this API: not only what you read, but what you do, whom you speak with, what you buy.

Third: Data protection is an ornament. Recall KVKK's "80s trend" warning: while the authority lectures the citizen on a photo filter, the state's own databases were being sold on query panels. In the period of the personal agent this picture grows heavier: email, calendar, payment, health app, smart home. An agent's access rights are the map of a person's life. In Turkey, which institution, under which law, under which judicial review, will protect this map? Today there is no answer. As the absence of data is itself a datum, the absence of a rule is itself a rule: if there is no rule, the owner lays down the rule.

Official Narrative and Class Reading

Official narrativeClass reading
"Any lab that isn't focused on alignment will fall behind."Safety is product differentiation; the market does not punish harm, it prices it. Meta's own record is the proof.
"We didn't ask everyone else to do the same; we did it ourselves.""No" not to slowing down, but to sharing the slowing-down decision. The brake under a single foot: 14 percent of the shares, 61 percent of the votes.
"We delayed Muse for months for safety."The product was not working; an agent reaching into iCloud photos without leave. An engineering necessity turned, after the fact, into a virtue.
"Superintelligence should be everyone's; there is no single, benevolent superintelligence."Access to everyone, property to us. The digital form of formal freedom.
"Open source prevents centralisation."The strategy of commodifying the complement. Open when behind, closed when ahead: Glimmer open, Spark closed.
"Contributor pricing: share your data, take a 95 percent discount."The price list of the theft of the general intellect. The return to labour is not a wage but a company-store token.
"Muse, a personal agent that works on your behalf."The enclosure of intention. The agent that bargains is the market's doorkeeper.
"AI will bring more employment."The same year, 8,000 workers at four in the morning; the savings to the data centre.
"A dynamic auction guarantees the lowest price."Intelligence becoming a commodity with a spot price; priority to ability to pay.
"Any policy that slows is a risk to American leadership."Yes to China's being slowed, no to Meta's. Openness ends at the border.
"The metaverse is the future of the internet."Reading a dystopia's consolation machine as a business plan. An 88-billion-dollar error of ontology; the bill to the teams laid off.
Not in the text: the worker, the union, the content moderator, the data centre's water.The next section of the piece.

Let us reduce the table to one sentence: Amodei asks "how fast should we go," Zuckerberg says "let the wheel stay with me"; we ask "why is the wheel with you."

What Is Being Covered Over?

There are four groups of people who never appear in Zuckerberg's manifesto, in the 15 September reply, or in the Muse announcement.

Content moderators. The labour that keeps Meta's platforms "safe" is mostly the labour of moderators working through contractors in countries such as Kenya, the Philippines, Colombia; people who watch images of violence, abuse, hate for hours. In Kenya, moderators founded the African Content Moderators Union in 2023; Kenya's Court of Appeal held in September 2024 that Meta was the de facto employer of moderators it worked through a contractor. Behind the words "alignment" and "safety" are these workers' nerves.

Those laid off. 8,000 in May, 600 from the superintelligence lab itself in October 2025. While the manifesto writes "more employment," these figures are not there.

The data centre's neighbours. Meta's 2026 capital expenditure is, according to the sources, between 115 and 145 billion dollars; a large part of this goes to data centres. A single project in Louisiana is 50 billion dollars (Fast Company). In Spain, in Talavera de la Reina, it was calculated that the planned Meta data centre would draw hundreds of millions of litres of water a year, and local people opposed it with the slogan "The Cloud is drying my river"; in the Netherlands, in Zeewolde, Meta's data-centre project was stopped in 2022 by local resistance. "Personal superintelligence" is no one's personal water.

Children. Muse shipped two weeks after a 17-billion-dollar child-harm settlement. The measures the settlement imposed (a default two-hour limit under 18, an access block from midnight to six in the morning, hiding the like count) are the form of harms known for years, accepted years later, under the pressure of a lawsuit. The market did not "sort" these for twenty years.

If you ask the owner about a technology's safety, you learn the owner's safety.

Three Common Errors

First error: "Zuckerberg is for open source, so he is with us." Part of the left and of computing workers look with sympathy on Zuckerberg's openness talk against Amodei's closed-cartel proposal. This is the error that the enemy of my enemy is my friend. Open weights can weaken the cartel's hand today; seeing this is right, and it is valuable that models like Glimmer be used by independent developers, universities, cooperatives. But the owner of the open weights decides alone when to close them, and has decided. If we are to defend openness, we must defend an openness that does not depend on Meta's discretion; with public compute infrastructure, a public licence, public governance.

Second error: "Zuckerberg is a bad man." Explaining Meta by one person's morals is comforting but wrong. Zuckerberg's decisions are not personal; they are positional: the owner of a platform that lives on advertising must look at every new technology as "how does this step between people and the market." Tomorrow someone else in that chair would make the same account. The problem is not Zuckerberg's 61 percent; it is the property order in which 61 percent is possible. Criticise the person, change the structure.

Third error: "The metaverse failed, AI will fail too." Bubble talk is misleading. The metaverse failed, but Meta did not; it moved the money and the people to AI, and dumped the loss on the worker. Muse's first six days show a demand the metaverse did not have in five years. Part of the AI investment will of course fail; but the money of the failed fund will remain, the GPUs, the data and the models will remain with the monopoly. The bursting of a bubble does not abolish the monopoly; it sorts it.

Concrete Tasks

None of these will change a board decision in Menlo Park. But they strengthen your hand at the next launch, the next "personal agent" story, the next slowdown debate.

  1. Ask every AI story the four questions. The template we built in the GPT-6 Astra piece: what is being sold, taken from whom, sold to whom, who pays the cost? Apply it to Muse: what is sold is the intermediation of intention; what is taken is your behavioural data; what is sold is to the advertiser and the seller; the cost is paid by your decisions.

  2. When you hear the word "access," add the word "property." Read the sentence "everyone will have access to superintelligence" together with the question "whose will the superintelligence be." If the second question is not asked, the first sentence is a marketing sentence.

  3. Defend open source, not the owner of open source. Use models like Glimmer, learn, work on them; but do not forget that the company that opens the weight can close it tomorrow. Develop a demand for public, community-governed models and public compute infrastructure. The guarantee of openness is not the licence; it is governance.

  4. Read the permissions you give a personal agent as a contract. Email, calendar, payment, health app, photos: every permission is a map of your life. There is a reason the "free" tier asks for a card. If you use it, use it with the least permission; if you will not, tell those around you why.

  5. See "contributor" prices as a labour question. If a company gives you a 95 percent discount for your prompts, your prompts have a value. If you are a computing worker, ask how company data and your own labour are transferred into the AI tools used at work; bring to the collective-bargaining table an "labour as training data" clause.

  6. Talk about dependence on WhatsApp as an organisational risk. Does your union branch, your association, your student group entrust all its communication to a single monopoly's infrastructure? As we told in the Jitsi piece, the organisation's communication should be in the organisation's own hands. Propose free-software and self-hosted alternatives to your organisation; take on the setup.

  7. Build the third party in the slowdown debate. You are not obliged to choose between Amodei's cartel brake and Zuckerberg's one-man brake. As we proposed in the Brake piece, the "right to stop" is a worker's right: the right to refrain from work in article 13 of Occupational Health and Safety Law no. 6331 should also be thought for the worker inside AI labs. The place of the brake is the foot of the one who stands under the risk.

  8. Do not forget the metaverse; document it. 88 billion dollars, thousands of layoffs, empty virtual squares. At the next promise of "the internet of the future," recall this balance sheet. To keep the record of capital's fictions of the future is to take back the right to speak about the future.

  9. Answer loneliness not with a product but with organisation. In a world where AI friends are sold, gathering around a real table is a political act. Start a reading group, join an association's event, drop by the union hall. The principle in the Event Organizing and Comradeship guide holds here too: to be organised is to take care.

  10. Follow and spread the struggle of content moderators and data workers. The moderators' union in Kenya, the ruling that Meta is the de facto employer, data-centre resistance in Latin America and Europe: these are the real safety front of the "AI safety" debate. Translate into Turkish, share, make a link.

  11. Criticise the person, change the structure. Carry every discussion of Zuckerberg from his personality to the dual-class share structure, the platform monopoly, the advertising business model. To say "Zuckerberg is bad" is not a finding; to say "it should not be possible to rule 61 percent with 14 percent" is a demand.

  12. Demand a rule for personal agents in Turkey. By which law will the data an AI agent can access, the decision it can take, the payment it can make, be limited? To whom will accountability belong? These questions should be on the agenda of the professional chambers, the bars, the unions, before the agents come to Turkey.

Dear Comrades,

To understand what Mark Zuckerberg wants, one must look at the product table more than at his speeches. The table has said the same thing for twenty-two years: friendship, gaze, speech, access, money, space, intention, perception. Each time a layer that stands between people, and between people and the world, is found, fenced, owned. The metaverse was this programme tried on the wrong layer; Muse is it tried on the right layer.

Zuckerberg's "no" in the slowdown debate is the natural result of this programme. Putting the brake on the cartel table means sharing ownership of the layer. Zuckerberg does not want this; because what he wants is not speed, but intermediation itself. Amodei's proposal is only the dividing of this intermediation among a few companies. Both camps ask the same thing: who should own the layer? Neither asks the same thing: why should it have an owner?

Young comrade, when you hear the next promise of "personal superintelligence," the question you will ask is clear: who decides the decision of the machine that decides on my behalf?

What Zuckerberg wants is to stand between us. What we want is that no one stand between us.

Comradely, Knowledge Commons

Reproduce it.

Related Pieces

All our AI pieces: bilgimusterekleri.org/en/tag/yapayzeka/

The slowdown-debate line

Meta and platform monopolies

Philosophical ground and fictions of the future

The commons and the counter-programme

The Turkey line

Sources

Zuckerberg's reply to the slowdown call

Zuckerberg's texts

The Muse family

Meta's structure, investments, record

The metaverse

Philosophy and Marxist frame

Tags:#mark#zuckerberg#markzuckerberg#meta#facebook#technocapitalism#artificialintelligence#agi#muse#pacingthefrontier#generalintellect#philosophy

Related Posts

Newsletter