On Monday, 10 August 2026, Mark Zuckerberg published a document titled The Future is for Everyone, and Meta simultaneously released a new open-weight model. We are writing about it today.

Errors or inaccuracies are possible. Spotted something that looks wrong? Write to me and I will correct it.

Why This Is Not Just Another Announcement

Documents like this appear constantly, and most of them are marketing. What makes this one interesting is that it does not try to argue that Meta's technology is better. It tries to argue that the whole question has been framed wrongly.

Public debate about AI generally runs along one axis: how dangerous, and how fast. Zuckerberg proposes an entirely different one: who holds it.

And the document is long and built as a structured argument - thirteen chapters touching employment, communities, national security, chip exports, model alignment, recursive self-improvement, and Meta's own governance structure. Most coverage focused on one sentence from it.

This piece is in four parts: a section-by-section reading of the document, what actually shipped, what it costs and who pays, and what it changes for an individual. Our own interpretation is kept to the end and marked separately.

---

Part One: The Document, Chapter by Chapter

1. The Three Principles

Individual empowerment	is the source of prosperity
Invention	is the primary purpose of superintelligence
Balance of power	is the foundation of safety

From which follow the two questions he defines as the defining ones of the age: who will have access to superintelligence, and what we will direct it towards.

2. The Critique of Doom Discourse

He expresses puzzlement that so much discourse from AI developers is filled with doom, and makes an argument from consistency: anyone who believes AI will eliminate most jobs and much of humanity's relevance - why are they rushing to build exactly that future.

"The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic"

And historically: hoping an absolute power will benevolently provide for humanity, if only sufficiently enlightened, has not previously led to safe or positive outcomes.

The values he names as guiding: liberty, open inquiry, free enterprise and equal opportunity.

And his chosen examples of invention are not incidental: the brothers in the bicycle shop who believed people could fly, the bookbinder's apprentice with no schooling who worked out how to generate electricity, and the kid in a garage who thought personal computers could be for everyone. All individuals outside established institutions.

3. Invention, Not Automation

Early AI could answer questions and do routine work. The claim is that superintelligence's great contribution will not be there, but in discovering new knowledge - from developing drugs to finding ways to improve a business.

And the reasoning is sharp: the number of questions a person can ask in a day is finite. The number of valuable things superintelligence can invent to advance their goals is not.

4. Six Concrete Goals

1. A personal agent for everyone. Working around the clock on relationships, health, career, finances, home management and hobbies. With a fully private mode in which even Meta cannot see the information or grant access to it - compared to how WhatsApp encryption is built. Accessible through any device, including glasses.

2. Creation tools. His eight-year-old daughter, he says, codes her ideas and produces videos in a single evening. Meta researchers are generating novel crystal structures suited to augmented reality glasses.

3. Tools to start businesses. Realising ideas without raising money or building teams. And his forecast runs against intuition: not only greater economic growth, but more employment over time.

4. A personal tutor and coach. In his phrasing, with a PhD in every subject and unlimited patience. The social angle: students will get help in areas today available only to those whose parents can pay.

5. Science. Biohub has already shipped open source biological models covering virtual cells and proteins. The goal of curing or preventing all diseases this century is now, he says, expected far sooner.

6. Free or affordable access - and a paid layer too. Free versions for billions, and for those wanting to pay for more compute: a dynamic auction mechanism guaranteeing each person the lowest possible price for the intelligence and compute they use, while ensuring capacity goes to whatever is collectively found most valuable.

5. The Claim About Alignment, and This Is the Heart

The prevailing industry view: with enough time, a single model can be "aligned" to be benevolent.

Zuckerberg's claim: that view is fundamentally flawed, because humanity is not a monolith. People hold different values representing contradictory tradeoffs. So no technological solution can align with opposing interests simultaneously. Any single model would have to prioritise some values over others - and would thereby cease to be benevolent to everyone.

The conclusion: there is no such thing as a singular benevolent superintelligence.

The Three Thought Experiments

Scenario	When only one holds it	When everyone holds it
A superintelligent lawyer	An unfair advantage in court, even when wrong on the merits	Justice carried out more fairly and efficiently
Cybersecurity	Almost any system can be broken into; a less secure world	All systems harden and update
A single business	That business outcompetes everyone; a less dynamic market	Everyone gets tools; a more dynamic economy

The structure is identical in all three: exactly the same capability, opposite social outcome - depending entirely on distribution.

6. Where Meta Positions Itself

The commercial sentence inside the philosophical document: Zuckerberg argues Meta is the company primarily focused on building personal superintelligence for everyone, while most other labs focus on AI for companies, governments and institutions. So if they lead, the balance of power tilts to large institutions over the individual.

7. Employment and the Economy

The starting point: there is a natural balance between companies automating for efficiency and individuals gaining skills to serve more advanced needs.

The common fear: automation outruns the growth in people's capabilities. His claim: there is no rule requiring that.

And an economic argument he adds: compute will always be finite, so there will always be an opportunity cost to using it. If AI can invent highly valuable things, it makes more sense to allocate it there than to automating existing jobs.

The new jobs he names: one-person product studios designing custom toys, furniture or clothing; world builders and experience designers; personal biologists formulating tailored treatments.

The historical comparison: before the industrial revolution 90% of people were farmers growing food to survive.

And on company size: he expects companies to shrink - but that this means more companies with fewer people each, not fewer jobs.

8. Building Infrastructure With Communities

A mechanism called a Community Compact, covering high-paying local jobs, investment in schools and public services, a commitment that energy prices will not rise, and environmental care. Plus a fund named the Future Is For Everyone Fund.

The measurable example: in Richland Parish, Louisiana, where Meta is building a large data centre, teachers received a $50,000 bonus this year on the back of increased tax revenue.

On workforce: contrary to fears about displacing knowledge workers, there is a shortage of skilled tradespeople. Meta established America's Workforce Academy for free training and guaranteed jobs in the areas where it builds.

On energy and water: the company says it builds its own energy generation wherever it invests, sometimes supplying surplus back to the community. And commits to being water positive by 2030 - restoring 200% of water used in high-stress areas.

9. Cyber, Bio, and an Unusual Proposal

On cybersecurity: the pattern is that defenders must hold the greater balance of power and resources. And widely deployed open source systems have proven more secure, because more people find vulnerabilities and upgrade.

On biological and chemical risk he calls for humility, noting few historical precedents. The ability to synthesise harmful compounds has existed for decades but has rarely become a significant issue - his conjecture being that the financial motivation present in cyberattacks is absent here.

His two recommendations: focus on limiting the physical production and distribution of harmful materials, and accelerate society's ability to develop responses, including streamlining regulatory approval. In the long run, he argues, the answer to risks from scientific discovery is to accelerate the pace, not slow it.

10. Freedom and Preventing Government Tyranny

The tension: we want both individual freedom and a government able to protect us.

The position: in a liberal democracy a person holds all rights initially and agrees to limit some for the common good. Likewise, a person should have access to personal superintelligence, with restrictions only when truly required. Hence the fully private mode in the personal agent.

11. American Leadership

This is the chapter that received the least attention in coverage, and it is among the most explicit in the document.

The opening claim: which nations lead in AI will determine a new geopolitical balance of power, deciding the prevalence of democratic values, prosperity and security.

The three policy levers: speed of building physical infrastructure, avoiding steps that slow American labs, and export controls that slow foreign labs.

On infrastructure: America holds an advantage in silicon design and a disadvantage in how fast it can build energy capacity and physical infrastructure. And the figure he cites: countries like China are bringing more than a gigawatt of nuclear capacity online every other week.

On chip export controls: he says they have succeeded in slowing foreign labs, so continuing them is right.

And on open source, here he surprises: he does not believe restricting access to foreign open source models is an effective solution. The goal should be for American models to be the best in the world - so remove the hurdles that make it hard for them to compete, rather than blocking rivals.

And on distillation - models learning from other models - an explicit position: all AI models are derived from human knowledge, and he considers it important to protect the principle that you can learn from anything you can observe.

12. Alignment and Existential Risk

The structure he proposes: not one centralised superintelligence, but as many people and businesses as possible with different agents aligned to their goals, checking and competing with each other.

And on alignment itself: most labs, he claims, treat alignment as a means of enforcing a centralised set of values. And his example is concrete: one leading model refused to help draft a letter to prospective parents at a school, because it considered standardised testing unethical.

What Meta proposes instead: alignment means the agent shares the person's goals and values, not those of the company that built it. With the commercial reasoning attached: people will not adopt agents they do not trust with sensitive details, and will not trust agents that act against their interests.

13. Control of Self-Improvement

The dilemma: once AI systems can improve themselves, any lab that does not allocate substantial compute to recursive self-improvement falls behind by definition.

His proposed solution is not to stop, but to balance: Meta, other frontier labs and clouds should collectively build enough compute to allocate some to self-improvement and stay competitive, while still committing the significant majority to people's goals.

And he concedes the limit: any AI engaged in self-improvement is by definition advancing its own goals. But he argues a system directing its own goals is not inherently harmful in itself, as long as a balance of power favouring people is maintained.

14. Four Policy Conclusions

1. What Meta will do. Train leading models then maximise availability: easy to use, free or as affordable as possible, sufficient compute, wide distribution. Plus a fully private mode, and alignment defined as helping each person achieve their own goals.

2. On open source. Strong support, and a claim that restricting the ecosystem would be a mistake. And a statement that now that Meta Superintelligence Labs are running, it will resume releasing open source models soon.

3. On governance, and this is the most concrete commitment in the document.

4. On government policy. Close cooperation between labs and government; accelerating infrastructure - energy, data centres and silicon; caution about slowing American competitiveness; and scepticism towards any proposal leading to centralised superintelligence.

The document is signed with his first name.

---

Part Two: What Actually Shipped

Muse Glimmer is an open-weight model. That term is worth getting right, because it is not the same as "open source": open weights means the trained parameters are published and downloadable, and can be run and modified without permission - but training data and process are not necessarily published.

And the significant technical point: the model is smaller than rivals' leading models, and was designed to run agentic tasks on a Mac or PC with a single graphics card.

Alongside it, Muse Spark 1.2 was offered, a more powerful model, with developer access.

---

Part Three: What It Costs, and Who Pays

What the Last Report Showed

In the second quarter of 2026, which we wrote about here:

	The quarter	A year ago
Revenue	$60.80 billion	$47.52 billion	+28%
Expenses	$42.03 billion	$27.08 billion	+55%
Operating income	$18.78 billion	$20.44 billion	-8%
Net income	$15.85 billion	$18.34 billion	-14%

The advertising engine has never been stronger. And profit fell.

EPS came in at $6.18 against consensus of $7.23, and operating margin fell from 43% to 31%.

Capital expenditure in the quarter: $31.08 billion, and the floor of the annual outlook was raised to $130 to $145 billion.

And Here Is the Point Easiest to Get Wrong

And How the Market Responded

The stock closed Monday at $594.92, up 0.48% from $592.10 on Friday.

Almost nothing. A long strategic document, a model release and a new product line - and the price did not move.

---

Part Four: What It Means for the Person at the Edge

If the thesis holds, the practical conclusion for an individual is not "learn AI" in general. It is more specific:

1. The difference will be between the person running it and the person being run. The same model in two different hands produces entirely different results - because what matters is which question gets asked and what is done with the answer. The expensive skill will be framing the problem, not operating the tool.

2. The layer that erodes first is the middle. Not the expert and not the beginner - but whoever's job was to move information from one point to another without adding judgement.

3. Individual advantage comes from attaching to a place, not to a tool. General professional knowledge becomes available to everyone at once. What does not is specific context: customers who know you, understanding of a local market, data only you hold, and accountability someone will sign their name to.

4. And financially, not as a recommendation: if this vision materialises, the money does not necessarily flow to whoever builds the model. It flows to whoever holds the bottleneck the model consumes - power, chips, cooling and data centres. Exactly the pattern we saw this week at Riot and Bitdeer.

---

אא

הזווית שלי

דעה אישית של אילן אברמוב - לא ייעוץ ולא המלצה

I read this document holding two thoughts that do not sit together, and I think both are true.

First, it is a serious document. I say that because it is easy to dismiss as marketing, and it is not. It builds an argument, defines terms, cites numbers, and admits explicitly in two places that there are things it does not know - on biological risk, and on whether a system directing its own goals can be relied upon. Corporate vision documents do not usually admit anything.

And the central argument is strong. The idea that there is no such thing as a single benevolent superintelligence, because humanity has no uniform values, moves the question from "how smart is the model" to "who decides what smart means".

Two statements stayed with me. The first is that the most dangerous scenario is a lab training a powerful model and keeping it to itself - a complete inversion of the common intuition, which also happens to describe his competitors precisely. The second is the example of the model that refused to draft a letter to parents over standardised testing. A small example, which is why it works.

And now what bothers me.

The big gap is between the weights and the gigawatts. Zuckerberg argues, correctly, that concentration of power is the danger. But the infrastructure producing that power is more concentrated than ever - and this comes out of his own document: compute is finite, every use has an opportunity cost, and you need to build "enough compute" just to compete. When a free model runs on infrastructure costing hundreds of billions, the balance of power has not moved to the individual. It has moved one layer up.

The second point is that philosophy and interest point the same way. And not in the simple sense usually attributed. Meta does not "only sell advertising" - it is building its own meter on compute and models. So it is not conceding its rivals' market, it is entering it from below: erasing their price and installing its own pricing mechanism in the same motion. That is a more aggressive move than "giving it away", not a softer one.

Which does not make the argument wrong - sometimes interest and truth coincide. But it does mean reading more slowly, particularly when the document presents free as a principle and payment as a footnote.

And another tension I hold: the American leadership chapter asks for two things that are hard to hold together. On one hand maximal openness and no restriction of foreign models. On the other, continued chip export controls that slow those same competitors. There is American logic in that, but it is not quite the principled position presented elsewhere.

And what I think is the genuinely important part, which almost nobody discussed: the governance structure.

Handing an independent board the authority to approve safety criteria for model releases is the only thing in the document that can be enforced. Everything else is a statement of intent. That is measurable: either the board stops a release at some point, or it does not. And if it never does, we will know what it was worth.

And that is also the test I would set for the whole document: whether Meta releases open weights when its model is the best in the world, and not only when it is closing a gap. Releasing a small model that runs on one graphics card is a fine move - and a cheap one. The real test arrives the day the concession hurts.