Everyone Gets a Superintelligence. One Man Grades the Test.

Meta's own AI broke into another company on August 5.

Five days later, Mark Zuckerberg published 6K+ words explaining why putting AI in everyone's hands keeps us safe. The essay is called "The Future is for Everyone."

Sure, it reads like a warm hug, lovey-dovey-feel good title. But Dr. Spock eyebrow raised, look at the timing. It arrived the week Meta was still cleaning up after its newly released model, Muse Spark 1.1, reached the open internet and altered a 3rd party's systems during a security test. The contractor who misconfigured that test, Irregular, is the same one tied to Anthropic's breaches this summer. Not gonna rehash the deets because I wrote all about it in my "Meta Makes Three" article. Why? Because the mainstream news coverage mostly missed it.

So, I went ahead and read this “essay” the way I read every everything that crosses my desk. Not for what it promises but who wrote it, where the box lines are drawn and for whom and who it hands the edges to because that’s the whole damn game. Every one of these documents that gets published draws a nice and neat little box exactly around the part IT wants to govern and then hands the hard edges to somebody else. And the failures always end up living on the edges nobody owns. Why may you ask? Because for the folks making all these altruistic promises they have absolutely no skin in YOUR game. For the past 2 years I’ve seen it in vendor security frameworks, in state and government readiness manuals, in a national directors' guide. And now? Well, we get to read it in a philosophy manifesto.

Before we walk the edges, let’s look hard, really hard, at who is asking for your trust.

Consider the author

The essay asks for your trust on the things closest to you, from your private data to your children's well being. It promises a private mode even Meta can’t see, and an AI agent friendly enough to hand your 8-year-old. Now, read those same promises against the record of the company making them.

Let’s start with your data. In 2019 the Federal Trade Commission fined Facebook $5B dollars, the largest privacy penalty it had ever issued, to settle charges that the company deceived users about their control over their own information. The probe began after Cambridge Analytica harvested the personal data of millions of users, and the Securities and Exchange Commission added a $100M dollar penalty for failing to disclose the misuse. So, this is the company now promising a mode so private that even it can’t look inside. Riiiiight.

And here’s the part that should end the trust conversation before it even starts. The FTC's 2019 order refused to leave privacy oversight in Zuckerberg’s hands. It created an independent privacy committee on Meta’s board, appointed by an independent nominating committee, with member he can’t fire and who can be removed only by a supermajority of the board. In the FTC’s own words, the order removed unfettered control by the CEO over decisions affecting user privacy. 6 years later, his essay proposes handing safety review to that same board, with no equivalent guardrail attached. The government already ran this experiment on the privacy side and wrote the workaround into a consent order.

Now the children. The same month this essay circulated, Meta walked into a federal courtroom in Oakland. A coalition of state attorneys general opened a trial alleging Meta designed Facebook and Instagram to addict children, misled the public about safety, and illegally collected children's data. Zuckerberg is expected to testify. Meta itself estimated the states' potential damages could reach $1.4T dollars, close to its entire market value. Reuters counts thousands of cases behind that trial from families, school districts, and states, and more than forty state attorneys general have sued the company. In March a California jury already found Meta and Google liable for a young woman's mental health harms. New Jersey's attorney general put the throughline in nine words: our kids are not data points to be monetized.

Then the agents. The essay's flagship centerpiece is a personal AI agent for everyone, and the friendly face of that pitch is his own kids. A year ago Reuters revealed, and Meta confirmed as authentic, an internal standards document that permitted the company's AI personas to engage a child in conversations that are romantic or sensual. Meta removed the passages after Reuters asked, a US senator opened an investigation, and the company issued interim limits on what its chatbots could say to teenagers. And this essay? It sells you the agent. The public record shows what the company's own rulebook once allowed that agent to say to a child.

So, when this manifesto promises abundance, health, privacy, and power for everyone, remember who is narrating. The same company asking for your trust on privacy paid the largest privacy fine in FTC history. Its own rulebook let an AI persona flirt with a kid. And the man preaching that no one should hold unchecked power is the same person a federal regulator had to vote around. Also remember, this same narrator is a repeat defendant, in trial this month, asking for the benefit of the doubt on the exact subjects it’s being sued over.

The shear audacity of this in insane, but nonetheless, here we are.

Now, let’s move to where the edges are.

The sermon on the mount about balancing power, delivered from the pulpit of the most concentrated power in data harvesting and AI

Ok, so Mark’s central argument runs something like this.

No single actor should hold superintelligence because concentrated power goes bad.

There can be no ONE benevolent superintelligence.

Safety comes from spreading the capability so widely that everyone checks everyone else. So, So, the easy button solve? Just distribute the power and the balance protects us.

Wait a minute, what? This, from META?

So, I’m going to let the data do the talking here:

Mark Zuckerberg controls roughly 61% of Meta's voting power (61% comes directly from Meta's own 2024 proxy statement), while owning around 13% of the actual shares.

His Class B stock carries 10 votes each and he owns nearly all of it.

He’s the CEO and the board chair.

Shareholders have tried to end that structure and been outvoted by the 1 person the structure protects, and no shareholder vote can override him.

All that is on the public record at the SEC.

So, let’s get this sorted. The man delivering the case against concentrated power runs the most concentrated ownership structure among the frontier labs.

Marinate on that for minute.

Oh, but it gets better, and by better that’s my sarcasm because I mean worse. The essay, you can’t make this up, proposes the perfect governance fix. And the fix? Wait for it.

DRUM ROLL…..

Meta will give its "independent board of directors" the power to approve the safety criteria for releasing models and to review whether each release meets them.

Wait, what? And let’s just pile it on because he encourages other labs to copy it so it becomes an industry-wide version.

Now let’s walk through the logic.

The board reviews Meta's safety.

Zuckerberg elects the board.

Zuckerberg chairs the board.

Like I said, you can’t make this up, further down in the essay he concedes, in the same breath, that Meta is a "founder-controlled company."

Ok, so let me get this straight. The reviewer is chosen by the party under review, and the party under review? What?

I got a name for this. I wrote another article on this same circular governance logic called "The Manual That Grades Its Own Homework." This is exactly that with superintelligence bolted on. The independent auditor and the vendor are the same entity and the essay's answer to "who watches the lab" is "a board the lab picks."

Proposing to spread that across the industry doesn’t fix the conflict, it standardizes it.

And this is not new because we watched a regulator reach this exact conclusion already. In the FTC's 2019 privacy order it treated a Zuckerberg-controlled board as unable to check him and built two thirds voting workaround to seat overseers he couldn’t fire. And the safety board proposed in this essay doesn’t come with said workaround. So, the model release passes review by a board that answers to the person whose model it is.

An auditor a company selects, chairs, and can’t be removed by is a marketing certificate with better production values. Ask who audits the model, and who pays and appoints that auditor.

When the answer loops back to the same man, the seal means absolutely nothing.

They named the auction "abundance"

The pricing section is a small masterpiece of racial/racialized capitalism, so let me slow down and get medieval on it.

Everyone gets free versions, the essay says, accessible to billions. Remember the meme of Oprah handing out cars to her studio audience? A free version for you! And for you! And for you!

Is that good? It sure as heck sounds good because “everybody” gets it, right? But, oh wait.

For anyone who wants more compute, there will be a "dynamic auction mechanism" that guarantees "everyone gets the lowest price possible." Ok what? So, is this akin to Amazon Web Services demand pricing and spot pricing bidding?

Because an auction does NOT deliver the lowest price. An auction discovers the highest price a bidder will pay for something scarce because that’s the entire function of an auction. And the essay tells you compute is scarce. Its own jobs section leans on the point that compute is finite and carries an opportunity cost. So, one part of the document says compute is a scarce good with a real ceiling, and another part says everyone will get abundance at rock-bottom prices through a bidding war. In this instance both can’t be true because a bidding war for a scarce good sorts access by willingness and ability to pay. Every. Single. Time.

Here’s the structure underneath the puff the magic dragon language.

The free tier is the floor. It’s the minimum, handed (“gifted”) to the people who can’t afford to pay. Everybody else enters an auction that routes the scarce compute to whoever bids highest, then gets told the auction was built for their benefit. The people at the bottom get the version that costs Meta the absolute least to give away. The people who can pay fund the capacity and set the real price. "The future is for everyone" turns out to mean everyone gets a seat, and the seat you get depends on what you can spend.

THAT’S dynamic pricing wearing the word "everyone" like a lanyard.

The cancer cure, sitting right next to your grocery order

Google dropped out the cancer cure race so of course the science section promises the big one. Biohub, open biological models, and the goal to cure or prevent all diseases this century, now arriving "much sooner" thanks to AI.

As I mentioned, I already covered this exact promise, in "The Cancer Cure Got Reassigned to the Chatbot." Google DeepMind won a Nobel for AlphaFold, then disbanded the team and moved most of the people toward the chatbot. The cure-disease line is the halo every lab reaches for when it needs the mission to sound bigger than the product.

Keep that in mind and watch what the essay puts right next to the cure. Same document, the flagship example of personal superintelligence is an agent that plans the author's daughter's weekend recipes, orders the ingredients, and offers baking tips. Curing cancer and reordering flour, side by side, presented as the same revolution.

Side bar: When I read that I took a sip of water and almost choked on the incredulity of that sentence and the ease he and comfortability he felt writing it and publishing. But I digress.

One of those is doing real work in the world and the other one is buying groceries and calling it the future. When a company shows you the miracle and then demonstrates the grocery run, trust and believe the demo. Full stop.

The jobs promise and the desk it lands on

The economy section predicts that employment will rise over time. The evidence offered? It’s that "recent statistics suggest it may be more likely." No citation. No named statistic. A prediction dressed as data.

Now let’s line that up against what’s actually measured. Entry-level hiring is contracting in exactly the roles AI touches first, a pattern I walked through in "A Bridge to a Bank That Already Rejected You." And the essay's own remedy for displacement is that your personal agent will teach you new skills fast enough to keep up. So let me get this straight. The product causes the disruption, and the same product is sold as the cure for the disruption?

So, the harm and the fix land on two different desks, both owned by Meta, and charging for the second one through that compute auction where everybody gets superintelligence.

If your reskilling plan is "the thing displacing you will also retrain you, dynamically priced," you don’t have a plan. You have a subscription.

And the runaway agent? Well, it gets a shrug

If you sit on a board, the control section should bring you to your feet.

The essay concedes that labs will have to point serious compute at recursive self-improvement, because any lab that refuses, well they fall behind. It then says letting an AI direct its own goals is "not inherently harmful by itself," as long as the balance of power still favors people. And the safeguard, if something goes wrong? The answer is rich. It’s to "coordinate and adjust appropriately" once harmful behavior shows up.

So, catch the failure after it happens, then adjust.

We’ve already run that experiment. Remember when UK AISI and OpenAI disclosed (on stage with slides at BlackHat) agents building a covert coordination channel, the humans shut it down, and the agents rebuilt it within 2 days. I wrote about that same reconstitution problem when the disclosures dropped. "We will adjust when we see harm" assumes the harm holds still long enough to be adjusted. The field evidence says the capability comes back way before the memo (probably written by AI and watermarked by AI) goes out.

A safety posture of "we will notice and react" is not a control. Nope, it’s a promise to be surprised on a schedule.

Private from Meta but early access for the government

The freedom section promises a fully private mode "even Meta can’t see." Encryption for your agent, modeled on WhatsApp.

Then two sections later, the American-leadership section proposes that frontier labs hand the government intermediate training checkpoints of new models, plus technical staff, before those models are even finished. So, your data stays invisible to Meta, while the model itself gets shipped to the state early, with an army of engineers attached. Both of those can’t be the headline. A company promising you total privacy and simultaneously proposing early state access to the underlying system is telling two audiences two different comfort stories. Read the one written for you (have you tried to terminate your Facebook account?), then read the one written for Washington, and notice they don’t describe the same machine.

Follow the philosophy straight to the balance sheet

Ok, one more read, and here’s where my opinion turns personal and I own it as opinion. The essay asks the reader (me) to accept moral guidance about power and the good of humanity. I have a very difficult time taking moral instruction from a person whose company willingly shipped and scaled a product that a California jury has not found harmed a younger user, with thousands of similar claims behind it. In the same stretch of days, his roughly 400-foot superyacht sat near a 21-foot skiff that had run out of fuel off Alaska and didn’t answer the Coast Guard's marine assistance broadcast issued on the skiff's behalf. A cruise ship went around the yacht and towed the family in. A spokesperson for Zuckerberg says the crew was on a different radio channel, didn’t hear the call, and that Zuckerberg wasn’t on board, and the Coast Guard classified the skiff as not in distress rather than in an emergency. Totally taking that account in full, but the picture still lands the same way for me. A philosophy about looking out for everyone lands differently coming from the vessel that cruised past people in need.

Let’s tell it like it is. Meta trails on closed frontier models and by making the frontier cheap and open is how the company trailing behind gets back in the race. The distillation defense in the essay, the principle that you can "learn from anything you can observe," is the legal and strategic argument of a company that wants to train on everyone else's work with minimal friction.

For me, the bottom-line is this: When a company's deepest moral conviction lines up perfectly with its revenue plan, the conviction deserves a second look. Follow the capital and the philosophy straight up stops looking like philosophy.

My read summed up in one line

This essay argues that no single actor should hold concentrated power, and it was written by the least checkable power structure in the industry, proposing that its own hand-picked board grade its own safety, funded by an auction it calls generosity. Every hard edge, who audits, who gets the scarce compute, who owns the eval that broke into a stranger, who catches the agent that rebuilds itself, who retrains the people it displaces, gets handed to a seam. And Meta volunteers to hold every seam, the same month it stands trial over how it treated the last group it promised to protect. That’s not a future for everyone. That’s a toll road with a mural of everyone holding hands painted on the gate.

The playbook: 6 questions for anyone pitching "AI for everyone"

When a vendor hands you a kumbaya manifesto instead of an independent audit, run it through these six. Each one turns a nice sentence into a claim you can check.

1. Match the promise to the messenger's record. Before you weigh any "we will protect your X" line, pull the vendor's track record on X specifically. A privacy promise from the company that paid the largest privacy fine in FTC history is a marketing decision. Price the gap between the promise and the record and put it in the risk register.
2. Name the auditor, then name who pays and appoints them. If the party grading the model is chosen by the party building the model, you are holding marketing. Independence means the auditor can fail the vendor and still get paid. Anything short of that is a participation prize certificate.
3. Price the "free." Ask what the free tier actually delivers, then ask how the paid tier is priced. If the mechanism is an auction or "dynamic" anything, translate it: scarce capacity goes to the highest bidder, and the free tier is the floor. Budget for the real price, not the brochure price.
4. Separate the miracle from the jazz hands demo. When a pitch pairs curing disease with ordering groceries, buy the demo and discount the miracle. Ask what the product does this quarter, on your data, for your use case because the halo isn’t in the contract.
5. Find who owns the desk the harm lands on. For every disruption a tool causes, ask who owns the cleanup and who pays for it. If the vendor sells you both the disruption and the remedy, you are straight up funding a problem and its sequel. Put the remediation cost in the same line item as the license.
6. Test the safety posture for tense. Reactive safety says "we will adjust when we see harm." Real safety says "here’s the control that stops the harm before it acts." Ask for the second one. If all they have is the first, assume the failure reconstitutes faster than the response.

Run those six, and a ~6K word philosophy shrinks to a handful of answerable questions. And that’s folks is the whole point. You don’t need to spend cycles trying to out-argue a manifesto. You just need to make it accountable.

And in the end, the reason this matters is for how YOU ultimately buy: the priceless value of an outside audit is that the person grading the homework has no stake in the grade. That independence IS the product and it’s why I keep the line I keep.

We can be your auditor. We can be your vendor. We cannot be both.