I spent most of last Tuesday morning with Mark Zuckerberg's essay "The Future is for Everyone." Six thousand words on superintelligence, published the day after I'd been elbow-deep in this year's layoff data, which made for strange reading — the man whose company cut 8,000 people in May, arguing that AI doesn't have to cost anyone their job.

The easy move here is the cheap shot. Half of my LinkedIn feed took it within hours: Meta is third in the model race, so of course Zuckerberg wants to change what winning means; the essay is a strategy memo in a philosophy costume; next.

I think that's lazy, and I want to do something different. I want to take the essay as seriously as it deserves — because parts of it are genuinely the strongest case anyone has made against AI doom — and then show you the exact sentence where it breaks. Both halves matter. If you only read the takedowns, you'll miss a real idea. If you only read the fans, you'll miss the hole.

Full disclosure before we start, as always: I run a company that builds AI agents. A world of widely distributed agents is good for my business. Zuckerberg's incentives are showing in his essay, and mine are showing in this one. Motives noted, on both sides. Now let's argue the merits, because arguments don't care who benefits from them.

The part he gets right

The core of the essay is a claim about alignment that almost nobody in the AI safety world wants to engage with, and they should.

The standard story goes like this: AI is dangerous, so we must align it — get its values right — and because that's hard and high-stakes, powerful models should be developed carefully by a small number of responsible labs. Safety through concentration, roughly.

Zuckerberg's response: there is no such thing as getting the values right, because humanity doesn't have one set of values. People disagree — genuinely, permanently, about real things. Any single superintelligence would have to pick winners among those disagreements, which means a "benevolent singular superintelligence" is a contradiction in terms. So the safety question isn't technical at all. It's political. And the political answer humanity already discovered is balance of power: lots of actors with competing interests, checking each other, the way democracies check factions and markets check firms.

His thought experiments are good, and I'll retell my favorite. One person with a superintelligent lawyer is a menace — they win even when they're wrong. Everyone with a superintelligent lawyer is something closer to actual justice, because today's courtrooms are decided by resource imbalances as much as by merits. Distribution doesn't just spread the power around; it cancels out the unfairness.

And there's one passage I found braver than it's gotten credit for. He says the most dangerous scenario isn't open models in the wild — it's leading labs training powerful systems and keeping them for themselves, however responsibly they narrate it, because that's the road to a singular unchecked superintelligence. Whatever you think of the conclusion, that's a real inversion of the standard safety story, argued rather than asserted. Thirty years in this industry has taught me to respect people who state a falsifiable position in public.

So: the strongest anti-doom case on record. Genuinely.

Now the hole.

The sentence that breaks

Deep in the essay, carrying the entire weight of the existential-risk argument, is this claim — I'm paraphrasing tightly: if billions of people end up using and scrutinizing personal superintelligence agents, then by definition we will have solved alignment with individual interests.

Read it again. Because that sentence converts adoption into proof of alignment. If billions use it, it must serve them.

The strange thing is that the single best counterexample in business history is the author's own company.

Billions of people adopted News Feed. Used it daily, for years, scrutinized it constantly. Does anyone — including, on the evidence of his own past testimony, Zuckerberg — believe that demonstrated News Feed was aligned with users' interests? The feed was optimized for engagement. People's revealed in-the-moment preferences and their actual interests over time turned out to be different things, and the machine served the first at the expense of the second, at civilization scale, while adoption numbers went up and to the right the entire time.

Adoption proves a product is compelling. It does not prove the product serves you. Every slot machine designer knows the difference. The essay's whole safety architecture — billions of aligned agents checking each other — rests on treating a market test as a safety certificate, and the author has personally run the largest experiment ever conducted showing those are not the same test.

That's the hole. And because the sentence is load-bearing, everything downstream inherits it.

The clearest casualty is the section on self-improving AI. Zuckerberg concedes, flatly, that there's no reason to expect a self-improving system to be benevolent. His answer is a ratio: keep the significant majority of the world's compute directed by people at people's goals, and humanity keeps the balance of power. But three paragraphs earlier, he's told us a self-improving system might wring 100x more intelligence out of each gigawatt. If that's true — his hypothetical, not mine — then a minority of raw compute can hold a majority of effective intelligence, and the ratio guarantees nothing. The essay hands the skeptic the exact number needed to break its own safeguard. I read that section three times looking for the reconciliation. It isn't there.

Smaller version of the same habit: the jobs section says "recent statistics suggest" people's capabilities may grow faster than automation. No citation. Meanwhile the payroll data actually on the table shows employment for 22-to-25-year-olds in AI-exposed jobs shrinking around 4% a year. When an essay this careful goes vague exactly where the evidence runs against it, that tells you where the author knows the soft spot is.

The swipe worth pausing on

One more passage, because it cuts close to what I do for a living. Zuckerberg takes an unnamed shot at a rival lab whose model refused to help a parent draft a school letter because the model disapproved of standardized testing. His conclusion: alignment should mean agents that share your goals and values, not a company's — helping people pursue their own diverse aims rather than enforcing a centralized dogma.

As a criticism of paternalistic refusals, this lands. I build on multiple LLM providers precisely because vendor-imposed values show up in production in ways clients can't predict, and a model lecturing your customer mid-task is a real business problem, not a hypothetical one. Score the point.

But watch the leap. From "models shouldn't impose a lab's politics" he jumps to "alignment means serving the user's goals" — and that quietly deletes the entire category of cases where the user's goal is the problem. The person the fraudster is calling doesn't care whose values the fraudster's agent was aligned with. Balance of power among empowered people is cold comfort to whoever's on the receiving end during the years before the defenders' tools catch up — a gap the essay waves at with "defenders will have more compute" and never actually closes.

Refusing to draft a school letter is a failure. Refusing to run a romance scam is not. A serious theory of alignment needs a principled line between those, and "share the person's goals" doesn't draw one — it erases it.

What I'd take from it

So where does that leave the essay? More right than its critics say, and broken exactly where it most needs to hold.

Zuckerberg is right that concentration is a real danger, right that the safety debate has undervalued distribution, and right that "trust us, we're the responsible ones" has a terrible historical track record as a governance model. Where he's wrong is in treating distribution as sufficient. There are two failure modes on the table — concentrated power abused, and distributed power misused — and a serious philosophy has to price both. This essay prices one, and covers the other with an adoption metric that his own company's history refutes.

And for those of us who run businesses rather than labs, there's a colder reading that has nothing to do with philosophy. This essay announces that the biggest consumer distribution machine on earth intends to give away capable personal agents to billions of people, free. Whatever that does to humanity's balance of power, I can tell you what it does to pricing power for everyone who sells agents — including me. The floor is about to move. If your AI product's value is "access to a capable agent," you have a Meta problem arriving on a schedule you don't control. If the value is what the agent is wired into — your workflows, your data, your operation — you'll be fine. Worth checking which one you are before he ships.

Read the essay yourself — genuinely, all of it, not a thread about it. Then find the adoption-equals-alignment sentence and sit with it for a minute. Tell me: did billions of people using a product ever prove it was on their side? You've been one of the billions. You already know the answer.

Sandeep Bansal is the founder of GoZupees, which builds AI agents on multiple LLM providers — and would benefit from the distributed-agent future this essay promotes, which is exactly why he's trying to read it honestly.

Next issue: The biggest story in the AI jobs data isn't the layoffs at all. It's who trains the seniors of 2032.

Sources worth your time:

  • Mark Zuckerberg, "The Future is for Everyone" (Aug 10, 2026)
  • Payroll research on 22–25-year-old employment in AI-exposed occupations, as reported by Yahoo Finance, July 2026 (verify primary study attribution before publish)
  • Challenger, Gray & Christmas, 2026 Job Cut Reports