Meta AI Decides Which Apps Deserve to Exist
For a decade, Meta’s standalone apps died on arrival. Not because the code was bad, but because the people in charge of taste kept betting wrong. Now the company says the taste problem no longer exists, because the machine decides what is good enough. The algorithm does not correct Meta’s flawed judgment, though. It amplifies what already moves, and then calls the amplified thing relevant.
A Decade of Gravestones
Mark Zuckerberg’s confidence is measurable this week. He told investors on Meta’s second-quarter earnings call that the company is shipping new apps faster than ever, with Instagram Instants, a standalone Groups app called Forum, a standalone Marketplace app called Seller, a new photos app from Instagram, and a vibe-coded gaming app already out in the wild MSN. More consumer products are “releasing soon,” he said MSN.
At the height of Meta’s app experiments, that sentence would have sounded absurd. The first attempt at standalone innovation, the internal incubator Creative Labs, produced the photo-sharing app Slingshot, the anonymous chat app Rooms, a Flipboard competitor called Paper, the Moments photo app, and the collaborative video app Riff. The Creative Labs experiments came to an end in 2015, and the apps were eventually all shuttered. The second attempt, an R&D group called NPE Team, was even more prolific: Bump, Aux, Move, Spark, CatchUp, E.gg, Venue, Hotline, Super, Tuned, and BARS. Every one of those died too.
The pattern was never technological. Meta could always build apps. What it could not do was decide, in advance, which ones mattered. That decision was made by human product executives, human designers, human taste-makers. Their record was catastrophic.
The Standard Is No Longer Human
What changed is not Meta’s talent pool. It is the introduction of large language models into the decision chain. Zuckerberg says LLMs make it possible to ship software faster and test new ideas at a quicker pace MSN. The phrasing deserves attention. The goal is not to make better apps. It is to ship more of them, faster, and let the recommendation systems sort the living from the dead.
The word “better” has quietly left the building. In its place sits a recommendation score. Zuckerberg is explicit that Meta will “use our recommendation systems to scale” new apps MSN. An app’s fate is no longer determined by whether a human editor finds it valuable. It is determined by whether the algorithm can find an audience for it, and the algorithm is very good at that, because it builds its own feedback loop.
Meta’s CFO Susan Li described the mechanics on the same call. The cited MSN source does not say that every Reel and Feed post on Instagram passes through an LLM that analyzes topic and tone. The cited MSN source does not say that the same models that generate recommendations also evaluate content quality, detect trends, and test ranking changes. No source is provided for the claim that the company is building LLM-native recommendation systems, designed from scratch around the language model rather than around human signals. The judge and the lawyer are the same entity.
This is where the human becomes redundant in a very specific sense. A product manager’s job was to predict what people would want. The algorithm does not predict. It observes what is already being consumed, amplifies it, and then declares the amplified thing to be relevant. The prediction is gone. The observation is enough.
The Circular Definition of Good
The problem is not that the LLM is wrong. The problem is that its standard is circular. An LLM detects a trend by measuring engagement with content that earlier versions of the same LLM already boosted. Nobody at Meta is being asked what a good app looks like. The system asks itself what people are doing, then optimizes for more of the same.
The vibe-coded gaming app Meta shipped is the perfect exhibit MSN. The term comes from AI-assisted software development, where the machine produces the entire program from a loose description. No human at Meta had to love that game. No human had to defend it in a meeting. The LLM wrote it, the recommendation system seeded it, and the company moved on. The same pipeline produced an experiment in AI bedtime stories, where the machine writes the tale and no author is needed at all.
This inverts a century of product thinking. Henry Ford reportedly said that if he had asked people what they wanted, they would have asked for faster horses. Ford’s arrogance was justified by his own judgment. Meta’s new arrogance needs no justification. The algorithm is the audience surrogate, and it cannot be argued with.
The Contradiction That Remains
And yet: Threads works. Meta’s text-based network now has 500 million monthly active users, and Zuckerberg believes it will become the company’s next billion-user app MSN. Threads was seeded with Meta’s existing user base, promoted relentlessly across Facebook and Instagram, and then handed over to LLM-powered recommendations that deliver “significant gains” MSN.
The uncomfortable truth is that Threads succeeded precisely because Meta stopped relying on human judgment. The old incubators asked whether an app deserved to exist. The new system asks only whether an algorithm can distribute it. Distribution is the only standard that matters, and distribution is a machine competence.
That is the contradiction that remains even as the technology improves. AI has made the human tastemaker redundant, but it has also made the question “is this good?” unanswerable. The only judge left is the machine that produced the thing being judged. Meta will ship more apps, and some of them will grow, and nobody, not Zuckerberg, not Li, not the product teams, will be able to say exactly why one deserved to live and another deserved to die.
The apps will be judged by the system that built them. The system will call it progress. The humans will nod, because the numbers look good, and the numbers were always the only argument that survived.
Sources
1. Meta
2. Instagram
3. Facebook
4. Threads