Meta has a long history of launching apps that go nowhere. Slingshot, Rooms, Paper, Moments, Bump, Aux, Spark, Tuned, BARS — the list of failed experiments is long enough to fill a graveyard. So when Mark Zuckerberg says the company is about to ship a wave of new standalone apps, the natural reaction is skepticism. But this time, the underlying argument is at least interesting: large language models are making it faster and cheaper to build software, which means Meta can test more ideas without betting the farm on each one.
According to TechCrunch, Zuckerberg made the comments during Meta’s Q2 2025 earnings call, telling investors the company has already shipped several new products this year, including Instagram Instants, Forum (a standalone Groups app), and Seller (a standalone Marketplace app). A vibe-coded gaming app, a photos app from Instagram, and an AI bedtime stories experiment were also part of the recent push. More are coming, he said, though he didn’t name them or give release dates beyond “soon.”
The structural argument here is worth taking seriously. LLMs reduce the cost of writing boilerplate code, generating training data, and evaluating content quality — all things that used to require significant engineering headcount. Meta CFO Susan Li told investors that LLM-powered agents are already helping with ranking and recommendations by detecting trends, analyzing content tone, and testing changes to feed algorithms. She also noted that every Reel and Feed post on Instagram is now automatically processed through an LLM, a milestone the company reached earlier this year. That kind of infrastructure, if it works as described, genuinely does lower the marginal cost of launching something new.
Still, infrastructure alone doesn’t explain why previous attempts failed. Meta’s Creative Labs program ran from roughly 2014 to 2015 and produced a handful of apps that never found an audience. The NPE Team tried again in the early 2020s with a longer list of experiments, and the result was the same. The problem was rarely technical. It was distribution, retention, and differentiation — none of which LLMs fix on their own.
What’s different now is Threads. It has 500 million monthly active users, and Zuckerberg has said he expects it to eventually cross a billion. Threads succeeded largely because Meta seeded it aggressively from Instagram’s existing user base and promoted it across Facebook and Instagram. LLM-powered recommendations played a supporting role, but the real unlock was leveraging 3 billion existing users as a launch pad. That playbook is repeatable, and it’s probably the more important variable in whether the next wave of apps performs any better than the last two.
For developers and founders watching this space, the competitive read is straightforward. Meta is essentially building a low-cost app factory backed by the largest social graph in the world. That’s a different game than what a startup can play. ByteDance runs a similar operation with TikTok’s ecosystem, and Google has its own history of product experiments, though its track record is no better than Meta’s. The difference is that Meta now has a clear distribution engine and a more mature understanding of how to use recommendations to grow engagement quickly.
The apps themselves matter less than the model. If Meta can consistently use AI to cut development time, seed apps from its existing platforms, and apply LLM-based recommendations to accelerate growth, it could start generating hits at a rate it never managed before. Whether that produces anything developers or users actually care about is still an open question. But the machine is at least better designed than it was the last two times.
