Software for One

On February 17, David Singleton published a blog post announcing Dreamer, the company he’d spent a year building with Hugo Barra and Nicholas Jitkoff. He described what it was for with a specific example. Every morning, an agent prepared him for his day and delivered it as a personal podcast: calendar, news, last 24 hours of Slack. A simple tool that fit his life exactly, which he built himself, without writing a line of code.

Five weeks later, Meta hired the team and Dreamer is gone. Barra, Jitkoff and Singleton will work at Meta Superintelligence Labs, building for 3.5 billion users.

Dreamer had raised $56 million at a reported $500 million valuation. Its founders are not people who need to be acqui-hired. They had a product in beta and thousands of users. The thing they were building was harder to describe than most startup pitches. Singleton called it a layer where “software should be personal, malleable, and shaped by the person using it”. Not an app, not an agent, but something that sat underneath everything else and got smarter about you over time.

That a company with that specific thesis just landed inside the one platform best positioned to build a unified software layer is a signal worth paying attention to.

For years, Western companies have tried and failed to build super apps. Facebook launched Messenger bots in 2016. Snapchat introduced Minis in 2020. Amazon tried to become a default commerce layer. Apple built App Clips. None of it worked.

In contrast, by 2017, WeChat had become the dominant interface to life in China — not just messaging, but payments, news, government services, travel bookings, medical appointments. More than a billion users conducted much of their daily lives through a single application.

The consensus explanation was culture: Americans prefer best-in-class apps for each task, but that explanation is incomplete. WeChat dominated because it solved a problem that didn’t exist in the West to the same degree. When WeChat launched, China had weak app store infrastructure and fragmented payment rails. Google Play was blocked, which meant Android in China was a patchwork of competing stores with no single distribution layer. Mini programs gave developers a distribution advantage that justified building there

In the US the App Store had solved distribution. Stripe and Visa had solved payments. And the big Western social apps had open APIs. Pre-Cambridge Analytica, apps like Spotify and Tinder built their growth loops on Facebook's social graph. Western apps could build on top of each other. There was no reason for developers to build inside a container rather than ship their own app, and no reason for users to route through a wrapper.

Once habits are formed, displacing them requires something that is 10x better.

OpenAI thought AI might be that. Instant Checkout launched in September 2025, letting users buy products directly inside ChatGPT. Walmart, Etsy, and Shopify signed on. Walmart’s conversion rates inside ChatGPT came in at roughly a third of its native channels. Users who discovered products through ChatGPT consistently bounced back to the retailer’s own site to finish the transaction.

In March, OpenAI retreated to being a discovery and referral layer and now the checkout happens elsewhere.

The through-line across all of these attempts is that the software itself never changed. Each app tried to become an aggregator — more things than before, organized in one place — but once deployed, it was fixed. It couldn’t notice you’d stopped using a feature. It couldn’t infer that what you actually needed was a workflow nobody had designed yet. It could surface things it thought you’d want, but it couldn’t change how it worked based on what it learned.

What Singleton was building isn’t a container for existing software. It’s a layer that adapts to a specific person, accumulates understanding over time, and uses that understanding to make everything it touches more useful.

This kind of software had never been economically viable before. Anish Acharya is an a16z partner who led the pre-seed of Wabi, a platform where anyone can describe and build their own personal apps without code. He described building a math game to help his kids earn screen time, and a mini Instagram for family photos of the cat. “These aren’t products or startups,” he wrote. “They’re not made to scale or monetize. They’re just little apps custom-fit for a specific audience, useful in a narrow context, and fun to build.” You wouldn’t write a thousand lines of code for an audience of three. Now you would. Now you could.

Like Singleton’s morning briefing agent, these apps weren’t going to scale. They weren’t meant to.

Not everyone is excited about this shift.

Last week, Apple blocked updates to Replit and Vibecode on the App Store. The rule it referred to isn’t new: apps can’t execute code that changes their own functionality after review. Apple maintained it was routine enforcement. Developers and analysts said Apple was protecting a revenue stream and maintaining its position as the distribution platform. Both things can be true at the same time. And there’s also a third thing: an app that evolves unpredictably after review is harder to hold to any quality or security standard, and Apple has always made the consistency of that standard central to its brand.

Motivation aside, Apple is blocking runtime autonomy — anything that executes arbitrary logic and evolves its behavior per user. An app that learns you specifically and reshapes itself around that understanding starts to look like an unbounded platform that violates Apple’s rules. The App Store was designed around a specific picture of software: fixed, reviewed, the same for everyone who downloads it.

Adaptive software, software that might never look the same twice, has no place in that model.

But the constraint only applies to software running on Apple’s operating system. Nine million people already wear the Ray-Ban glasses. Meta’s own devices run Meta’s own runtime. The platform problem that blocks adaptive software on iOS doesn’t exist there.

What does this mean for bringing Dreamer’s original vision to life?

Meta has tried building super apps before, more than once. Facebook Home tried to make Facebook the default interface to your phone. Facebook Credits tried to become the payment layer across apps. Portal tried to own the living room. Messenger tried to become a platform. Distribution was never the problem. Meta has five platforms with over a billion users each. What every prior attempt lacked was software that could adapt. They could only be what someone had designed them to be.

Alexandr Wang’s Superintelligence Labs has been absorbing AI teams in quick succession. Manus, Moltbook, and now Dreamer. Wang’s framing for all of it is agents: “truly personalized and always-on, with the ability to integrate across surfaces and wearables.”

Singleton’s, when he confirmed the deal, was broader. He and Zuckerberg, he wrote, “share the same vision of the future: one where billions of people have the power to create software that makes their lives better.” This goes beyond better agents and is a description of what Dreamer was actually building: software that reshapes itself around the person using it.

Whether Meta pursues the narrower vision of personal agents or the broader one of truly personal software is the question the acquisition leaves open.

Every super app attempt has been a bet that a unified layer could be more valuable than the sum of its parts. Every static attempt lost that bet because the software couldn’t compound. It could only accumulate. Adaptive software is the first proposal with a mechanism for something different: a layer that doesn’t just get better at predicting what you want, but changes what it is based on how you use it.

Meta’s acquisition of Dreamer is an acknowledgment that the next attempt could be built on that premise. Whether it works is still an open question. But for the first time, the people trying to build a unified layer are taking adaptive software seriously as the mechanism for doing it.