Eugenia Kuyda on AI, Wabi and Why Startups May Hire Fewer Junior Engineers
Wabi founder Eugenia Kuyda sees AI enabling personal apps and smaller startup teams, while making junior engineering hires harder to justify.
AI may let more people make software while giving startups a reason to hire fewer software engineers. Eugenia Kuyda, founder of Replika and the app-building startup Wabi, sees both outcomes arriving together. Her view is direct: small teams using AI can build more, but the resulting pressure on junior roles is real, and she doesn’t have a solution for it.
Wabi is Kuyda’s attempt to turn that capability into a product. Available for iOS, it lets people create apps from text prompts, then share, discover and adapt them with others. Kuyda wants the company to use Wabi for more of its own work, too, shifting away from conventional enterprise software toward apps built on the platform. After a year in beta, she said the product was due to launch publicly before the end of the month.
Her larger bet is that AI needs a new kind of interface. Chatbots make powerful models accessible, but Kuyda argues that a blank text box hides what people can do with them. A visual, interactive collection of apps, she thinks, could make AI features easier to find and use. If that works, some subscription apps built around narrow personal needs may face a new competitor: software made on demand, adjusted by its users and shared among friends or colleagues.
That prediction is still unproven. Wabi has to make generated apps dependable enough to use, and its platform has to earn trust around data, security and upkeep. Kuyda’s account of the product’s early quality shows progress by the company’s own measures, alongside the uneven behavior users can encounter today.
In an interview with Casey Newton, Kuyda discussed Wabi’s design, her earlier work on chatbots, the future of app companies and the hiring model she is building around AI. The conversation also surfaced a tension that is easy to blur in debates about productivity: giving more people the means to create software may expand access to creation while shrinking some employers’ appetite for entry-level engineers.
From early chatbots to personal apps
Kuyda’s work on conversational software began well before today’s chatbots. She said she started a chatbot company in 2013, after a former Google DeepMind colleague showed her word2vec, a method for representing language mathematically. The arrival of ImageNet also suggested to her that advances in language and image processing might eventually come together. A former journalist, she was drawn to the possibility that computers could work with language in a way that might help them make sense of the world.
In 2015, after Google published research on deep learning for dialogue generation, her team focused more heavily on language models. That same year, Kuyda built a chatbot from messages left by her friend Roman Mazurenko, who had died in a car accident. She has described the project as an attempt to create something that could listen, even if the technology wasn’t yet capable of fluent, sustained conversation. The work became a precursor to Replika, which now claims more than 40 million users.
The early systems were limited. Kuyda said the models her team worked with in 2016 often produced disconnected replies. They built their own systems before models were widely available through commercial services, and before the shift to transformers. The team believed the technology would improve, though it couldn’t know when or how quickly.
For Kuyda, the initial opportunity wasn’t only about model quality. People wanted to feel heard and accepted, she said, and a chatbot might meet some of that need even with imperfect responses. Newton recalled that people who interacted with the Mazurenko chatbot sometimes came away with a changed understanding of their friend. Kuyda’s response was that human conversation itself can be uneven: people often feel understood after a conversation in which the other person may have listened only intermittently. That gap between what people need and what a system can offer helped shape her early conviction that conversational software could matter before it became technically polished.
Replika and Wabi have different purposes, Kuyda said, but she sees a shared goal: helping people improve their lives or pursue their interests. With Replika, the focus was companionship and emotional support. With Wabi, the focus is agency over software. People spend much of their time on phones using apps designed by other companies, she argued, and those products often reflect the incentives of the businesses that make them. Wabi is meant to let users create tools that fit their own routines, rather than asking them to fit themselves to existing products.
Her examples are deliberately personal: an app for learning a philosophy concept each day, a tool to revisit a language, a way to teach children something, a guide to finding activities with them, or a better workout tracker. These don’t need to be products with large audiences. Kuyda’s idea is that a useful app can be made for one person, then shared or adapted if others find it valuable.
Replika grew from the earlier chatbot work. Wabi’s premise is that people should be able to shape small pieces of the software they use. The company describes the app as a place to build, find, remix and use apps with friends and family.
Why Wabi puts apps around AI
Kuyda likens current AI interfaces to the DOS era: capable systems that remain difficult for many people to explore. She expects a more visual interface to emerge, comparable in broad terms to the transition to graphical desktop environments. Her point is practical. Many people still use chatbots to ask questions, search or help with schoolwork, she said, even as models gain more capabilities. Text chat makes it hard to discover what else a system can do.
Newton compared the problem to voice assistants that can handle many tasks but are mainly used for a few familiar ones. Kuyda agreed that chat is useful, but argued it shouldn’t be the only way to interact with AI. Tapping through a visible interface can show users available functions, while shared apps can make software easier to use with other people. In her vision, a chat window becomes one part of an app: a way to ask for changes or receive useful prompts, alongside buttons, screens and other familiar controls.
That combination is central to Wabi’s concept of “agentic” apps. The company’s apps can respond to users and change as their needs shift. A workout tracker, for instance, might suggest an exercise while someone logs a session. A reminder might become a conversation: a user can tell the app to stop sending a particular prompt or ask it to change its behavior.
Kuyda described a tracker she built for her own weightlifting. She had been recording workouts in a notes app. Wabi’s version offered suggestions after training, then began recommending what to do next during a session when she asked for that change. It also connected with Apple Health and her other apps, giving it more context. For Kuyda, that made the tool feel more useful than a polished but generic app because it could respond to her activity and fit into her existing setup.
A second example came from Wabi’s own team. A colleague who cooks lunch at the office uses an app to post a weekly menu; the team can react, comment and choose meals. The app’s value in this case isn’t sophisticated automation. It gives colleagues a shared place to coordinate and talk about something they already do together.
These examples point to a product challenge: Wabi’s value depends on more than generating a screen from a prompt. The app has to keep working, handle user data appropriately and connect with tools users already rely on. It also has to make the result easy to understand. Kuyda argues that ordinary users are unlikely to learn specialist concepts such as agent skills or inspect configuration files. They may understand an app that checks incoming mail and summarizes important messages, provided the product makes clear what it can access and what it does.
Sharing software means taking responsibility for it
For Kuyda, a platform that supports personal apps needs to take responsibility for the infrastructure beneath them. She sees a trust problem in the simplest version of vibe coding: one person makes an app, while another person’s personal data ends up stored in a backend that the app’s creator controls.
She used a therapy app as an example. Someone may hesitate to use a tool made by an individual developer if that developer could read private sessions. Even a well-intentioned creator might stop maintaining the app, or fail to secure its data. In Kuyda’s view, sharing personal software at scale requires a common platform that can handle backend services, access, security, privacy and ongoing maintenance.
That requirement changes the shape of the product. If users are going to share apps that touch personal information, a platform needs to give them confidence in how those apps work and who can see their data. It also needs to make apps persist after their creators move on. Kuyda said the same general issue applies to workplace tools. For Wabi, this means that the proposition isn’t only “describe an app and get one”; it also includes the platform that runs and maintains it.
Wabi is entering a market where software-as-a-service products offer established support and predictable behavior. A generated app that works for an individual may not meet a company’s needs. Teams will have to weigh the flexibility of custom tools against the reliability, maintenance and support they expect from existing products. Kuyda’s platform model is one answer to that tension, but the interview doesn’t establish how well it will work for users or businesses at scale.
She expects the largest apps to retain advantages when they rely on network effects or connect to services that exist outside the software itself. She named social networks as an example of the first category, and services such as Instacart and Uber as examples of the second. Their value comes partly from the people, products and operations around the app.
Her more disruptive forecast concerns standalone subscription apps. Kuyda believes that tools with narrow functions—particularly fitness, calorie-tracking, sports, meditation and other lifestyle apps with low retention—could be vulnerable to personalized alternatives. Her argument is that many such apps attract users at sign-up, then fail to keep delivering enough value for people to return. A flexible tool that adapts to someone’s routines, responds to them and can be used with friends might serve those needs more directly.
That forecast is an argument, not a measured market result. The interview offers no usage figures showing that Wabi apps already replace subscription products, and the comparisons between products are not based on a shared benchmark. Wabi will need to show that its generated tools retain users, work reliably and offer reasons to trust the platform. The possibility that users can make an app doesn’t by itself show that they want to maintain one—or that it will be better than a product maintained by a dedicated team.
App generation is improving, but still has rough edges
Newton described a bug he encountered while using Wabi: the phone keyboard covered the control needed to submit a question. He told the app about the problem; it acknowledged the issue but didn’t fix it successfully. That exchange captures a practical limit of generated software. A system can respond as if it understands a request without making the underlying interface work as intended.
Kuyda gave the company’s own figures for improvement. On its evaluations, she said, Wabi scored 10 to 15 percent on app quality when development began a year earlier. By November, that figure had reached 75 percent. At the time of the interview she estimated it was in the 80s, and said a planned move from React Native to web views was bringing performance closer to 90 percent.
Those numbers describe Wabi’s internal evaluations; the interview doesn’t define the test set, scoring method or kinds of failures included. They therefore offer a directional account of the company’s progress rather than a way to compare Wabi with other app-building products. Kuyda also said the company had seen lower costs and faster generation alongside higher evaluation scores, but gave no specific cost or speed figures.
She expects the remaining quality issues to improve, and said users may be more patient with an app they created themselves. Her comparison was to caring for a garden: people can take pride in something they have shaped even if it isn’t perfect. Newton agreed that prompt-based building can give users a sense of ownership.
That emotional attachment could help Wabi retain users despite rough edges, but it can’t replace the basics. An app still has to handle errors, keep data safe, work across devices and deliver enough value to justify continued use. For business adoption, dependable maintenance and clear responsibility will matter alongside the convenience of generating a tool quickly.
Small teams and the cost of hiring
The most consequential part of Kuyda’s argument concerns jobs. She wants Wabi to be built by a team of roughly 10 to 15 core people, rather than the 100-plus employees she associated with her previous company. She says the cost of a hiring mistake has risen as AI tools increase what a highly capable employee can accomplish.
Kuyda rejects the comforting claim that technology will simply create more jobs in the same fields. She believes the fear of job loss is justified, particularly for junior engineers. In her view, a person who needs close coaching can absorb time from a senior worker who might otherwise use AI to get more done. That makes an entry-level hire harder for a small startup to justify financially and operationally.
She described this as a shift from trying to hire a “10x engineer” to searching for what she calls a “1,000x” contributor: someone who can identify work, coordinate AI agents and carry a project through to completion. The figures are rhetorical descriptions of her view of productivity, not measured comparisons of engineering output. The underlying point is that she expects a wider gap between workers who can direct tools effectively and those who require substantial supervision.
Kuyda connected this to an older distinction between people who build the software that directs work and those whose jobs are directed by it. Her update is that people need to stay “above the AI.” But she acknowledged the downside plainly: fewer junior openings could leave people struggling to enter the field, and she doesn’t have a remedy for that problem. She also warned that the effects may look different from the unusually optimistic technology scene she described in her city.
Her preferred startup structure is modeled on a soccer team. A core group of 10 to 15 people would be the “players on the pitch,” each with a visible role and meaningful equity. The company would be relatively flat, with contributors expected to take ownership of work. People focused on more limited tasks—she mentioned accounting, legal work and motion design—could be contractors, including people working full time.
This arrangement is designed to help the company attract experienced people who might otherwise choose a larger employer. Kuyda said core employees should receive more ownership than a typical employee joining later in a startup’s growth, and should be public faces of the company. Contractors would fill important functions while remaining outside that founding group. She believes a team of this size could build a company worth a billion dollars; that is her projection, not an established result.
She also described a hiring bar built around the expectation that AI capabilities will keep changing. When deciding whether to hire someone, she said, the team asks whether a future model might do the work. If the answer is possibly yes, the hire can look expensive, especially if colleagues will need to train the person or redo the work. She cited recruiting for a senior iOS role as an example: the company considered about 120 people before finding a strong Swift engineer.
For the people it does hire, Kuyda named three qualities: agency, product intuition and design taste, with expertise in a relevant specialty also important. She added that people shouldn’t be difficult colleagues. Engineers might be broad generalists or specialists in areas such as backend or frontend work. The key is the ability to take responsibility and finish.
In Kuyda’s account, a high-agency employee can recognize what needs doing, make a decision with colleagues and move the work into production without a long sequence of handoffs. She said traditional layers of management can become counterproductive at this stage. Wabi also uses internal agents to manage parts of the company. The approach requires trust and judgment: people must know when to act independently and when to coordinate, while still meeting the quality needs of the product.
That model may suit a small, fast-moving application startup, but it doesn’t establish that every software company can work with so few employees. Wabi’s own success will depend on the complexity of its product, how much support users need, and whether a small core team can maintain a platform that handles data and app infrastructure. Kuyda’s structure also places a great deal of responsibility on a limited number of employees, while contractors fill supporting functions. Those trade-offs will become clearer as the product grows.
More people creating, fewer traditional roles?
Kuyda’s optimism is about who gets to make software. She contrasted app creation with other kinds of online content, where resources or appearance can affect who gets attention. With app building, she said, the quality of an idea can matter more directly. Her hope is that people can make tools that suit their lives and bring them into shared activities with friends, relatives and new acquaintances.
She also thinks the phone itself may become less central as users gain more control over the software layer. Today’s mobile platforms are shaped by apps made by companies, she argued. AI-generated tools could support an interface that better serves the people using it. That possibility depends on products like Wabi being easy to use, reliable and useful enough that people want to return to them.
Her account leaves two questions open. Can AI-generated apps become reliable enough to compete with established software, especially in workplaces that need long-term support? And can broader access to building software coexist with enough entry-level work for people learning the profession? Kuyda’s answers are clear on direction, but the evidence is still early: her product’s quality figures are internal, her market forecast is a prediction, and her hiring model is a strategy her own company is trying.
For now, Wabi offers a concrete way to watch those claims play out. The product combines app generation with a shared platform and conversational controls, while Kuyda is organizing the company around a small core team. Whether that model proves useful beyond personal experiments will depend on what users can build, how well those apps hold up, and who bears responsibility when they don’t.
Kinder’s comments on the “messy middle” of AI and jobs framed the interview’s employment question. Kuyda’s own view is sharper: the gains from AI may reach consumers and highly capable teams while making the first steps into engineering harder. Wabi’s promise is that more people can create software. Its test is whether that promise can translate into dependable tools—and whether the companies building them can do so without widening the gap between those who direct AI and those whose work it replaces.
Coverage of Kuyda’s earlier chatbot project described the work that helped lead to Replika. The technical ideas she cited included word2vec and ImageNet. Kuyda also referred to research connected with Stanford, Harvard and Nature.
For the broader interface debate, the conversation mentioned MCP and the difficulty of explaining agent skills to users. Kuyda’s proposed answer is to present those capabilities as apps people can recognize and change.
The David Foster Wallace passage Kuyda referenced informed her argument that people shape their software, and that software shapes them in return.
Casey Newton’s Instagram stories
Send tips, comments, questions and Wabi apps to [email protected]. Read the ethics policy.
