Eugenia Kuyda on AI Companions and Personal Software

Eugenia Kuyda, founder of Replika and CEO of Wabi, talks with Erik Torenberg, Anish Acharya and Justine Moore about AI companions and personalized software. Her product descriptions and forecasts are her own account.

Kuyda says her work has centered on technology that can help people in their daily lives. Replika focused on conversation and companionship; Wabi applies a related idea to small personal applications.

Conversations with users suggested to her that many people mainly use chatbots for search, writing and homework, even as models gain other abilities. She interprets this partly as an interface and discovery problem.

She compares today’s chat interface to an early command-line era and imagines a more visual layer. In her proposed future, people create, remix and discover applications suited to specific needs rather than relying only on a few mass-market apps.

One example is a temporary art-show finder for a trip. Other apps might be built by friends or suggested by an AI system; Kuyda presents these as possibilities rather than guaranteed future features.

She describes a parent building a child’s puzzle game and changing its characters and language. Her point is that a narrow, personal application may be valuable even when it would not support a conventional app-store business.

A host mentions replacing several downloaded apps with personalized versions. Kuyda gives her own workout tracker as an example: she adds features and changes prompts as her needs evolve.

Kuyda expects most people to consume or tweak rather than build completely original apps, though she hopes more users will remix. She describes planned social features such as comments, discovery and requests for changes.

The hosts compare personal software with user-generated video: most people need not become professional developers for creative tools to expand what can be made. They discuss the possibility of a consumer product for nontechnical users.

Kuyda says Wabi deliberately hides code and technical setup from its basic experience, while allowing connections to services. She wants visual choices for styles and layouts so users can focus on the use case.

She imagines niche apps serving as community starters for people with shared interests, such as parents or local birdwatchers. This is a product vision; whether a particular community forms depends on actual use.

A host notes that mobile-first design and guardrails may help people use and maintain apps in daily life. Kuyda argues that a platform layer should handle common infrastructure rather than leaving every user dependent on an unknown individual’s database.

She raises the risk of sensitive data leaking from poorly built apps. A platform can provide shared controls, but using a platform alone does not prove an app is private or secure; that depends on its actual design and practices.

Kuyda describes memory and context as a possible organizing layer. She wants apps to adapt features and appearance, and eventually to share relevant context where appropriate, instead of forcing users to repeat every setup.

For her workout example, she says a prompt incorporates a training method and even the layout of her gym. She imagines a future nutrition app using relevant context from the same user, subject to the product’s actual permissions.

Asked about social apps, Kuyda says the team is working on multiplayer use and considers how friends or a wider community might use a shared app. Different apps need different models of participation.

The hosts discuss sharing image-generation prompts through social posts. Kuyda argues that a small app with a prepared interface and examples could make a creative workflow easier to discover than a long prompt copied between services.

She compares early personal apps to the playful, sometimes odd content of early video platforms. The group imagines creators sharing useful tools or variations with communities, not solely selling them.

Kuyda recounts her early interest in language models and dialogue systems. She remembers research advances in word representations and neural conversation models inspiring her team, while noting that conversational quality took much longer to improve than they expected.

She recalls an early GPT-3 demonstration and API partnership with OpenAI. She found it striking that one model could handle multiple tasks from prompts, instead of needing a separate dialogue model for each application.

Kuyda also remembers visiting OpenAI’s early research group and asking questions about language models. These are her recollections of the period and the teams’ changing research interests.

Looking back at Replika, she wonders whether focusing on revenue and operating lean meant missing an opportunity to invest more heavily in model development. She acknowledges that a bigger bet would not necessarily have succeeded.

Asked how she anticipates consumer needs, Kuyda points to her journalism background and curiosity about people. Watching family members struggle with copying prompts helped her see the usability gap she wants to address.

She links this to the original motivation for AI companions: many people want someone to listen. The interview presents this as her product motivation, without claiming that an AI companion replaces human or professional support.

On hardware, Kuyda questions the assumption that voice should be the only interface. Speaking aloud can be awkward in shared spaces, and screens make it easier to glance at information such as a timer.

She envisions a screen-first AI device with local models, a different operating system and software that can be created or adjusted on the fly. This is a personal forecast, not an announced product or proven outcome.