This video features AI pioneer Eugenia Kuyda, CEO of Wabi, alongside Erik, Anish, and Justin, discussing how personal software will transform from a developer-monopolized structure to a medium anyone can create with. She explains why command-line AI interfaces are the new MS-DOS and how mini apps can be shared like TikTok videos. She also details her journey spanning over a decade, from training language models in 2012 to building a platform where even a mom can create custom apps in minutes, along with behind-the-scenes stories from the early days of OpenAI and what voice-only devices are missing.

1. The Evolution of AI Software: From Too-Early Beginnings to Personalized Apps

Eugenia Kuyda reflects on her career, noting that on the journey toward personal software like Wabi, she always had a tendency to start too early. She has been immersed in AI research since 2012, captivated by the idea of having meaningful conversations with machines that could positively impact our lives. Initially, she focused on AI companions like Replika, helping users live better lives and feel good. But now she's applying the same idea to personal software -- building mini apps or software that help us live better lives in personal ways throughout the day.

While running Replika, she observed how users were using AI products like ChatGPT, Gemini, and Claude, and was surprised that most people were using them only for very simple purposes like search, writing, or homework help. Despite the models' incredible potential, people weren't actually utilizing them.

"When people look at chatbots, they seem to think of them as just a search tool, a writing tool, maybe a conversation partner. They're not properly utilizing the amazing capabilities of the models."

Seeing this, Eugenia felt that the "interface problem" was clear. Current chatbots are like computers in the MS-DOS era, and soon an innovative interface moment comparable to Windows or Mac OS will come to AI. Although current chatbots have nearly a billion users, usage patterns remain at a simple level.

2. The Future of Apps: From Developer Monopoly to the User Creation Era

Eugenia explains that the future software world will be completely different from today. Just as we moved from a few TV channels to a world overflowing with YouTube, Reels, and TikTok -- the world of user-generated content (UGC) -- apps will shift from being made by a few professional developers to an era where we all create apps for each other, and AI suggests apps for us.

"Right now we're trapped in apps made by a small number of professional developers, but eventually we'll move to a new world where we all create apps for each other and AI suggests apps for us."

She imagines opening a future operating system to see not just popular apps, but cool apps made by friends, apps modified for yourself, or apps suggested by AI. For example, AI could create a custom app for an art lover traveling to New York that finds art exhibitions near their Airbnb. This represents the emergence of deeply flexible and personalized software -- like an operating system built on the platform of "me" rather than a fixed environment.

3. The Rise of Ephemeral Software: Building Apps Just for You

Until now, software has been bound by the assumption that it needs to be durable because development is expensive. But Wabi overturns this thinking by demonstrating the possibilities of "ephemeral software." Eugenia shares examples of Wabi users creating very small, personalized, niche-market apps that could never exist in an app store.

• Motivational quotes app: An app that pulls quotes only from a specific show and displays them at 5:30 AM.
• Children's puzzle game: A puzzle game with Elsa and Jasmine princess themes, created in 2 minutes for her daughter. She even switched the language to Italian for her daughter attending an Italian kindergarten, turning it into a learning tool.

"Apps that could never exist in an app store -- too small and personalized -- are being created. I made a puzzle game for my daughter in 2 minutes and switched it to Italian as a learning tool based on our needs."

These apps are hard to find in existing app stores and come with issues like complex onboarding, paid subscriptions, and insufficient personalization. Wabi lets users create and modify what they want on the spot, delivering personalized experiences that existing apps couldn't provide. Erik shared that thanks to Wabi, he deleted numerous paid apps for migraine tracking, restaurant recommendations, hyper-personalized notes, and specific-style image conversion, making better versions himself in Wabi.

Eugenia shares her experience creating an app to track her beginner weightlifting. Existing app store apps had too many unnecessary features, so she built a simple workout tracking app reflecting the exercise methods from a book she was reading and her gym's environment. She kept adding and modifying features each time she visited the gym, developing it into her own custom app.

"Existing apps had way too many unnecessary features. I built a simple workout tracking app reflecting my book's exercise methods and my gym environment, adding and modifying features each gym visit to develop my own custom app."

4. The Social Function of Apps: Mini Apps as Community Catalysts

Wabi emphasizes the social function of apps beyond individual app creation. Eugenia estimates that fewer than 10% of all users will build apps "from scratch," with most people tweaking existing apps. Wabi will soon update its social graph feature, allowing friends to share which mini apps they download, how they use them, leave comments, and give "likes."

"Over 90% of people will 'tweak' existing apps. Soon through a social graph update, friends will share which mini apps they use and how, and they'll even be able to request modifications from each other."

This enables users not only to create apps but also to discover new apps, use apps with friends, and request feature modifications from creators. Anish draws a YouTube metaphor, noting that just as we moved from a few cable channels to countless YouTube channels where people directly create and share content, a similar transformation will occur in the software space.

Currently, app stores lack social features due to privacy concerns, but Wabi's mini apps can serve as catalysts for community formation. For example, mothers whose children attend Italian kindergartens could create, share, and form communities around related apps, or birdwatchers in a specific neighborhood could gather through apps.

Justin evaluates Wabi's product design as effectively reflecting the "vibe coding" or AI-powered building trend. Wabi provides guard rails that let users easily create and modify apps, enabling people without technical knowledge to unleash their creativity without worrying about mistakes.

5. Wabi: A Framework for Memory, Context, and Expression

Eugenia explains that Wabi is not merely a collection of apps but a framework for memory, context, and expression. Every time users create and share apps, Wabi learns and understands who that user is.

"Wabi isn't just a collection of apps -- it's a framework for memory, context, and expression. Every time I create and share an app, Wabi learns and understands who I am."

This aligns with Andreessen Horowitz (a16z)'s "Software 3.0" concept, which refers to deeply personalized software. Wabi mini apps achieve deep personalization through the user's prompt, going beyond mere functionality and appearance. Like Eugenia's workout app, adding specific exercise methods or gym environments to the prompt allows AI to understand that context and generate workout routines.

Additionally, the Wabi platform remembers and shares personal information like the user's age, location, whether they have children, and health goals. For example, a workout app and a nutrition app can communicate with each other, sharing user context to provide a more integrated and personalized experience. This solves the inefficiency of existing "walled gardens" where apps cannot share information and users must repeatedly grant consent and integrate with each developer.

6. The Era of Prompt Sharing: Enhancing AI Discoverability

Justin points out that current "prompt sharing" is happening in a very inefficient manner. For example, popular AI image-generation prompts on TikTok get listed in lengthy comment threads, and users often struggle not knowing which app and model to use. This inefficiency hinders AI discoverability and reduces user motivation.

"Current AI prompt sharing is too inefficient. Long prompts listed in TikTok comments, or users not knowing which app and model to use. This inefficiency hinders AI discoverability and reduces motivation."

Wabi solves this problem by making prompt sharing much easier and more intuitive. Clicking a link in TikTok comments opens a fully configured mini app where users can just add photos, try various styles, and see others' creations. This represents a transition from the era of text prompts -- like MS-DOS commands -- to an era where anyone can easily use AI through a graphical user interface (GUI).

7. A 100x Richer Software World and the Creator Economy

Eugenia says that currently the world has only built 1% of the software it needs, and the remaining 99% will be built within the next 5 years. Just as YouTube's early content -- home videos and lip-sync clips -- seemed somewhat quirky and toy-like but eventually grew into a massive platform, Wabi's mini apps may start simple and fun but will ultimately evolve into something much larger.

She imagines a future where apps can be treated like "content."

• Health/fitness influencers: Can create 5 mini apps containing their fitness protocols to share with fans and monetize.
• Diverse app styles: Even identical functionality will come in diverse styles and perspectives, like a beautifully designed Pomodoro timer from a specific designer.
• Interest-based communities: People can gather around specific interests -- like Reddit -- to create, share apps, and form communities.

"Imagine a future where apps can be treated like 'content.' Health influencers create mini apps with their workout routines to share with fans, and people seek out apps with a specific designer's unique style."

Wabi will birth a new creator class through these changes. Currently, creators make videos, articles, and various content, but they still can't create software. Wabi helps people without technical knowledge turn their ideas into mobile apps and easily share them with fans without app store approvals or distribution processes. This provides existing creators with new revenue opportunities and fan engagement, while also opening paths for potential new creators to express themselves.

8. AI Evolution from 2012 to 2025: Behind the Scenes of Replika and OpenAI

Eugenia says she became fascinated with AI in 2012 when she encountered Word2Vec technology at Google DeepMind. As language became capable of interacting with computers, she thought -- like Wittgenstein's saying "the limits of my language are the limits of my world" -- that if computers could learn language, they would understand the world and become truly intelligent. Since image recognition models like ImageNet were emerging at the time, she decided to start a company focused on language models and conversation generation.

At the time, there were no relevant papers or algorithms, but when Google published the first deep learning conversation generation model paper in 2015, Eugenia's team went all in on reproducing this model. This resulted in the launch of Replika, the first generative AI chatbot, though she recalls this was 7 years before Transformer models would emerge.

In 2020, when OpenAI offered a partnership before the GPT-3 API launch, Eugenia was stunned. Previously, training conversation models required collecting massive conversation data, but GPT-3 could perform any task through zero-shot/few-shot learning.

"GPT-3 was truly magical. Previously, each model needed specific datasets for training, but GPT-3 could handle any task. Ask it to write tweets like Sam Altman or translate, and it just did it."

Replika was one of the first partners of the GPT-3 API, and at a time when major companies were reluctant to launch generative AI products, it was the only chatbot leveraging generative AI. Other companies were fearful after Microsoft's "Tay" chatbot devolved into a Nazi chatbot within an hour.

Eugenia also shares early OpenAI anecdotes. As a Y Combinator alum, her team would visit the OpenAI team researching in Greg Brockman's apartment to ask questions and share experiences. Ilya Sutskever and Andrej Karpathy were there -- Eugenia calls them "superstars" with deep respect. But she was very disappointed when OpenAI pivoted from language model research to video games.

As Andrej Karpathy later admitted that was the wrong research direction, Eugenia's team was right, but they learned the lesson that being right alone isn't enough. Replika was built into a large business with only a million dollars in investment over a long period, but she regrets not having the "guts" to invest more capital in building their own language models.

"Being right isn't enough. Sometimes you need the bold decision to 'go big or go home.' We didn't have the guts to invest more capital in language model development, and we missed the generational shift."

9. Future AI Hardware: The Trap of Voice-Only Devices

Eugenia presents an interesting perspective on future AI hardware, pointing out that voice-only devices are a tremendous "mind trap." Many developers misinterpret the movie Her, thinking voice interfaces are the ultimate interface, but this misses the film's core point.

"Voice-only devices are a tremendous 'mind trap.' Many developers misinterpret the movie 'Her' thinking voice interfaces are the ultimate interface, but they're missing the core of the film."

Voice interfaces aren't perfect for the following reasons:

• Environmental constraints: Difficult to use in bedrooms with others sleeping, noisy spaces, offices, or even while walking.
• Inefficient information acquisition: Having to ask about remaining time every time you set a timer is inefficient; seeing it on screen is far more intuitive.
• Disruption and discomfort: Having push notifications or text messages read aloud is mostly disruptive and unpleasant.

Eugenia emphasizes that making screen-less devices is the biggest mistake, and advocates for building "screen first" devices. AI devices should be about an AI-first operating system, not voice-driven.

• Locally-running models: Models must run locally on the device.
• Flexible and personalized OS: No fixed apps, users can create and modify software on the fly, with personalization far deeper than today's offerings.

Currently AI is merely an app on smartphones, but she believes there's room for a truly AI-first smartphone to emerge.

Conclusion

Eugenia Kuyda, with Wabi, demonstrates a pioneering approach -- viewing the future of AI not merely from the perspective of technological advancement but deeply considering what meaningful impact it can have on human lives and society. She is confident that AI will open an era where not just a few developers, but everyone creates and shares software reflecting their own needs and desires. This change will go beyond simple app usage, leading to personalized experiences, new community formation, and the birth of a new creator class. Ultimately, Wabi is expected to become a platform that helps us all live richer, more personalized digital lives through AI.