a16zSeeing The Future from AI Companions to Personal Software
CHAPTERS
- 0:00 – 2:19
From AI companions to “personal software”: Eugenia’s through line
Eugenia frames her work from Replika to Wabi as a consistent mission: using AI to improve daily life through meaningful interaction. The shift is from an AI “friend” to mini-apps that help you moment-to-moment with highly personal workflows.
- •Started working on AI in 2012; long-term belief in human-machine conversation
- •AI companions as early focus: emotional support and life improvement
- •Wabi extends the same idea into personal, utility-driven software
- •Mini-apps as a way to make software feel personal and helpful throughout the day
- 2:19 – 4:23
“It must be an interface problem”: why chatbots underdeliver
Despite rapid model capability gains, most users rely on ChatGPT-like tools for basic tasks (writing, search, homework). Eugenia argues the limiting factor is the command-line/chat interface, which doesn’t naturally reveal richer possibilities.
- •Observed users mostly doing simple tasks with powerful models
- •Chat UI behaves like a “DOS-era” interface: limited affordances
- •Research and usage patterns reinforce writing/search as dominant behaviors
- •A new interface layer is needed to unlock model capabilities
- 4:23 – 5:54
The “Mac moment” for AI: an OS built on you
Eugenia predicts a shift akin to TV-to-YouTube: from a handful of professional apps to mass user-generated software. In the future OS, people discover, remix, and receive AI-suggested apps tailored to their context, goals, and plans.
- •Software will follow a TV→UGC evolution similar to YouTube/TikTok
- •Apps built by “all of us for all of us,” sometimes by AI for you
- •Home screen mixes mainstream apps, friend-discovered apps, and self-made apps
- •AI suggests situational mini-apps (e.g., NYC art finder near your stay)
- •Core concept: personalization anchored on the user’s context
- 5:54 – 7:55
Ephemeral vs durable software: niche apps that shouldn’t exist on the App Store
Wabi enables tiny, bespoke tools—things too niche to justify a traditional app business. Eugenia shares examples like a hyper-specific motivational quote app and a bedtime puzzle game customized to her child’s preferences and language learning.
- •Lower cost/time makes “ephemeral” software viable
- •Examples: niche quote app tied to a specific TV show
- •Rapidly built bedtime puzzle game with character-themed and language tweaks
- •Avoids App Store friction: onboarding, paying, limited personalization
- •Future: OS should proactively suggest ready-to-use mini-apps based on context
- 7:55 – 10:03
Replacing paid apps with Wabi creations: tracking, notes, and personal workflows
Anish describes deleting many old paid apps because Wabi-made versions were better tailored and ad-free. Eugenia explains her own “aha” moment building a weightlifting tracker that evolves as she tweaks and republishes it.
- •Wabi can replace long-tail utility apps (migraine tracking, restaurant recs, notes)
- •On-the-fly creation beats ad-heavy, bloated App Store alternatives
- •Eugenia’s example: beginner weightlifting tracker built and iteratively improved
- •Apps evolve from simple trackers to generators (new workouts based on inputs)
- •Publishing and remixing creates a lightweight “mini-app store” dynamic
- 10:03 – 11:22
Who creates vs who consumes: remixing, comments, and a social graph
Eugenia expects original creators to remain a minority, but wants most users to tweak existing apps. Wabi’s upcoming social graph aims to make mini-app discovery social and collaborative, including feedback loops between users and creators.
- •Prediction: <10% will be original creators; many more will tweak/remix
- •Launching social graph: see downloads, usage, and community feedback
- •Comment-driven iteration: users request creator tweaks or remix themselves
- •Apps become shared utilities among friends/family, not just personal tools
- •Discovery becomes a core product surface, not a separate marketing problem
- 11:22 – 14:10
Investing thesis: “software as content” and the YouTube analogy
Anish and the team connect Wabi to an investing theme: software creation is restricted by the small number of developers, so enabling consumers expands the software universe dramatically. The product goal is mass-market creation without “text-to-app developer tooling” vibes.
- •Only ~20M developers: most software reflects their preferences
- •If creation becomes easy, more people will make software for themselves/others
- •Wabi positions as a consumer product, not a dev-adjacent tool
- •No code shown; integrations are simplified as “power-ups”
- •Canva-inspired approach: visual controls and guardrails for delight and safety
- 14:10 – 19:06
Mini-apps as community catalysts—and the need for an “organization layer”
The discussion shifts to why UGC software needs a platform with guardrails, distribution, and trust—similar to YouTube/TikTok for video or Shopify for commerce. Eugenia argues links to random vibe-coded apps are unsafe and unreliable, so an organizational layer must host apps, data, and social discovery.
- •Apps can spark local/niche communities (parents, hobbies, neighborhoods)
- •Guardrails matter: non-pro developers can leak sensitive data unintentionally
- •UGC content succeeded on platforms, not via link-sharing chaos
- •Analogies: GeoCities→LinkedIn; custom stores→Shopify; videos→YouTube/TikTok
- •Wabi as a hosted layer with social graph, integrations, and shared context
- 19:06 – 21:52
Wabi as memory, context, and expression: toward Software 3.0 personalization
Eugenia frames AI’s unique value as deep personalization, beyond “old-school” software wrappers. Mini-apps can encode personal prompts, preferences, and environment—then share context across apps (e.g., fitness and nutrition) without repeated integrations.
- •AI’s “mobile-era breakthrough equivalent” is deep personalization
- •Personalization layers: features, aesthetics/skin, and prompt customization
- •Context inputs can include photos of environments and specific constraints
- •Platform-level memory: age, location, goals, routines shared across apps
- •Shared integrations/context reduce repeated setup across individual apps
- 21:52 – 23:10
Multiplayer and community apps: shared feeds and collaborative experiences
Wabi explores “multiplayer” mini-apps where people use tools together or contribute to a shared community space. Eugenia gives an example: a dog portrait generator that could become a community feed rather than isolated outputs posted elsewhere.
- •Building multiplayer primitives is complex because apps vary widely
- •Goal: friends/family co-use, and open community participation options
- •Example: dog portrait generator app with a universal feed
- •Shifts sharing from external platforms to in-app community interactions
- •Envisions social experiences built natively around mini-app functionality
- 23:10 – 28:12
Prompt sharing emerges as consumer behavior—and why it’s broken today
Justine highlights real-world prompt-sharing on TikTok/Reels (especially among younger creators) as a strong signal. Eugenia argues passing long prompts is like worse-than-DOS commands; mini-apps can package prompts, models, examples, and UI into one “tap-to-try” artifact.
- •Prompt sharing already happens via messy comment threads
- •High friction: finding the prompt, correct app, model choice, and inputs
- •Mini-app links could open a ready-to-run experience with examples/styles
- •Reduces drop-off caused by copy/paste and failed first attempts
- •Applies beyond images to text utilities (e.g., bloodwork analysis workflows)
- 28:12 – 33:46
100x’ing meaningful software: creators, niches, and “weird internet” energy
The group explores a future where apps are treated like content: influencers publish mini-app bundles as part of their “protocols,” and communities form around using them. They emphasize a return to experimental, niche creativity—software that wouldn’t be viable as a standalone App Store business.
- •Thesis: the world has far less software than it needs; creation will explode
- •Apps as content: creators distribute useful tools, not just videos/courses
- •Mini-apps can be monetized or simply expressive (style, taste, worldview)
- •New creator class: designers/experts shipping their own “takes” on utilities
- •Nostalgia for early internet: less polished, more experimental niche artifacts
- 33:46 – 39:25
How AI evolved since 2012: from Word2vec to GPT-3’s “magic” moment
Eugenia recounts early excitement around representing language computationally (Word2vec) and the long road to viable dialogue generation. She describes seeing GPT-3 before launch as a turning point: a general-purpose, few-shot model that changed what was possible for products like Replika.
- •2012: Word2vec made language manipulable; philosophy influence (Wittgenstein)
- •Early years lacked clear dialogue-generation algorithms and models
- •2015: Google dialogue generation paper catalyzed major internal bet
- •Surviving until transformers; Mina paper as another major milestone
- •2020: early GPT-3 partnership; shift from task-specific training to general models
- 39:25 – 43:31
Inside early OpenAI: YC Research days, RL detour, and lessons on execution
Eugenia shares what it was like visiting OpenAI early on—initial openness, then a shift away from language toward reinforcement learning and games. She reflects on being “right” about language not being enough; you also need capital, conviction, and execution to seize generational moments.
- •Early OpenAI access via YC Research; Q&A with key researchers
- •OpenAI moved focus away from language to RL/game environments for a time
- •Replika’s constraints: only $11M raised; survival required revenue focus
- •Lesson: being right isn’t enough—must also execute and raise/bet boldly
- •“Go big or go home” as a core takeaway for the current era
- 43:31 – 50:25
Predicting consumer behavior through empathy—and future AI hardware beyond voice-first
Eugenia attributes product intuition to journalism training and deep empathy with non-technical users. On hardware, she warns of a “voice-first” trap: voice is useful but insufficient for discovery and productivity, so future AI devices should remain screen-first with an AI-first OS and more local models.
- •Empathy and observation (journalism background) as edge in AI product design
- •Many AI builders optimize for themselves; mainstream users get left behind
- •Voice-only devices are constrained by context (privacy, noise, speed)
- •Screens are essential for discovery, productivity, and control
- •Future: AI-first smartphone/OS, more local inference, fewer fixed apps, dynamic creation
