a16zSeeing The Future from AI Companions to Personal Software
CHAPTERS
- 0:54 – 2:55
From AI companions (Replika) to personal software (Wabi): the through-line
Kuyda explains the continuity from Replika to Wabi: using AI to improve daily life through meaningful interaction. The focus shifts from emotional companionship to “personal software”—mini-apps tailored to an individual’s needs, context, and routines.
- •Core mission: AI that materially improves a person’s life day-to-day
- •Replika emphasized relationship/companionship and emotional support
- •Wabi applies the same personal focus to tools and workflows
- •Being early is a recurring pattern—sometimes too early
- 2:55 – 3:43
“It must be an interface problem”: why chatbots under-deliver on AI’s potential
User behavior around tools like ChatGPT clusters around search and writing help, despite models being capable of much more. Kuyda argues the chat/command-line metaphor constrains imagination and prevents mainstream users from discovering richer use cases.
- •Observed usage is dominated by simple tasks (search, writing, homework)
- •Chat/command-line interfaces bias users toward limited affordances
- •Model capability is outpacing how people can practically apply it
- •The next leap requires a more interactive, visual, guided interface
- 3:43 – 4:50
The “Mac moment” for AI: what the next operating paradigm could look like
Kuyda frames current AI UX as the “DOS era” and predicts a Windows/macOS-like breakthrough. She describes an OS oriented around the individual, where AI suggests and composes apps dynamically based on upcoming plans and personal preferences.
- •Current chatbots resemble DOS; a GUI-level shift is coming
- •Future OS homescreen mixes big apps, friend-discovered apps, and self-made apps
- •AI can generate situational tools (e.g., an art-show finder for a NYC trip)
- •The organizing principle becomes “you” (deep personalization), not fixed apps
- 4:50 – 5:54
When apps become like YouTube videos: UGC software and the long tail
The discussion uses TV-to-YouTube as an analogy for what happens when creation becomes accessible. Wabi’s premise is that software will shift from being made only by professional developers to being created, remixed, and shared by everyone.
- •Software today resembles “a few TV channels”: limited supply by pros
- •UGC content platforms show what happens when creation friction collapses
- •There are only ~20M developers; most software reflects their preferences
- •A mass-market consumer product can let non-technical people make software
- 5:54 – 7:55
Ephemeral vs. durable software: tiny, personal, throwaway apps that still matter
Kuyda highlights “ephemeral” software: apps too niche, small, or personal to justify App Store development. She shares examples like a show-specific motivational quote app and a bedtime puzzle game customized for her daughter’s interests and language practice.
- •Many valuable apps are too niche to exist as products on the App Store
- •Personal context (kids, language, interests) drives unique app needs
- •Creation time drops from hours/days to minutes with AI-assisted building
- •AI could proactively suggest situational mini-apps based on context
- 7:55 – 11:37
Replacing paid apps with Wabi creations: making, tweaking, and iterating daily
Anish and Eugenia describe replacing downloaded/paid apps with bespoke Wabi mini-apps (e.g., migraine tracking, restaurant recs, personalized notes, image transformations). A key behavioral shift is continuous iteration—users tweak prompts and features as needs evolve.
- •Users can delete long-tail utility apps and rebuild better versions personally
- •Wabi reduces ad-ridden friction and irrelevant feature bloat
- •Eugenia’s workout tracker evolves into a generator via incremental prompt edits
- •Creation becomes an ongoing loop: build → use → tweak → republish
- 11:37 – 14:17
Investing in “software as content”: mini-apps as shareable media objects
The investors connect the thesis to a new category: “software as content,” where mini-apps function like posts or videos—distributed socially, consumed quickly, and shaped by communities. This reframes software from a static product into a creative medium and distribution game.
- •The YouTube metaphor extends to software distribution and consumption
- •Creation can be intrinsically fulfilling, not purely commercial
- •Software supply expands when non-developers can create
- •Opportunity: mass-market consumer platform, not ‘text-to-app for developers’
- 14:17 – 16:41
Mini-apps as community catalysts: from non-social app stores to social niches
Kuyda argues app distribution needs to become social, enabling interest-based micro-communities to form around mini-apps. A mini-app can be the “starter” for finding others with shared local or niche interests (e.g., specific neighborhoods, parenting contexts, hobbies).
- •Traditional app stores aren’t designed for social discovery (and may never be)
- •Mini-apps can act as conversation/community starters in specific niches
- •Local and identity-driven interests become organizing primitives
- •Community emerges around shared usage, not just shared content feeds
- 16:41 – 19:07
The “organization layer” for vibe coding: guardrails, mobile-first, and platform trust
Wabi positions itself as the organizational layer for consumer-made software—like YouTube for video or Shopify for stores—rather than a web of untrusted links. The team emphasizes guardrails, non-technical UX (no code, no API keys), and platform-level safety for data and persistence.
- •Wabi avoids code exposure; uses consumer-friendly abstractions (e.g., “power-ups”)
- •Mobile-first matters because daily life happens on phones
- •A platform layer is needed so UGC apps aren’t random links with fragile backends
- •Trust/safety: non-pro developers can accidentally leak sensitive data; guardrails help
- 19:07 – 23:11
Wabi as memory, context, and expression: toward Software 3.0 personalization
Building on the mobile-app evolution analogy, Kuyda says AI’s ‘native superpower’ is deep personalization through shared context and memory. Mini-apps can incorporate user prompts, preferences, and environmental specifics, and eventually share context across apps (e.g., fitness ↔ nutrition).
- •AI-native software should personalize beyond UI: behavior, goals, environment, context
- •Users personalize by editing prompts and adding reference data (e.g., gym photo, method book)
- •Platform-level context (age, location, goals) can power all mini-apps consistently
- •Long-term vision: shared context across apps instead of isolated walled gardens
- 23:11 – 28:11
Prompt sharing as emergent behavior: turning messy text prompts into usable apps
Justine and Eugenia describe how people already share prompts socially in clumsy ways (e.g., long prompts in TikTok comments). Wabi reframes prompts into mini-apps with a GUI, examples, model settings, and a one-click “try it” path—reducing friction and improving discovery.
- •Prompt-sharing demand is real, especially in creative subcultures (image/video edits)
- •Today’s workflow is brittle: prompts + app confusion + model selection + copy/paste
- •Mini-app packaging can embed prompt, model, UI, examples, and sharing in one place
- •Lower friction turns fleeting curiosity into repeat usage
- 28:11 – 33:55
100x’ing meaningful software + creator economy: professional mini-app creators emerge
The conversation projects a future where software volume and variety explodes, starting with playful ‘toy’ apps and evolving into serious utilities and creator-driven offerings. Mini-apps become a new creator product—like courses, templates, or merch—enabling style, community, and potentially monetization.
- •Early ecosystems look toy-like, then mature into durable categories
- •Creators can publish mini-app ‘bundles’ as part of their brand/protocol (e.g., fitness)
- •Value isn’t only monetization—also style, taste, and identity in software form
- •Wabi can provide feedback loops: see what people used, remixed, and achieved
- 33:55 – 42:01
How AI evolved since 2012: Word2vec to GPT-3, and surviving the ‘too early’ years
Kuyda recounts entering AI in 2012 through Word2vec and the idea that language learning unlocks world understanding. She traces milestones: early dialogue generation papers, the long gap before transformers, and the shock of GPT-3’s few-shot generality—plus Replika’s early position as a generative AI chatbot when others feared Tay-like failures.
- •2012 spark: Word2vec and the philosophical significance of computable language
- •2015 dialogue generation research prompted a major bet—years before payoff
- •Transformers and later Mina marked step-changes in conversational quality
- •GPT-3 introduced general-purpose few-shot capability; Replika became an early large API user
- 42:01 – 46:26
Lessons from the OpenAI office and founder mindset: be right, but execute
Kuyda describes visiting early OpenAI/YC Research, observing shifts away from language into RL/game worlds, and reflecting on strategic tradeoffs. Her takeaway: scrappiness and profitability can be defensive; sometimes founders must raise big and commit to a generational bet—or risk missing the moment.
- •Early OpenAI felt like a small, star-studded research group; later became a giant
- •OpenAI’s temporary pivot away from language models created uncertainty for language-focused builders
- •Capital constraints can push teams toward revenue optimization over frontier R&D
- •Founder lesson: ‘go big or go home’—being right isn’t enough without execution
- 46:26
Future AI hardware: avoiding the voice-first trap and building AI-first, screen-first OSes
Kuyda argues that voice-only devices are a strategic mistake: they fail in common social settings and are poor for discovery and information intake. She envisions AI-first smartphones/OSes with local models, dynamic software creation, and deeply personalized experiences—where AI is not an app, but the substrate.
- •Voice is situationally constrained (sleeping partner, crowded spaces, office)
- •Even voice assistants ship with screens because visual feedback is essential
- •Discovery and productivity need screens; voice-only harms UX (e.g., spoken notifications)
- •End-state: AI-first OS + local model execution + fluid, personalized software creation
Why AI shouldn’t be “just an app”: setting the stakes for a new interface era
Eugenia Kuyda opens with a strong claim: AI as a standalone phone app is the wrong end-state, and many builders are falling into an interface “mind trap,” especially around voice-first assumptions. The conversation tees up a broader thesis: the next wave of consumer AI will be defined by interface and product packaging, not raw model capability.
- •AI being siloed as a single app on a phone is a temporary, limiting paradigm
- •Builders often over-rotate to voice because of sci-fi narratives (e.g., Her)
- •The real bottleneck is how people access and discover AI capabilities
- •This moment rewards bold product bets (“go big or go home”)
Who creates vs. who consumes: remix culture, social graph, and lightweight collaboration
Kuyda predicts a familiar dynamic: a small minority will create from scratch, while many more will tweak and remix. Wabi leans into this with social features—seeing downloads, comments, and remixing—so users can request changes and co-evolve apps with creators.
- •Original creators may remain under ~10%; most will tweak/remix
- •A social graph enables discovery through friends and community activity
- •Comments can drive iterative improvements or creator-requested tweaks
- •Apps become shared artifacts, not just private utilities
