Key Takeaways

  • William Imoh built Mural, a real-time voice app for learning Norwegian, in Swift and Kotlin without writing native code himself.
  • Imoh skipped cross-platform wrappers like React Native, using Codex to write native code directly for platform-specific iOS features.
  • Mural runs on OpenAI's GPT-Live-1 model for low-latency speech and conversational interruptions, costing $3 to $5 per 30-minute session.
  • Claire Vo highlights GPT-Live as an underappreciated tool for voice control and instant interactions because raw speed unlocks conversational modalities.
  • Imoh runs solo operations at Hackmamba through a Lord of the Rings-themed Grok Bot fellowship, while noting the existential dread of operational AI dependence.

Your Stack No Longer Depends on What You Know

When William Imoh wanted to build Mural, a voice app to practice Norwegian, standard developer logic dictated React Native or Flutter. Without knowing Swift or Kotlin, founders default to cross-platform layers. Imoh rejected that path.

He wanted a native feel and system-level iOS features, so he let Codex generate the codebase. As Imoh explained: “I guess when we have AI agents building stuff like the language doesn't matter like the language I know doesn't matter anymore. I just choose the best possible language for the tech that I'm building. Uh so that's why I went with swift and cotlin even though I haven't written a line of swift in my life.” (play from 7:02)

For founders, this breaks an old hiring constraint. Early technical choices almost always reflect the founder's resume. You built in Python or JavaScript because that was what you typed at your last job. When agents write the code, picking a framework based on personal familiarity produces inferior software. You pick the target runtime that delivers the cleanest user experience, then let the agent build to that specification.

The Unit Economics of Real-Time Voice

Native UI was only half the equation. Language apps like Duolingo rely on static prompts and recorded audio. Mural needed to simulate a live tutor who listens, pauses, and responds without delay.

Imoh plugged Mural into OpenAI's GPT-Live-1 model right after its release. As Imoh put it: “I described the experience I want which was basically I want this application to run on GPT live one because GPD live one was just released I think a week after. So GPD live one was the technology that I tested that enabled me build this.” (play from 9:25)

The unlock is interruption handling. Human conversation is messy; speakers talk over each other and correct themselves mid-sentence. Claire Vo noted that GPT-Live handles real-time audio and immediate interruptions with minimal latency, making it uniquely suited for voice control and live speech testing.

The trade-off is cost. Imoh noted that running Mural costs $3 to $5 for every 30-minute conversation. At ten hours of monthly practice, a single active user generates $60 to $100 in model fees alone. That breaks standard consumer software pricing. You cannot charge a flat $10 monthly subscription when an active learner burns through that budget in ninety minutes. Founders building on real-time voice must charge per minute, sell high-ticket educational packages, or wait for inference costs to collapse.

The Operational Trap of Autonomous Agents

Coding Mural was not Imoh's only AI experiment. At Hackmamba, he runs operations using a Lord of the Rings-themed fellowship of Grok bots with full tool access across his workflow.

Yet Imoh admitted the dark side: the dread of operational dependence. When autonomous agents run your core workflows, you lose the muscle memory of how your business actually functions. The speed is immediate, but the vulnerability is real.

What to Do With This

Audit your product roadmap for features where you picked a familiar wrapper instead of native code. Re-prompt a code-generation model with the native target framework (Swift for iOS or Kotlin for Android) and build a native prototype this week. If you plan to introduce live voice via models like GPT-Live-1, calculate your unit cost at $0.10 to $0.17 per minute and set up usage-based pricing before launch.