Moonshots EP 199: OpenAI vs Grok — The Race to Build the Everything App
Summary
Rapid-turnaround episode covering OpenAI Dev Day, with Diamandis, Dave Blundin, Alexander Wissner-Gross, and Immad Mostaque. The key numbers: OpenAI grew from 2M to 4M weekly developers, 100M to 800M weekly ChatGPT users, and 300M to 6B tokens per minute on the API. Mostaque puts this in striking context: 6B tokens/minute equals ~3 quadrillion tokens/year; all humans speak ~50 quadrillion tokens/year. At current growth rates, OpenAI alone will overtake total human speech output within 1-2 years. Wissner-Gross notes 4 billion smartphone users still aren’t using any superintelligence — massive saturation gap. The panel covers three Dev Day announcements: (1) Apps inside ChatGPT (SDK for Booking.com, Figma, Zillow integration), which Mostaque frames as the WeChat super-app battle for attention between OpenAI, Meta, Google, and Elon/X; (2) Agent Builder for visual multi-step workflows without coding, which Blundin and Wissner-Gross both dismiss as “vaudeville-stage” transitional — voice will replace node-based visual programming quickly; (3) Codex voice mode controlling physical devices (cameras, lights on stage), which Diamandis sees as finally solving AV integration pain. Key strategic insight from Blundin: OpenAI agent builder was itself built end-to-end in 6 weeks with Codex writing 80% of PRs, matching AI 2027 report predictions. Mostaque notes Grok model pricing at $0.50 for tasks that cost $5 on OpenAI — 10x cost compression. Wissner-Gross offers the most interesting take: we’re approaching “negative speed” recursive self-improvement where software is written preemptively before humans realize they need it.
Key Segments
- [00:01-07:00] OpenAI Dev Day numbers — 800M users, 6B tokens/min, approaching human speech output volume
- [08:00-12:00] Apps inside ChatGPT — WeChat super-app analogy, attention as finite resource, “every pixel will be generated”
- [12:00-18:00] Agent Builder — vaudeville-stage criticism, voice as future interface, org-chart-to-enterprise replication
- [18:00-26:00] Codex voice mode, AV control demo, recursive self-improvement, Codex getting 2 updates/week
Notable Claims
- OpenAI: 6B tokens/minute = ~3 quadrillion tokens/year; humans speak ~50 quadrillion tokens/year total
- OpenAI agent builder built in 6 weeks, 80% of PRs written by Codex itself
- 4 billion smartphone users not yet using any AI superintelligence
- Grok model pricing: 10x cheaper than equivalent OpenAI tasks ($0.50 vs $5)
- Dario Amodei’s prediction: 90% of code will be written by AI (panel agrees this is materializing)
- Codex CLI gets 2 updates per week — unprecedented for a $500B+ company
Bias/Framing Notes
Panel is almost uniformly bullish on OpenAI despite the episode title suggesting a competitive framing. Grok competition gets minimal actual analysis. The “million startups died” narrative is pushed back on by Blundin — refreshingly contrarian. Mostaque’s token math (OpenAI overtaking human speech) is compelling but conflates API tokens with meaningful communication.