06-reference

moonshots ep209 gemini3 missed

Wed Nov 19 2025 19:00:00 GMT-0500 (Eastern Standard Time) ·reference ·source: Moonshots Podcast ·by Peter Diamandis
gemini-3googlebenchmarksvending-benchai-agentsautonomous-businesssingularity

Moonshots EP 209: What Everyone Missed About Gemini 3

Summary

The panel dissects Google’s Gemini 3 release, which Alex calls the biggest model release since OpenAI’s O3. Key highlights: Gemini 3 scores nearly 50% on Humanity’s Last Exam, integrates agentically across all Google Workspace products, and exhibits “big model smell” — capabilities that can’t be replicated through extended reasoning alone. Alex one-shotted a cyberpunk FPS game with a 140-character prompt. The VendingBench Arena benchmark steals the show: AI agents manage a simulated vending machine economy with email, banking, and inventory tools, and Gemini 3 delivers nearly 3,000% more profit than GPT-5 or Claude Sonnet. The panel argues this proves AI can function as first-class economic actors, with Dave noting the internet advertising business ($300B/year) is already fully non-human. Google’s Anti-Gravity IDE (staffed by former Windsurf team members) and Gemini Live’s improved natural voice also get coverage. Duolingo is down nearly 50% over the past year as AI translation improves.

Key Segments

Notable Claims

Guests

RDCO Mapping