Google I/O 2025 AI Bombshell, Nvidia’s $50 Trillion Robot Bet & Coinbase Hacked

Discussion of Mark Zuckerberg's rejection of traditional delegation advice, staying deeply involved in Meta's engineering details. Contrasts founder-led companies (Zuck, Chesky, Musk) with professional CEO culture and the McKinsey-driven delegation industrial complex. Includes a practical framework for deciding when to delegate vs stay hands-on.

• •Zuckerberg doesn't believe in delegating - gets into any level of detail with any engineer or team he wants
• •Founder mode (Airbnb's Chesky) vs professional CEO delegation culture - middle management stack at risk
• •With AI/agents, being a 'one person department' is now viable - don't just forward prompts
• •If your job is 'if this, then that' forwarding, you'll be automated quickly
• •Three-step audit: Can AI do this task? Can I leverage technology instead of people? Am I adding unique value?

Brian Armstrong's response to social engineering attack where hackers bribed overseas support staff for customer data. Instead of paying $20M ransom, Coinbase invested $400M in security hardening, bounties, and customer reimbursement. Demonstrates long-term thinking and setting cultural precedent over short-term fixes.

• •Hackers used social engineering on overseas support team to get customer crypto holdings data
• •Armstrong refused $20M ransom, invested $400M in security, bounties, and reimbursements instead
• •Share price reacted positively - playing the long game protects shareholder value
• •Contrast with Australian corporate responses (Optus, Medibank) that negotiated and still got leaked
• •Founder-led principle: 'We will not negotiate with terrorists' sets cultural standard

Google's massive product launch including Gemini 2.5, agent mode, Veo-3 video AI, Project Astra live demo, and Jules code assistant. Discussion of Google's branding crisis with Gemini vs just calling everything 'Google', and how Larry Page and Sergey Brin's return is driving aggressive innovation to compete with OpenAI.

• •Agent mode launched - autonomous task execution competing with Manas.im
• •Veo-3 video generation, Jules code assistant, Project Astra (real-time camera + voice interaction)
• •Google finally integrated AI chat into search and called it 'Google' - solving branding crisis
• •Larry & Sergey back in engineering rooms driving product-first culture after years of protecting search monopoly
• •Plugging Gemini into Gmail, Maps, Calendar, Tasks - full ecosystem activation

Microsoft CTO Kevin Scott validates the Model Context Protocol (MCP) as the 'HTTP for the agentic web' - a standardized language allowing different AI tools and agents to communicate. Discussion of pattern matching from previous tech revolutions and understanding protocols at first principles level.

• •MCP is a protocol (like HTTP) that allows different AI agents and tools to speak to each other
• •Microsoft CTO Kevin Scott publicly endorsed this framework at Build conference
• •Pattern matching from previous tech evolutions helps predict future developments
• •Understanding things at first principles level (what IS a protocol?) enables better strategic thinking
• •VCs succeed by pattern matching successful models to new revolutions

OpenAI's Codex announcement - a coding-specific model that can work autonomously for 48+ hours on tasks. Acquired Windsurf to compete with Cursor. Discussion of OpenAI's strategy to own the consumer layer of all AI interactions and the race for distribution vs raw capability.

• •Codex is coding-specific model (not general purpose) - can work autonomously for 48 hours without breaks
• •Enables 'one person billion dollar company' concept and 'one person department' reality
• •OpenAI acquired Windsurf (Cursor competitor) to own consumer coding layer
• •Distribution now matters more than raw AI capability - race to integrate into ecosystems
• •Coding seems to be first domain where true autonomous agents work at scale

Practical demonstration of using Relevance AI (Australian startup, $24.5M raised) to automate personal workflows. Example: forwarding rental listings to email, auto-extracting details to Notion database. Emphasis on mindset shift - ask 'how can AI do this task?' before doing anything manually.

• •Relevance AI - Sydney-based, drag-and-drop workflow automation (easier than n8n)
• •Real example: rental listing emails auto-parsed and sent to Notion database in 10 minutes
• •CalAI app (photo food for calories) doing $2-3M/month with <10 employees - vibe-coded into existence
• •DripMax app (rate outfits with AI) shows how simple ideas can scale quickly
• •Mindset shift required: before doing any task, ask if AI can do it better or for you

NotebookLM now available as mobile app, allowing interactive podcast creation from any content. Original founders left Google to build stealth morning podcast app that scans your ecosystem (calendar, tasks, transcripts) to give personalized daily briefing with encouragement and advice.

• •NotebookLM mobile app released - create personalized podcasts from documents, YouTube, legal files
• •Can interrupt and interact with AI podcast hosts in real-time to refine learning
• •Ask it to apply insights to YOUR business specifically (e.g., 'what does Chesky's advice mean for W Build?')
• •Original founders left to build morning podcast app - scans your tools, gives daily briefing while brushing teeth
• •Simple idea: personal podcast on your day, commitments, encouragement - sold NotebookLM for extraordinary amount

Jensen Huang declares humanoid robots a $50 trillion industry because the world is optimized for human form factor. NVIDIA's Isaac Groot model based on 'Thinking Fast and Slow' - robots trained by humans, then thousands of simulations run in Omniverse. Robots born with years of experience from day one.

• •Humanoid form factor wins because world is built for humans - scalable like humans (8B+ units possible)
• •NVIDIA infrastructure layer: Isaac Groot uses 'slow reasoning then fast action' model from Kahneman's book
• •Humans train robots once, then thousands of physics-based simulations optimize in Omniverse
• •Robots born with equivalent of human lifespan of experience and context from day one
• •NVIDIA working with Disney/DeepMind - shift from animation to simulation based on real-world physics

Jensen reveals OpenAI's internal KPIs are 'tokens per quarter' - treating intelligence as manufacturing. Ilya Sutskever's insight that predicting next token requires deep understanding of reality. Philosophical connection: humans are also just predicting their next output (token) based on all data captured in life.

• •Token = how LLMs break up information (words, video, images) to process and measure throughput
• •OpenAI's KPIs now 'tokens per hour' - manufacturing intelligence at scale
• •Ilya: 'All AI is predicting next token, which requires deep understanding of underlying reality'
• •Human parallel: we capture data (experiences, learning, risks) to produce our next output/token
• •Buddhist-like mantra: don't stress mistakes - there's always another prompt, another chance for excellent token