AI can't even turn on the lights

What It Covers

Nilay Patel returns to discuss why major tech companies cannot reliably build AI-powered smart home assistants, arguing this failure reveals fundamental limitations in LLM technology that threaten the entire AI investment bubble beyond financial concerns.

Key Questions Answered

• •LLM Technical Limitations: Apple, Google, and Amazon all fail to reliably turn lights on/off using AI assistants despite massive resources. Each step in the automation chain needs 99.999% reliability, but current LLMs achieve roughly 95% per step, compounding into 50% overall failure rates across multi-step tasks.
• •Smart Home Integration Reality: Matter and Thread devices require complex workarounds like pairing to HomeKit first, then generating codes for Google Home. Multiple competing thread networks force router resets to base configuration. The promised seamless multi-platform harmony remains unreliable despite standardization efforts.
• •AI Product Economics: OpenAI plans trillion-dollar compute spending while companies add AI girlfriends and erotica features for user retention. Circular investment deals between AI companies, cloud providers, and chip manufacturers mirror 1990s telecom bubble patterns. Revenue models depend on features that drive engagement over utility.
• •Vision Pro Strategy Confusion: Apple ships heavier M5-equipped Vision Pro at $3,500 while market moves toward $1,500 competitors and lightweight smart glasses. Meta Ray-Bans succeed primarily as cameras with headphones, not AI interfaces. Supply chain management and chip production cycles now dictate product release timing over user needs.
• •Ambient Computing Gap: Microsoft advertises Windows 11 as "the computer you can talk to" while executives admit using words like "should" and "ought to" instead of demonstrating working products. Natural language input works, but the 10+ step chain from command to execution breaks down without error detection and self-correction capabilities.