AI can't even turn on the lights | The Vergecast

The latest AI advancements, particularly Large Language Models, are struggling to deliver reliable smart home functionality, indicating a potential bubble in the AI industry despite significant investment and hype.

• Current AI, especially LLMs, struggles with basic real-world tasks like smart home control, indicating a reliability gap that challenges the industry's significant investment.

• Apple's M5 chip updates and the Vision Pro's development reveal product strategies driven by supply chain logistics rather than user-centric innovation.

• The push for AI-generated content, including 'erotica,' reflects a drive for user engagement and revenue, raising concerns about content quality, ethics, and the proliferation of convincing fake information.

The podcast hosts discuss the challenges with smart home technology, noting that even major tech companies like Apple, Google, and Amazon struggle to implement reliable AI assistants based on LLMs to perform basic tasks like turning lights on and off. This difficulty highlights a critical limitation of current AI, which, despite its ability to understand natural language, cannot consistently execute multi-step commands in the real world. This unreliability in fundamental applications suggests that the current AI investment boom might be a bubble, as the foundational technology is not yet capable of delivering the envisioned 'ambient computing' future.

Baby Tech & Smart Home Devices

• 00:03:03 During the hosts' parental leaves, they experimented with baby tech and smart home devices. One host attempted to upgrade baby monitors to split-screen versions but found the user interface too complex, opting instead for two separate, non-internet-connected monitors. Another tried an 'Owlet' smart sock for newborns, but returned it due to its reliance on outdated 2.4 GHz Wi-Fi and frequent false alarms, especially for very small babies.

Smart Home Automation Challenges

• 00:05:14 Smart home automation, particularly with Matter and Thread devices, involves writing complex 'if-then' statements for conditional actions, such as shades closing to trigger lights off and fans on. While these systems appear to work well within Apple Home, Google Home often struggles with consistency due to conflicting thread networks and the need to repeatedly reconfigure the system. This instability suggests current smart home technology often feels like a 'hack' rather than a seamless, integrated experience.

Apple's M5 Chip Releases

• 00:18:48 Apple recently updated the iPad Pro, 14-inch MacBook Pro, and Vision Pro with the new M5 chip. These incremental updates are considered unexciting, with analysts suggesting consumers wait for the next generation. The hosts speculate these releases are driven by Apple's supply chain management, potentially shifting chip production capacity and leveraging cheaper new architecture rather than significant product innovation. Apple's press releases even suggest that M4 users are 'fine,' highlighting the minimal performance leap for recent buyers.

The Vision Pro's Identity Crisis

• 00:24:28 The Vision Pro, updated with an M5 chip, has paradoxically become heavier, moving in the 'wrong direction' from user desire for a lighter, cheaper device. Apple continues to roll out immersive content like NBA games, which are met with disinterest as they fail to offer a compelling viewing experience compared to traditional TV. The product's high price and ongoing challenges, alongside upcoming competition from Meta and Samsung, suggest the Vision Pro is increasingly becoming an isolated and puzzling product in the market, with Apple reportedly trying to pivot towards lighter smart glasses.

AI's Core Problem: Unreliable Execution

• 00:41:50 The core problem with current AI, particularly LLMs, is their inability to reliably execute multi-step commands in the real world, as evidenced by major tech companies failing to create functional smart home assistants. While LLMs excel at 'fuzzy search' and understanding natural language input, they lack true understanding and a reliable feedback loop to correct errors in complex chains of action. This leads to high failure rates for practical tasks, making them unsuitable for the 'ambient computing' vision that would justify current AI investments.

The AI Bubble and Content Moderation

• 01:06:50 The AI industry is seen as a bubble, driven by financial speculation and a lack of reliable, real-world products. Companies like OpenAI are relaxing content restrictions, even allowing 'erotica' with ChatGPT, to maintain user engagement and revenue. This move, echoing past tech industry stances on content moderation, leads to platforms being flooded with increasingly convincing, yet often fake or problematic, AI-generated content. This approach prioritizes stickiness and monetization over ethical considerations and the actual utility of AI tools, contributing to a distrust in digital information.