This episode focused on the shift toward local, always-on AI systems and the tools making that possible. The conversation started with Pokémon Go as an example of users generating valuable spatial AI data, then moved into NVIDIA GTC, inference hardware, and the broader push toward on-device agents. The second half centered on building workflows with Claude Code, Open Jarvis, mobile coding limitations, Google’s new embeddings model, and how agent permissions change the way people work with coding tools.
Key Points Discussed
00:01:38 Pokémon Go as unpaid spatial AI field work
00:07:13 NVIDIA GTC and the shift from training to inference
00:10:08 How chipmakers plan for agentic AI and local inference
00:24:16 Stanford Open Jarvis and fully on-device personal AI agents
00:30:05 Beth’s Podcast Buddy build and weekend app experiments with Claude Code
00:31:45 Claude’s one million token context window discussion
00:34:06 Claude usage limits doubling outside peak hours
00:35:45 What Claude Code on a phone can and cannot do
00:38:34 Google’s new embeddings model for locating objects and multimodal search
00:41:05 Brian’s cruise ship hot-and-cold app idea using geolocation and embeddings
00:43:29 How Claude remote works from a phone
00:50:41 Bypass permissions mode and the risks of letting coding agents run freely
00:55:37 Codex full access mode and why Carl prefers its UI
00:58:57 Brian’s story about building for fun versus building on deadline
The Daily AI Show Co Hosts: Brian Maucere, Andy Halliday, Beth Lyons, and Karl Yeh
