MBW 1044: Live, Love, Re-Org - The Strength of the M5 Mac Studio Running Local Agents
Summary
Federico Vittici joins MacBreak Weekly to review the M5 Ultra Mac Studio, focusing on its performance for local AI agents. He details his experience running a 24/7 local model in a storage unit and benchmarks the new hardware against previous generations. The discussion covers the rapid improvement of MLX-based inference engines, the trade-offs between local privacy and cloud subscriptions, and the practicality of high-end Macs for enterprise or hobbyist use.
Topics discussed
Intro: Federico Viticci joins for M5 Ultra and local AI
Banter: Apple merch and the M5 Ultra review unit
Local AI setup: Running DeepSeek 24/7 and cooling issues
Benchmarking methodology and the challenge of review units
MLX framework, model selection, and prompt processing speed
M5 Ultra for video editing vs. AI workloads
Cost analysis: Local AI hardware vs. cloud subscriptions
Clustering Macs with Thunderbolt 5 and RDMA
The AI upgrade cycle and the 'hamster wheel' of hardware
OpenAI's shift to API billing and Apple's AI strategy
John Ternus reorg: Streamlining Apple's hardware teams
Apple's AI lag: Siri, agents, and trust issues
Speculation: AI-generated mobile OS and third-party marketplaces
Linux on Mac: Asahi project and personal Linux experiments
AI agents fixing network issues and the Tailscale story
Verizon and T-Mobile eSIM pricing and plans
iOS 26.0.1 patch: Fixing a critical unlock crash
Meta's new VR glasses vs. Apple Vision Pro
Picks of the Week: KVM, OpenMinis, and Cerebus
Picks of the Week: Retro Mac apps and Keychron K3 keyboard
Outro: iPhone Duo accessories and show credits
Listen ad-free on Castria