706: No One Is Getting a Good Deal Tonight

What It Covers

ATP episode 706 covers Apple's surprise announcement of new Mac Mini and Mac Studio models featuring the M6 and M5 Ultra chips, macOS 26 Golden Gate UI refinements in beta six, iCloud shared album storage policy changes, LLM non-determinism from floating-point arithmetic, and John Siracusa's deliberate decision to purchase a mid-tier M5 Max Mac Studio as a transitional machine while awaiting future AI-optimized chips.

Key Questions Answered

• •M6 Chip Architecture: The M6 is Apple's first two-nanometer chip and uniquely doubles the neural engine count compared to previous no-suffix chips, delivering up to 10.3x faster long-form audio transcription versus M1. However, it reduces super cores from four to two while adding four new medium performance cores, making direct CPU comparisons complex. No M6 Pro, Max, or Ultra variants are planned, making this the sole M6 chip in the lineup.
• •Mac Mini Pricing Reality: The M6 Mac Mini starts at $900, up $200 from the M4, but inflation-adjusted pricing shows the 2018 model was actually the most expensive at $1,076 in today's dollars. Storage upgrades remain punishing: a two-terabyte SSD add-on costs $1,000, exceeding the base computer price. The base model consistently delivers the best value; any storage or RAM upgrade rapidly doubles or triples the total system cost.
• •Mac Studio Ultra Scaling Inefficiency: The M5 Ultra connects two M5 Max chips via UltraFusion, delivering only 1.3x multi-threaded CPU performance over the M3 Ultra despite two additional chip generations. GPU compute for AI tasks jumps 4.5x over the M3 Ultra due to architectural changes, not core count increases. For pure CPU workloads, buying two networked M5 Max Mac Studios with RDMA over Thunderbolt 5 likely delivers better performance-per-dollar than a single Ultra configuration.
• •LLM Non-Determinism at Temperature Zero: Setting an LLM temperature to zero does not guarantee identical outputs across runs. Floating-point arithmetic is non-associative, meaning concurrent execution across cores produces different rounding outcomes depending on which core completes operations first. This accumulates across millions of matrix operations inside a model. Developers relying on temperature-zero reproducibility for testing or watermarking should treat outputs as statistically stable but not mathematically guaranteed to be bit-identical across runs.
• •3D Printing for Micro-Problem Solving: Semi-rigid TPU filament, specifically Bambu's AMS-compatible TPU, solves real-world hardware replacement problems that metal parts cannot easily address. Printing replacement window screen clips in TPU rather than replicating metal originals provides flexibility that prevents finger cuts during installation, stronger layer adhesion than rigid filaments, and near-indestructibility under repeated stress. Orienting prints vertically on the build plate maximizes layer adhesion strength for load-bearing clips and fasteners.
• •iCloud Shared Albums Storage Policy Change: In the OS 27 generation, shared album photos and videos count against the album owner's iCloud storage, ending the previous arrangement where up to 5,000 items stored at reduced resolution did not consume quota. Critically, photos added by invited collaborators also count against the owner's storage. A free alternative exists: temporary shared albums that auto-delete after 30 days. Shared albums created on OS 27 are incompatible with devices running older operating systems.
• •Strategic Computer Purchasing Timing: John Siracusa's framework for major hardware purchases applies equally to Macs, TVs, and cars: identify whether the current generation represents a meaningful architectural leap or a transitional product. The M5 Ultra is explicitly a stopgap — two M5 Max chips fused together with no novel die architecture. Rumors suggest the M7 generation will be the first chip designed from the ground up around local AI inference, making the current cycle a poor time to invest in top-tier configurations.