The category I've been writing about for two years arrived faster than the broader industry conversation acknowledges and slower than the early enthusiasts expected. The shape of what's actually here is more nuanced than either timing claim alone suggests.
The Copilot+ PC marketing pushed NPU silicon into mainstream Windows hardware in 2024-2025. The hardware shipped; the software story hasn't caught up to Apple's. Worth being honest about where each platform actually stands at the end of 2025.
DSPy lets you optimize prompts the way you'd optimize a model, programmatically, against an objective, with structured evaluation. The default DSPy workflow sends a lot of prompts to a hosted optimizer. The local-first version is doable, faster than expected, and worth the setup.
The default for AI in 2025 is hosted. The hosted default makes sense for some workloads and is structurally wrong for others. Worth being explicit about which workloads belong on-prem and why the inertia is hard to break.
The cloud bill for serious AI training is still aimed at the hyperscaler tier. The patterns for doing meaningful training without that bill have matured. Worth being explicit about what's gettable on a budget that doesn't require enterprise procurement.
Most home AI setups die after the novelty wears off. The ones that survive into year two share a small set of operational properties, boring, durable, owner-friendly. Worth being explicit about what makes a stack worth keeping rather than just worth building.
Retrieval-augmented generation works well when the corpus fits on one machine. The honest version of what to do when the corpus outgrows that, without rebuilding the whole stack on cloud, is more interesting than either the all-local or all-cloud framings suggest.
Two years past the encoding-a-person framing. The thesis held in the parts I expected and bent in the parts I didn't. Worth being honest about what survived contact with the actual technology and what was just well-aged speculation.
The Apple Neural Engine doesn't get featured the way the GPU and unified memory do. It's the part of the chip that does the most for the on-device personal-AI story Apple is supposedly betting on. Worth being plain about why it matters and why nobody talks about it.
Spent a weekend doing actual training work on the Mac Studio rather than the usual inference-only experimentation. MLX is meaningfully more capable for training than it was six months ago. Worth being plain about what works, what doesn't, and what it means for the Apple Silicon training story.
Three platforms running the same models on the same prompts. The Studio numbers are mine; the 5090 and Threadripper numbers are well-published comparables. The takeaway isn't which one wins, it's that the answer depends on which workload you actually have.
Two and a half years ago I wrote a piece arguing personal AI would be the durable category, not enterprise chatbots. The 2025 version of that bet is partway right and wrong in interesting ways. Worth being clear about what landed and what didn't.
