Sandboxed - On-Device AI for iOS

We move from theory to practice by building a real-time image classifier from scratch. We explore the architectural challenges of integrating Core ML into an iOS app, using the Vision framework as the critical "transmission" between the camera sensor and the model engine.

We explore the architectural discipline of knowing when to use Machine Learning and when to stick with deterministic code. We dissect the "AI Everywhere" trap, compare the costs of Code 1.0 vs Code 2.0, and establish the "Certainty Gap".

We explore the architectural tension between AI capability and user confidentiality. From the Secure Enclave to Private Cloud Compute, we breakdown how Apple builds trust into the silicon itself.

Sam and Alex peel back the layers of Apple's machine learning stack. They navigate from the high-level magic of the Vision framework down through the engine room of Core ML and finally to the bare metal physics of Accelerate. The goal: teaching developers how to choose the right tool for the job to save battery, thermal headroom, and sanity.

We strip away the "Interpreted AI" myth to reveal the reality of the Core ML Compiler. We explore Layer Fusion, the ANE's Systolic Array architecture, the "Ping-Pong" performance trap, and why shipping a `.mlpackage` is actually shipping a blueprint for a silicon factory.

We strip away the biological metaphors of AI to reveal the mechanical reality: differentiable decision trees. We explore the Forward Pass, the role of ReLU as a logic gate, and why "Deep" hierarchies work. Finally, we break down the hardware reality—why CPUs are Ferraris, NPUs are Bullet Trains, and why memory bandwidth is the only metric that matters.

As iOS engineers, we respect the science of training models, but we live in the trenches of inference. In this episode, we explore the "Great Divide" between creating a model and running it. We break down the memory mechanics of O(N) training vs O(1) inference, dissect compiler optimizations like kernel fusion, and explain exactly how the Apple Neural Engine cheats bandwidth physics using quantization.

We deconstruct the .mlmodelc bundle to understand exactly what we're shipping. Does a model have logic? No. It has settings. We explore the "Factory" analogy, the difference between architecture and weights, and why we must treat ML models as frozen assets, not readable code.

We stop treating AI as a "magic black box" and start treating it as a function with a strict return type. Sam and Alex break down the three fundamental primitives of Machine Learning—Classification, Regression, and Clustering—and explain how to map vague product requirements into rigorous engineering specifications.

We dismantle the mystical "black box" of AI to reveal simple mechanical gears: Function Approximation and Optimization. We explain why code is dying, how "learning" is just curve-fitting, and why your iPhone is built to run these models, not build them.