Mike Wong (@artixels)

애플 iPhone의 Neural Engine(ANE)에서 동작하는 소형 diffusion 모델을 CoreML 패키지로 변환해, iPhone 15 Pro에서 3.78초 만에 이미지를 생성하고 메모리 사용량을 140MB 미만으로 줄였다. 로컬 실행, 개인정보 보호, 에너지 효율 측면에서 인상적인 온디바이스 AI 사례다.

https://x.com/artixels/status/2042877445662216458

#apple #coreml #diffusion #ondeviceai #neuralengine

For the last decade, "smart" meant "connected." We built better networking libraries and faster serialization, but we were still fighting physics.

The rise of capable silicon has rewritten the rules. The default is no longer "call the API." The default is "run it locally."

#EdgeAI #CoreML #AppleSilicon #iOSDev #MachineLearning

Why can't you train an LLM on your iPhone?
It's not just the speed. It's the Memory Physics.

Ep 6 of Sandboxed is out: "Training vs. Inference."
We break down the wall between "Learning" and "Doing."

https://logicbridge.dev/sandboxed/6

#OnDeviceAI #CoreML #iOSDev #MachineLearning #AppleSilicon

App Swift dùng CoreML (2‑6× nhanh hơn, tiêu thụ ít pin) tích hợp YOLO12n, nhận diện cảm xúc, bảo vệ màn hình khi phát hiện 2 người, và Pomodoro theo dõi chú ý bằng Vision. Hoạt động toàn bộ trên thiết bị, không gửi dữ liệu ra ngoài. 🚀📱

#AI #Swift #CoreML #MacVision #CôngNghệ #BảoMật #Pomodoro #VisionFramework #NhậnDienCảmXúc

https://www.reddit.com/r/SideProject/comments/1qcjnzo/mac_vision_tools_a_menu_bar_app_for_fun_tasks/

🚀 Mac Vision Tools: ứng dụng thanh menu macOS dùng mô hình CoreML chạy trên Neural Engine. Tính năng: phát hiện vật thể (YOLO12n), khóa màn hình khi phát hiện 2 người (Privacy Guard), nhận diện cảm xúc khuôn mặt, đồng hồ Pomodoro theo dõi chú ý. Hoàn toàn xử lý cục bộ, tiêu thụ ít pin. #MacVisionTools #AI #Swift #CoreML #NeuralEngine #Privacy #Pomodoro #CôngNghệ

https://www.reddit.com/r/SideProject/comments/1qcjnzo/mac_vision_tools_a_menu_bar_app_for_fun_tasks/

🎧 Most Core ML “failures” are task mismatch failures.

Classification = identity (what)
Detection = location (what + where)
Segmentation = pixel masks (which pixels)

The simplest task that satisfies your UI is usually the best architecture.

Listen: https://logicbridge.dev/sandboxed/4

#iOSDev #CoreML #OnDeviceAI #Vision

Your Core ML model isn’t a black box.
It’s adjustable logic.

And if your accuracy is *suspiciously* high… it might be cheating.

In Episode 3 of Sandboxed, we translate ML jargon into an iOS-developer mental model:
weights + biases as knobs you tune, loss as a measurable error signal, and training as a feedback loop that feels a lot like build-and-test.

https://logicbridge.dev/sandboxed/3

#iOSDevelopment #CoreML #OnDeviceAI #MachineLearning #Sandboxed

🎧 99% accuracy? Your model might be cheating.

Episode 03 explains “learning” as: forward pass → loss → nudge weights.
Then why models fail in real apps: overfitting + data leakage.

What’s the sneakiest shortcut you’ve seen in data?

Listen now: https://logicbridge.dev/sandboxed/3

#iOSDev #CoreML