the-art-of-core-ai - GitLite

KO | EN

gitlite — search

#cli #editor #codex #csharp #containers #golang #bash #web #mysql #privacy

the-art-of-core-ai

(★ 9)

The Art of Core AI — a free hands-on book on shipping AI models on iPhone/Mac with Apple's Core AI framework (iOS/macOS 27). 13 chapters + 7 labs, from first export to custom Metal kernels. Japanese edition on Zenn.

파일 탐색기

ZIP 다운로드

- 01-it-wasnt-compute.md
- 02-eleven-percent.md
- 03-ane-misconception.md
- 04-sixteen-levels.md
- 05-idle-experts.md
- 06-growing-past.md
- 07-grading-guesses.md
- 08-silent-seconds.md
- 09-write-all-at-once.md
- 10-watts-currency.md
- 11-law-of-memory.md
- 12-the-same-seat.md
- 13-off-the-rails.md
- 14-epilogue.md
- lab-0-environment.md
- lab-1-first-conversion.md
- lab-2-swift-runtime.md
- lab-3-porting-an-llm.md
- lab-4-shipping-to-iphone.md
- lab-5-honest-measurement.md
- lab-6-recipe-index.md
- book.css
- index.html
- llms.txt
build.sh
cover.png
README.md

# CDN으로 사용하기

jsDelivr

jsDelivr는 공개 GitHub 리포지토리를 별도 설정 없이 CDN으로 즉시 서빙합니다. 버전과 파일을 고르면 웹페이지에 바로 붙일 수 있는 링크와 예시 코드가 만들어집니다.

버전

파일

링크

예시

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<img src="cover.png" width="320" alt="cover">

**Field techniques for putting AI on your phone — written entirely from real measurements.**

On-device AI is governed not by how fast the chip computes, but by **how many bytes you move**.
This book builds that thesis one measured fact at a time: 30+ models shipped to real iPhones and
Macs on Apple's Core AI framework (iOS/macOS 27) — LLMs, MoE, diffusion LLMs, speech, music,
vision — every number reproducible, every wrong guess kept in the text.

📕 **Read online:** https://john-rocky.github.io/the-art-of-core-ai/ (single page) — or browse the chapters below.
🇯🇵 The original Japanese edition is published [on Zenn](https://zenn.dev/mlboydaisuke/books/coreai-textbook).

## Chapters

| # | Title |
|---|---|
| 1 | [It Wasn't Compute That Caught Up](chapters/01-it-wasnt-compute.md) |
| 2 | [11% of the Ceiling](chapters/02-eleven-percent.md) |
| 3 | [The Dedicated-Chip Misunderstanding](chapters/03-ane-misconception.md) |
| 4 | [Sixteen Levels](chapters/04-sixteen-levels.md) |
| 5 | [The Idle Experts](chapters/05-idle-experts.md) |
| 6 | [The Ever-Growing Past](chapters/06-growing-past.md) |
| 7 | [Grading Guesses](chapters/07-grading-guesses.md) |
| 8 | [The Silent Seconds](chapters/08-silent-seconds.md) |
| 9 | [Writing It All at Once](chapters/09-write-all-at-once.md) |
| 10 | [Watts, the Other Currency](chapters/10-watts-currency.md) |
| 11 | [The Law of Memory](chapters/11-law-of-memory.md) |
| 12 | [The Same Seat as First-Party](chapters/12-the-same-seat.md) |
| 13 | [Off the Rails](chapters/13-off-the-rails.md) |
| — | [Epilogue: How Many Bytes Did You Move?](chapters/14-epilogue.md) |

## Labs (hands-on)

| Lab | Title |
|---|---|
| 0 | [Environment Setup](chapters/lab-0-environment.md) |
| 1 | [Your First Conversion](chapters/lab-1-first-conversion.md) |
| 2 | [Driving It from Swift](chapters/lab-2-swift-runtime.md) |
| 3 | [Porting an LLM](chapters/lab-3-porting-an-llm.md) |
| 4 | [Shipping to iPhone](chapters/lab-4-shipping-to-iphone.md) |
| 5 | [Honest Measurement](chapters/lab-5-honest-measurement.md) |
| 6 | [Recipe Index](chapters/lab-6-recipe-index.md) |

## Where the measurements came from

Every number in this book was measured on the ports in the Core AI model zoo, and the working
notes behind each chapter are published:

- [**Core AI knowledge base**](https://john-rocky.github.io/coreai-model-zoo/) — the notes, a page
  per topic: stateful KV cache, AOT and specialization, ANE vs GPU authoring rules, the
  performance ceiling, custom Metal kernels. Where a chapter compresses a finding into a
  paragraph, the note has the run that produced it.
- [**coreai-model-zoo**](https://github.com/john-rocky/coreai-model-zoo) — 57 models, each with
  the recipe that produced it and a card stating what was gated and on which hardware.
- [**CoreAIKit**](https://github.com/john-rocky/coreai-kit) — run any of them from Swift in one
  line, if you would rather use a model than convert one.

## Author

[**Daisuke Majima (MLBoy)**](https://github.com/john-rocky) — the ports this book measures live in the
[Core AI model zoo](https://github.com/john-rocky/coreai-model-zoo) and on
[huggingface.co/mlboydaisuke](https://huggingface.co/mlboydaisuke); the same author runs
[devicemark](https://devicemark.github.io/) (on-device LLM leaderboard) and posts as
[@JackdeS11](https://x.com/JackdeS11).

## License

Text © the author. Code snippets are provided under the MIT license.