苹果重磅发布iOS 18.1开发者测试版，首次引入Apple Intelligence。
The AI system includes an upgraded version of Siri, a smart writing tool, and a view tool to provide a more powerful user experience.
The new version of Siri not only has a dazzling edge glow effect, but also understands the user's ambiguous instructions, supports text-to-speech, and can even answer Apple product malfunction questions.
The writing tool supports text generation, proofreading, and summarization features.
The View tool enhances image search and movie production.
In the future, Siri's personal information analysis and external application linkage, ChatGPT integration, etc., will be added.
At the same time, a 47-page technical report was released, detailing Apple's self-developed large model "AFM".
The report reveals that the AFM-on-device model has 3 billion parameters and is optimized for devices such as iPhones. The AFM-server model runs on Apple's private cloud and is designed for complex tasks.
According to the report, 10,240 Google TPUs were used for the training of these models, and no NVIDIA GPUs were used. Apple's AI outperformed GPT-4 in the instruction following and text summarization tests, demonstrating Apple's ambitions in the AI field.
Netizen test experience: iOS 18.1 developer beta
Apple's iOS 18.1 developer beta introduces features such as Apple Intelligence, Siri, and writing tools for the first time. Tech bloggers like Brandon Butch have a full experience of the new features.
Siri has a new look, not only with a more appealing appearance, but also with the ability to switch freely between voice and text.
User tests have shown that Siri can quickly answer complex questions and understand contextual commands with precision. For example, ask the old and new Siri answers when resetting the iPhone:
The writing tool is also highly acclaimed, with features such as grammar checking, tone adjustments, text proofreading, and shorthanding. In tests, it quickly corrects errors and improves the elegance of expression.
The email feed feature is also popular, as it automatically generates a summary before reading it, so users can quickly understand the content of the message.
However, the photo search feature still needs to be improved. Some users have reported that Siri still gets misidentified or irrelevant search results when processing a particular request. This problem has caused many users to complain.
Apple Intelligence also excels at transcription. It converts phone recordings into text, making it easy for users to record and review important conversations. This feature performed well in the test with high accuracy.
Dubbed "Reduce Distractions," Focus significantly improves the user experience by filtering through AI for important notifications.
In-depth interpretation of Apple's self-developed large model technical report
The 47-page technical report released by Apple explains in detail the architecture and performance of the self-developed large model. According to the report, AFM-on-device contains about 3 billion parameters and is suitable for on-device operation, while AFM-server runs on Apple's private cloud.
AFM model performance
In terms of instruction compliance, AFM-server outperforms GPT-4, and AFM-on-device outperforms open-source models such as Llama-3-8B and Phi-3-mini. In terms of text summarization, the two versions of AFM also have a significant lead in performance.
Model architecture and design
The AFM base model is based on the Transformer architecture, including the SwiGLU activation function and RoPE position embedding. Apple uses techniques such as Group Query Attention (GQA) and pre-normalization to improve the stability and efficiency of model training.
Training and post-training processes
The AFM model was trained using 10,240 TPU chips, not NVIDIA GPUs. Post-training includes supervised fine-tuning (SFT) and reinforcement learning based on human feedback (RLHF), using new algorithms to improve model performance. Apple specifically emphasizes strict adherence to user privacy protection principles during training.
Quantization & Adapter Technology
In order to run efficiently on devices with limited memory, Apple has developed advanced quantization methods and adapter architectures. By training the accuracy recovery adapter, the AFM model achieves efficient quantization while maintaining near-lossless performance.
Application & Evaluation
The AFM model is used in Apple Intelligence to support multitasking.
评估显示，AFM-on-device与Phi-3-mini相比性能更好，AFM-server在多项基准测试中表现接近或优于GPT-4。
Responsible AI principles
Apple has always followed responsible AI principles when developing its AFM models, ensuring that the models are robust when dealing with adversarial prompts. Human rater evaluations show that the AFM model performs well in terms of privacy protection and compliance.
If you find this article helpful, please like, bookmark, retweet and share. In the meantime, please follow me for more updates and insights on artificial intelligence!
reference