Apple's AI Rollout Will Likely Stretch Into Next Year
Why will Apple Intelligence only be available in American English initially?
Apple Intelligence will initially be available in American English because it allows the company more time to train AI models on other languages. Training AI models for different languages is a major undertaking, and starting with American English gives Apple the opportunity to gradually expand its language support. This approach also enables the company to ensure the AI features work effectively and efficiently before rolling them out to a broader audience.
What major challenges does Apple face in training AI models on languages other than American English?
Training AI models on languages other than American English presents several major challenges for Apple. One significant hurdle is the need for more data to train robust models for languages with less representation. Traditional translation models based on large language models (LLMs) perform well with languages abundant in data but struggle with underrepresented languages4.
To address this issue, Apple has developed a novel concept called contrastive alignment instructions, or AlignInstruct. This approach focuses on maximizing the utility of available resources by introducing a cross-lingual discriminator crafted using statistical word alignments to strengthen the machine translation process. AlignInstruct leverages the strengths of both machine translation instructions (MTInstruct) and advanced cross-lingual understanding, improving translation quality across various language pairs, particularly in zero-shot translation scenarios4.
Another challenge is the need to maintain the cultural nuances and context of different languages while training AI models. As language models are trained primarily on English text, they may have a narrow, western-centric lens, potentially leading to miscommunications and misunderstandings. Enhancing AI models with culturally diverse knowledge and understanding is crucial to avoid erasing cultural differences and perpetuating stereotypes.
Apple also faces the task of scaling and updating AI models for multiple languages and dialects, which can be a considerable effort. The company aims to overcome this challenge by consolidating models for various dialects of English, American, British, and Indian into one versatile neural network language model (NNLM). This approach simplifies the development process and significantly reduces the environmental impact of training multiple models.
In summary, the major challenges Apple faces in training AI models on languages other than American English include data scarcity for underrepresented languages, maintaining cultural nuances and context, and scaling and updating models for multiple languages and dialects.
How does Apple plan to manage the development and release of different features within Apple Intelligence?
Apple plans to manage the development and release of different features within Apple Intelligence by taking a slow and steady approach. According to a Bloomberg News report, Apple Intelligence won't be available for developers to test out until late summer, and it won't be part of the first beta releases of Apple's new operating system updates. The features will only work on some devices and only in American English to start with, and customers may have to sign up for a waitlist to try them.
This approach has a few benefits. For instance, it makes staffing easier, as Apple can assign its engineers to certain features, release that technology when they're finished, and then move them to other features. By starting off with American English, Apple also gets more time to train AI models on other languages, which is a major undertaking.
As for the development process, Apple has created a set of Responsible AI principles to guide how they develop AI tools and the models that underpin them. These principles focus on empowering users with intelligent tools, representing users authentically, designing with care to avoid potential harm, and protecting privacy by not using users' private personal data when training foundation models.
In summary, Apple plans to manage the development and release of Apple Intelligence features by taking a gradual, controlled approach that prioritizes careful engineering, language model training, feature development, and quality evaluation. This strategy allows Apple to ensure that its AI tools are developed responsibly and in line with the company's core values.