You might be a little confused by Apple Intelligence, Sam Altman’s presence in WWDC 2024 and everything that includes new AI for Apple devices. After recovering from a hangover Keynote addressthe company wanted to explain how your new artificial intelligence system workswhich is based on two new models a language created and trained from scratch.
Why not, some Apple Intelligence features are linked to ChatGPT. does not mean your AI uses OpenAI models such as GPT-4.. In fact, the Cupertino company did not even take advantage of the latter’s potential for create a base capable of offering its functions based on artificial intelligence.
Once this became clear, Apple developed two different models for different purposes. The first one is responsible for run inside devices – iPhone, iPad or Mac – and does not “go out” and does not interact with external servers to perform tasks. That is, everything necessary is provided by the performance of the devices themselves.
The second one does the opposite. This runs on the company’s servers, which, by the way, use thousands of M2 Ultra chips to create a huge data center. All those Apple Intelligence features that require more processing power are solved externally using a different language model than the previous one.
In fact, this difference is one of the few that both models have, since they were trained the same way considering its capabilities. As you can imagine, the model running on devices is much less powerful than the other, although they both follow very similar rules in terms of privacy, data, management of personal information or harmful content.
What data are Apple Intelligence models trained on?
Well, we’ll have to start from the beginning. To create a language model, Apple needs several things, and among them data. The company calls this phase “Pre-Training” and explains a few fundamental things to keep in mind. information you use to provide content to your models.
The most important thing is that Apple does not use any personal information or personal communications in training.. Just check the data online so that nothing is leaked to users. According to the Cupertino-based company, the volume of data is so enormous that it could be accessed to credit cards or social security accounts. But as we already told you, filter all this content to remove it your language model.
In addition, at this stage of data collection, filters for malicious words are also applied, duplicate information and focuses on everything that is considered high quality content.
Post-training
After completing the first phase, Apple realized that Data quality was important to create an effective language model. and with good scalability for Apple Intelligence, so he set to work creating two algorithms that would again check all the information collected and could process the data. That is, still finalizing all the content who already knew the models.
The good news is that both algorithms base their work on human learning and offer human-learnable capabilities that Apple says translate into higher quality language model instructions. Which in turn gives Apple Intelligence more human potential.
Optimization
Once the models have received the necessary data and they have been verified with millimeter accuracy, it’s time optimize them so its performance was quite powerful. Apple focuses on prioritization Speed close up token, achieving a latency of 0.6 milliseconds on the iPhone 15 Pro. In addition, this same device offers generation speed 30 tokens per second.
Both models responsible for moving Apple Intelligence use grouped queries that reduces memory and performance and significantly improves efficiency.
Adapting models to Apple Intelligence features
Apple was tasked with optimizing performance customize the capabilities of two models for the main tasks of Apple Intelligence. And it wasn’t easy since most companies open the door to everything, but when the idea came up to use these models on iPhone, iPad or Mac, it is necessary to adapt them to the tasks they can perform such devices.
These adaptations allow models to have ability to specialize on the fly in very specific tasks, such as email summaries or Siri recommendations, and the ability to scale easily and efficiently.
In short, they managed to create and train two completely new AI language models, as well as The ability to improve the way you express yourself, work, and communicate using Apple products..
Source: Hiper Textual
I’m Ben Stock, a highly experienced and passionate journalist with a career in the news industry spanning more than 10 years. I specialize in writing content for websites, including researching and interviewing sources to produce engaging articles. My current role is as an author at Gadget Onus, where I mainly cover the mobile section.
