Apple의 AI 전략: Siri 업그레이드와 하이브리드 AI 구조

Apple은 개인화된 사용자 경험과 보안 강화를 위한 온디바이스 프로세싱을 강조하는 고유한 AI 전략인 Apple Intelligence Agot을 도입했습니다.

### WWDC 2022에서 공개된 애플의 AI 전략: 애플 인텔리전스 및 주요 기능 소개
* 🍏 Apple은 WWDC 2022에서 고객에게 특별한 AI 경험을 제공하는 데 중점을 둔 **'Apple Intelligence'**라는 브랜드명으로 인공지능 전략을 소개했습니다.
* 📱 Apple의 AI는 주로 **22억 대의 기기와 그 기기에 설치된 앱**에서 작동하며, **10억 명이 넘는 Apple 사용자의 데이터**를 활용하여 개인화된 서비스를 제공합니다.
* 🔊 Apple의 AI 비서인 Siri는 대규모 작업 모델(LAM)로 알려진 여러 앱을 탐색하여 작업을 수행할 수 있는 **'앱 에이전트'로 변신**하는 등 AI 기반의 대대적인 업그레이드를 거쳤습니다.
* 🤝 Apple은 개발자가 **Siri의 앱 에이전트 기능을 Apple 에코시스템 내의 타사 앱으로 확장**할 수 있도록 API를 공개할 계획입니다.
* 🔒 보안 문제를 해결하기 위해 Apple은 대부분의 AI 기능을 전용 뉴럴 코어를 사용하여 기기 내에서 처리하고, 기기 내 용량을 초과하는 작업은 **별도의 프라이빗 클라우드 서버인 'Enday'**를 활용하는 하이브리드 AI 접근 방식을 채택했습니다.
* 🤖 추측과는 달리, Apple은 Siri의 AI 기반에 **GPT-4를 통합**하여 사용자가 Siri의 앱 에이전트를 통해 **GPT-4 응답에 액세스**할 수 있도록 했습니다.
* 🌐 궁극적으로 Apple은 보안을 우선시하면서 사용자의 맥락, 행동, 개인화된 결론을 이해하여 **완전히 개인화된 AI 경험**을 제공하는 것을 목표로 합니다.
* ⚡ 그러나 앱 에이전트로서 **Siri의 응답 속도, 기기 내 처리 속도, 발열 및 배터리 소모와 관련된 문제**와 같은 문제가 있을 수 있습니다.

At Apple's developer event WWDC 202, many people were curious about Apple's AI strategy, which was revealed. Apple quickly touched on its artificial intelligence strategy named Apple Intelligence Agot at the event. Hello, this is Jae-won Lee. At this year's Apple's annual developer event, WWDC 20, the focus was on the keynote of the first day. Among global big tech companies, Apple, which has been evaluated as lagging behind in artificial intelligence technology, was expected to introduce services and technologies related to AI, especially generative AI. As expected, Apple brought up AI during the keynote. Of course, this time they did not use AI, in other words, artificial intelligence. Instead, they presented a brand as Apple Intelligence, showing their insistence on providing not just common artificial intelligence, but a unique brand.

It's also a scene, oddly enough, unlike other big tech companies, there was hardly any mention of artificial intelligence itself in the announcements of artificial intelligence companies. So, there was no talk about how many parameters a certain language model or multimodal model had been developed. Instead, as Apple has always done, they focused on what experiences and benefits Apple's customers could have. So, while watching the presentation, I thought very much like Apple. On the other hand, I didn't feel any sense of significant innovation or performance overwhelming other big tech companies. So, there wasn't a feeling that Apple was coming out swinging. Nevertheless, on the other hand, it felt like Apple, which seemed lost after the debut of Siri, finally found the direction it needed to go. Let's briefly look at the content. So, what Apple has introduced.

What characteristics does Apple Intelligence, or AI, have? First of all, AI operates primarily in Apple's devices, numbering over 2.2 billion worldwide, and the apps installed on them. It operates using the data left by over 1 billion Apple users on these devices. So, what does it do? Apple introduced AI capabilities in terms of language, image, action, and context. In simple terms, it understands human language, images, and actions, and based on this, provides services tailored to individual users' contexts. It understands who you are meeting by analyzing the text in emails and calendars, recognizes who a person is in a photo, remembers what you did with that person, schedules appointments, and informs you of tasks to be done. Ultimately, the core of AI is actively utilizing user data floating within the Apple ecosystem to enhance users' artificial intelligence experiences.

It can be said that one of the things that sets Apple apart is that, compared to MSN and Google, its artificial intelligence performance may lag behind, but it plans to show a high level of artificial intelligence service experience by integrating artificial intelligence in detail and perfectly. So, what changes will Apple bring with this AI? Although they have introduced various functions, there are two main points to pay attention to. I named them in my own way: one is the advancement of Siri, and the other is Apple's unique hybrid AI. First, the artificial intelligence assistant Siri, which appeared 13 years ago, will undergo a tremendous upgrade based on AI. The current Siri is said to receive and process around 1.5 billion voice requests a day. However, the problem is that the content is very flat. So, you have to give specific commands like Siri, set an alarm for 7 o'clock or Siri, call my father, and it only performs that specific task. The artificial intelligence...

It's like a secretary, but it's actually a voice recognition app. However, Siri, based on AI, is completely different. It learns and understands the data stored in devices and apps, and when called, it can find the necessary information on its own and perform tasks. It's like being promoted from a subordinate who just did what they were told to oversee the entire app and report the results as a manager. In Apple's words, Siri is now described as a secretary who understands the contents of the screen. Let's take a look at an example mentioned in the presentation. Let's say a meeting is delayed, and your daughter might have difficulty arriving on time for her ballet performance. If you ask Siri what to do if your daughter is late for her ballet practice, Siri, as well as chatbots like GPT-3, offers very general alternatives. When I actually asked GPT-3, it advised me to calculate the departure time correctly and to check the tickets before leaving the house. This is something anyone can do.

Please tell me what you have. Of course, it may be different if we talk for a long time, but it doesn't help much in the current situation. However, let's see how Siri, upgraded AI, is. First, Siri checks my contacts and text messages to figure out who my daughter is and understands. Then, it checks the poster related to the performance my daughter sent me a few days ago, checks the place and time, returns to my calendar, checks when the current meeting started, where this place is, and what time it is now. It predicts the traffic situation when moving to the performance venue and informs me. Well, the scenes from the presentation end here, but if we imagine a little more, it will automatically send a text message to my daughter saying that I might be late, or if I want, it will call a taxi for me. In the midst of busyness, without the need for me to go back and forth between email apps, messaging apps, and taxi apps, now all I have to do is speak.

An artificial intelligence assistant for apps is called an app agent, and the way they operate is collectively referred to as the Large Action Model or LAM. It is a concept where a single artificial intelligence within the app navigates through various information remaining in the app, instead of individual app AIs, and checks it. An app agent learns behavior patterns and takes actions on behalf of the user by moving through one or several apps, which is the concept of LAM. Although this concept was presented as early as 2024, Apple really showcased it this time. Of course, Apple did not use the terms app agent or Large Action Model, but based on Apple's mention of AI as action-packed in the WWDC invitation, using the term action blockbuster in Korean, one can guess how much Apple is emphasizing the weight of actions that Siri, as an app agent, can produce. By incorporating an action button in the iPhone 15 Pro series released last time, Siri can now be launched with a physical button.

It seems that all of this was prepared for this. In addition, Apple introduced Siri's app agent function. When a situation arises where you need to enter your driver's license information, Siri automatically goes into the mobile ID app, retrieves the driver's license information, and comes back. Also, if you say you need to pick up your mother and go to the airport, Siri checks the messages with your mother, confirms the flight, checks the location of the flight in the real-time flight information app, and informs you of the arrival time of the plane. It also turns on the map app to inform you of the time it will take to get there. This app agent function is expected to mainly operate with apps developed directly by Apple in the early stages, but they are planning to release APIs for developers, so the number of apps that Siri can access within the Apple ecosystem will increase. Siri's capabilities will gradually expand, and the number of actions Siri can perform will increase. From this perspective,

When it comes to the recent collaboration between OpenAI and Apple that has garnered attention, there seems to be some ambiguity. Contrary to the speculation that Siri will replace GPT, Apple has actually incorporated GPT-4 as a basis for Siri. GPT is also integrated as one of the options accessible by the app agent Siri, to the extent that users can receive responses from GPT when requesting a search. Therefore, experts suggest that with this approach, services like entertainment and news, including Claude, may also play a role in Apple's ecosystem. Ultimately, companies with robust platforms that serve as gateways to consumers will inevitably wield greater power. We will update on this matter later. Returning to the point, I understand that Siri based on AI is powerful, but I wonder if there are any security concerns if it delves into all my personal information. To address this, Apple has prepared something.

It is Apple's own hybrid AI structure. Apple reportedly processes most AI functions using on-device AI methods. This includes the latest A17 chip installed in iPhones, as well as the M-series chips in MacBooks and iPads, which contain dedicated artificial intelligence processing cores, namely Neural Cores. This explanation states that on-device AI performance can be sufficiently enhanced without data leaving the device, eliminating concerns about security. However, Apple acknowledges that it cannot perform all AI functions solely based on on-device performance. Eventually, there are times when they need to rely on cloud AI power based on servers. Nevertheless, the cloud still raises concerns about security breaches. Therefore, Apple has established a separate server and cloud specifically for Apple Intelligence's inference operations. This so-called private cloud computer, known as Enday, is also built on Apple's proprietary Apple Silicon. Naturally,

Apple's devices and optimized servers for the data sent from those devices will be built in a hybrid AI structure. In this setup, when Apple Intelligence receives a user's request, it determines whether to process it on the device or in the cloud. If the workload exceeds the capacity of the device, it is simply thrown to the cloud. The server responsible for processing does not store data to prevent personal information leakage. Additionally, access to computational tasks is encrypted, making it impossible even for Apple developers to access. It is said to be a structure where even the entity requesting the computation is unknown. By going through rigorous security measures and borrowing computational power in this manner, they have created a safe artificial intelligence. Technical details will be discussed again when more information is released in the future.

While watching the presentation, I got the impression that Apple has put a lot of thought into what is important for AI to be useful to consumers, beyond just the technical quantifiability revealed in numbers. Ultimately, the answer seems to revolve around context, behavior, and personalized conclusions based on them. Apple needs to understand the most personal conversations and behaviors to provide a fully personalized experience, showing that AI is useful. To delve into the most personal areas of users, they have focused more on security. These features are also provided by Microsoft and Google's artificial intelligence, but Apple, based on a solid ecosystem, seems to be able to exert a stronger force. Of course, there are still many aspects to watch out for on Apple's presentation day. Issues like the response speed of Siri transformed into an app agent are expected, and similarly.

In on-device environments, there may be issues with processing speed, as well as heat generation and battery consumption. We will provide more detailed information on these issues after WWDC. That's all for the stories we prepared today. If you enjoyed today's content, please subscribe and like. Thank you.

WWDC 2024 애플 AI 전략을 설명하는 키워드 ‘액션’

Technology(기술), Trend(트렌드), Transformation(혁신)에 관해 
꼭 알아야 할 지식.  

일하는 사람들의 지적 동반자 티타임즈TV. 

www.ttimes.co.kr

Contact
주소  서울시 종로구 청계천로 11(서린동,청계한국빌딩)
e-mail  ttimes6000@gmail.com
Tel  02-724-7718

티타임즈TV

WWDC 2021에서 공개된 Apple의 AI 전략

Apple은 WWDC 2021에서 고객에게 특별한 AI 경험을 제공하는 데 중점을 둔 Apple Intelligence라는 AI 전략을 소개했습니다.

언어, 이미지, 사용자 행동을 이해하는 AI 기능을 활용해 개별 상황에 따라 개인화된 서비스를 제공하는 것을 강조했습니다.

Apple의 AI는 22억 대 이상의 기기에서 작동하며, Apple 생태계 내 10억 명 이상의 사용자 데이터를 활용하여 고객의 전반적인 AI 경험을 향상시키는 것을 목표로 하고 있습니다.

Apple은 사용자의 장치와 앱에 저장된 데이터를 기반으로 더 복잡한 요청을 이해하고 처리할 수 있도록 인공 지능 비서인 Siri의 대대적인 업그레이드를 계획하고 있습니다.

업그레이드된 Siri는 사용자의 필요에 따라 교통 상황 예측과 같은 구체적이고 유용한 정보를 제공할 수 있으며, 사용자의 딸에게 메시지를 보내거나 택시를 부르는 등의 조치를 취할 수도 있습니다.

이러한 발전은 Siri를 단순한 음성 인식 앱에서 보다 능동적이고 통합적인 AI 비서로 변화시키는 것을 목표로 합니다.

Apple의 앱 에이전트 및 대규모 액션 모델 소개

Apple은 다양한 앱을 탐색하는 중앙 인공지능 역할을 하는 Siri를 통해 LAM(대규모 작업 모델)과 앱 에이전트의 개념을 선보였습니다.

앱 에이전트 기능을 통해 Siri는 여러 앱에서 다양한 작업을 수행할 수 있으며, 향후 GPT-4와의 통합 및 OpenAI와의 협업을 통해 기능을 확장할 계획입니다.

또한 Apple은 Siri의 개인 정보 액세스와 관련된 보안 문제도 해결하고 있습니다.

온디바이스 프로세싱을 갖춘 Apple의 하이브리드 AI 구조

Apple은 전용 AI 처리 코어로 구동되는 온디바이스 AI 방식을 사용하지만, AI 추론 작업을 위해 Enday라는 프라이빗 클라우드도 사용합니다.

Apple은 하이브리드 AI 구조를 통해 워크로드 용량에 따라 데이터를 기기에서 처리할지 클라우드에서 처리할지 결정할 수 있으며, 엄격한 보안 조치와 데이터 암호화를 보장합니다.

이 접근 방식은 보안 및 데이터 프라이버시에 대한 우려를 해결하면서 향상된 AI 성능을 제공하는 것을 목표로 합니다.

소비자를 위한 Apple의 AI 접근 방식

Apple은 기술적인 정량화 가능성을 넘어 맥락, 행동, 개인화된 결론을 강조하며 소비자를 위한 AI에 집중하고 있습니다.

Apple은 개인의 대화와 행동을 이해하는 데 우선순위를 두는 동시에 보안에 중점을 두고 있습니다.

탄탄한 에코시스템의 지원을 받는 Apple의 AI 노력은 마이크로소프트와 구글에 비해 경쟁 우위를 점하고 있는 것으로 보입니다.

그러나 Siri의 응답 속도, 기기 내 처리, 발열 및 배터리 소모에 대한 잠재적인 우려는 여전히 존재하며, 이는 WWDC 이후 더 자세히 다뤄질 것으로 보입니다.