The rise of AI Agents: the intelligent future of mobile operating systems
How does AI technology redefine mobile operating systems? This article will give you an in-depth understanding of the rise of AI Agents and their revolutionary impact on mobile phone interactions.
Core content:
1. Pain points of mobile phone users in the pre-AI era and development bottlenecks in the smartphone era
2. How AI technology solves these pain points and promotes the transformation of mobile phone interactions
3. The core capabilities of AI Agents and their application prospects in mobile phone operating systems
1. Development Background of AI Mobile Phones
Pain points in the non-AI and smartphone eras : In the non-AI era, users were submerged in the system ecosystem and had to bear a lot of memory burden. From the simple feature phones (which are actually rudimentary compared to today) to the complex smartphone era, a tsunami of complex functions and services flooded consumers. Mobile phone manufacturers racked their brains to develop many practical but unused or even unused functions for consumers. An important reason is the lack of a bridge for automatic connection between service design and user scenarios. In the smartphone era, most functions still rely on simple machine learning algorithm rules to achieve information retrieval and judgment, which is a low-dimensional function implementation.
2. Core Concepts of AI Mobile Phones
AI's transformation of mobile phone interaction: With the explosive development of AI in recent years, this pain point demand will be met unprecedentedly. AI will become a transportation hub for interaction between mobile phones and users. The development path of AI mobile phones is to move from localized AI functions to service ecology, using Agent capabilities to reach out to all directions like an octopus.
The core of AI mobile phone : Agent and service ecology are integrated. The mobile phone operating system has returned from decentralization to a centralized system with AI as the core. For example, with Apple's voice assistant Siri and Samsung's voice assistant Bixby as the center, users can make basic local function calls and generate answers through multi-modal input.
AI Agent Core Capabilities - Automation Function: The core capability of AI Agent is automation function, which can shorten the user's interaction path. Under traditional operating systems, to create a new appointment and meal schedule, you need to open the calendar app, fill in the information, share the schedule, open WeChat, select contacts, and send it. In actual operation, it is very likely to be interrupted by other operational obstacles, and the interaction efficiency is relatively low. With AI empowerment, only voice input is required, and AI completes the information based on historical information. All actions will be completed silently in the background, and users will be informed at important core interaction contact points. Such use cases will be spread throughout the user usage system. In terms of people's food, clothing, housing, transportation, and personal affairs arrangements, some people have already practiced it, such as Brain AI and J1Assistant by Professor Luo Youhao.
AI's change to the technical ecology of mobile operating systems: The introduction of AI agents will completely change the technical ecology of mobile operating systems. Information integration, search, and screening will be more intelligent and accurate. For example, the filtering of junk information such as notifications, emails, and text messages will become extremely accurate, which is conducive to supervised learning, information classification, priority sorting, etc., semantic and intent recognition plus user dynamic information (location, time, interactive feedback) to achieve more accurate personalized information processing.
3. Specific AI mobile phone product example - J1Assistant
J1Assistant overall situation : J1Assistant is centered around Luo Yonghao, and uses it as a reference point to connect multiple devices to form a complete AI OS ecosystem. The initial version of the experience is not satisfactory. It integrates information tools from Google, ChatGPT, and Perplexity, and realizes the functions of to-do, memo, and schedule.
J1Assistant framework design and functional coverage: J1Assistant is designed to build a self-enclosed AI assistant OS system in terms of basic framework design, with functions such as agent, memo, in-app SMS social networking, information search, Chatbot, etc. It basically covers most of the scenarios in which users search for and generate data. Because of its idol burden, the entire assistant inherits the GUI and interaction logic of the Hammer system, especially the memo subsystem. So far, the Hammer Note has a good user activity.
J1Assistant Current integration and problems: The current integration is still in a preliminary verification stage. Users can accurately distinguish tasks by inputting specific instructions, such as generating to-do items or notes with clear instructions, and the system can accurately generate corresponding results. I use the first version. With the accumulation of users and feedback, the system will be optimized to a certain extent. At present, the interactive entrance is too complicated, which will cause setbacks to the user's interactive link.
J1Assistant search module expansion: In the search module, we can see that the assistant has extended to information retrieval in various fields, including Google, X, Reddit and other different branches, and also extended to e-shopping. I think all of these are entry points from the information layer, and in the future they will definitely go deeper into specific AI agent tasks, not just an information retrieval tool.
Currently, due to the limitations of R&D costs and historical reasons, many functions of Mr. Luo's Assistant are in a semi-automated state. As the capabilities of large models improve, it is expected that the model will be able to identify user intentions and make inferences on its own, automatically giving users the results they want.
4. Cooperation between mobile phone manufacturers and big models
Cooperation trend between mobile phone manufacturers and large models: Recently, Apple and Alibaba reached a non-exclusive cooperation agreement, and Samsung and Zhipu reached a deep cooperation agreement, which brought a great siphon effect to large model manufacturers. In the past year, Alibaba, Baidu, iFlytek, Tencent, Zhipu and domestic mobile phone manufacturers have not reached very deep cooperation. On the contrary, the world's two largest mobile phone brands have chosen to cooperate deeply with domestic model manufacturers, because it is difficult for overseas brands to develop large models that adapt to the domestic local market. I believe that after Apple and Samsung successfully landed large models in China and received positive market feedback, domestic mobile phone brands will continue to follow up and even go overseas.
Details of the cooperation between Apple and Alibaba: According to the latest news, the cooperation between Apple and Alibaba focuses on e-commerce and life services: Shopping experience optimization: Combined with Alibaba's e-commerce ecosystem, it provides users with smarter shopping recommendations and search functions. Life service access: It may access more local life services such as food delivery, taxis, payments, etc. through AI assistants. This cooperation will help Apple improve the integration of domestic mobile phone system tools and service systems. The strong combination will distance itself from other manufacturers and provide a model for other brands. AI drive is a connection hub. Without AI drive, many services cannot be connected.
Samsung's cooperation with Zhipu and related product functions: Similarly, Samsung Galaxy S25 uses the capabilities of Zhipu Agentic GLM, and the "Voice Chat Vision" supports calling system-level functions through voice in the FunctionCall mode, which can expand the personalized usage scenarios of a series of personal assistants such as schedules and travel route planning.
5. Deepseek model analysis and application
Deepseek's influence and follow-up: Deepseek has become a phenomenal success, both overseas and domestically, which has given great support to mobile phone manufacturers to integrate the Deepseek model. Huawei, Honor, OPPO, Meizu, and VIVO have all quickly followed suit. It is quite possible to take Deepseek overseas in the future, and Luo Yonghao's AI assistant R&D team may also be actively following up on the integration of Deepseek.
DeepSeek V3 performance analysis: DeepSeek V3 performs well in difficult academic and reading comprehension tests (such as MMLU, DROP, MMLU-Redux), even surpassing GPT-4o and Claude-3.5. It performs relatively poorly in code-related tasks (such as Codeforces, LiveCodeBench) and some tasks such as SimpleQA. Overall, it is at the leading level in difficult tasks such as language understanding, reasoning, and mathematics, but there is still room for improvement in some open domain question-answering tasks. For example, comparing Deepseek v3 with GPT-4o:
SimpleQA (Correct): GPT-4o leads by 13.3, indicating that GPT-4o is more accurate in basic open question-answering scenarios (such as user questions - direct answers). FRAMES (Acc.): GPT-4o leads by 7.2, indicating that GPT-4o is more stable in multi-round dialogue scenarios.
For mobile phone AI OS systems, there are many open-domain question-answering scenarios, such as information query (search engine, news, encyclopedia), social communication (text messages, social media, instant messaging), entertainment (music, video, games), life services (navigation, takeout, shopping, payment), AI assistants (Siri, Google Assistant, Xiao Ai, etc.). However, in the current open source model field, DeepSeek can meet the needs of real mobile phone AI OS user scenarios. When compared horizontally with the model used by Xiao Ai, DeepSeek has a crushing advantage.
DeepSeek model size and application : It is not enough to talk about the model size without talking about the actual scene application. The parameter scale of the DeepSeek-R1 series models ranges from 150 million (1.5B) to 671 billion (671B), of which the largest model has 671 billion parameters. In contrast, Xiaomi's self-developed large model MiLM mainly has two versions with 1.3 billion (1.3B) and 6.4 billion (6B) parameters. Xiaomi has done a lot of optimization work in scene optimization and local deployment. In the short term, DeepSeek is still a long way from being deployed as a local model on the mobile phone. At present, some major brands have placed Deepseek as the entrance to the intelligent body in the mobile assistant. There are also some flagship products like Oppo that deeply integrate the full-blooded version of deepseek into the assistant, which can be directly invoked by voice.
The strategic dilemma of mobile phone manufacturers in applying DeepSeek: A mobile phone brand's product line covers a variety of models from low-end to flagship high-end. It hopes that the AIOS system will cover all models, but there is a contradiction that cannot be reconciled in the short term. Although open source models with strong performance such as DeepSeek have been launched, low-end models that want to be AI-enabled can only be completed by using ready-made cloud API + rule matching methods based on low hardware configuration. Over time, costs accumulate and benefits decrease. Therefore, most manufacturers currently use the strategy of flagship phones + data sets + model optimization to build AIOS. Although the initial investment is large, the cost will be reduced in the later stage, and the high premium of flagship phones can cover part of the local model training cost investment.
In summary, with the continuous improvement of model capabilities and the continuous expansion of application scenarios, the development and evolution of AI operating system (AIOS) will be further accelerated. The value brought by AI is far more than giving users the ability to automate operations. More importantly, it greatly enhances users' actual ability to solve problems, enabling users to complete tasks that previously required a lot of energy in an efficient way while saving time and effort, significantly improving user experience and life and work efficiency. The reduction in model price and size is believed to benefit more user groups.
