Looking at the end of the year, what is the trend of the surging AI wave? What kind of foreshadowing does it leave for 2025?

Three perspectives and ten major annual trends are clearly and thoroughly presented in the "2024 AI Ten Major Trends Report" released by Qubit Think Tank today.

There is no doubt that we are now in an era deeply affected by all-round changes in AI.

Different from other think tanks and research institutions, Qubit Think Tank is based on Qubit’s long-term understanding and profound accumulation in the field of artificial intelligence. It continues to track the innovation, reshuffle and dynamics of the field in industry, academia and research, and combines the Nearly a hundred start-up companies, research institutes, and investment institutions had in-depth exchanges to outline the current situation of AI and look forward to future trends from the three dimensions of technology, products, and industry.

The report not only provides an in-depth analysis of how this cutting-edge technology iterates technical capabilities, reshapes the business landscape, and leads industrial upgrading, but also has a keen insight into the trend of change and provides a forward-looking outlook on the future path.

This report has also received support from many institutions in the field of industry, academia and research. Not only did the public wisdom on trend nominations, but also shared excellent judgments and comments on specific technologies. This gives the report a broader perspective and a deeper industrial ecological foundation. I would like to express my gratitude——

Now, let’s focus on AI and take a sneak peek at the top ten trends of the year:

Large model Innovation: Architecture optimization is accelerating, and integration and iteration are the general trend. Scaling Law generalization: Reasoning ability becomes the crown jewel, forcing computing and data revolution AGI exploration: Video generation ignites the world model, spatial intelligence unifies the virtual and real AI application landscape: The first round of shuffling is over, focusing on the five major scenarios of the 20th track Competition in AI applications: Operations are more important than technology in multi-field competition, and AI assistants must compete with each other. AI application growth: AI+X empowered products are booming. On the Internet, it is difficult to find popular native AI products. AI product trends: multi-modal launch, Agent sweeping everything, highly personalized AI is ready to transform thousands of industries: the left hand changes productivity, the right hand reshapes the industry ecology. AI industry penetration rate: determined by data basis Initial speed, user demand becomes accelerated AI Venture Capital: The Matthew effect of investment and financing is obvious, the frequency of national team’s shots increases Technical perspective

Large model innovation: Architecture optimization is accelerating, and integration and iteration are the general trend

In 2017, the paper "Attention Is All You Need" was published, and the Transformer architecture came out and gradually became a natural language Process the mainstream technology paradigm in the field. But Transformer is not perfect, and there has always been a voice in the industry, academia and research circles: new breakthroughs are needed in the architecture field to build a powerful and efficient new generation of basic large models.

Who will innovate or even subvert Transformer and replace it?

Since 2023, a large number of innovative large model architectures have emerged, trying to retain the advantages of Transformer whileSolving the problem of too high computing power overhead is expected to achieve breakthroughs in performance and efficiency, posing a strong challenge to Transformer's absolute dominance.

Recurrent neural network-like model (represented by RWKV) state space model (represented by Mamba) hierarchical convolution model (represented by UniRepLKNet) multi-scale preservation mechanism model (represented by RetNet) liquid neural network model ( Represented by LFM)...

A variety of representative technical paths, on the basis of retaining the advantages of the Transformer architecture to varying degrees, combined with innovative developments made by ideas such as RNN and CNN, have also enabled the presentation of large model architectures There is an increasingly obvious mixed trend, More innovative architectures have the characteristics of “drawing on the strengths of others”.

Generalization of Scaling Law: Reasoning ability has become the crown jewel, forcing changes in computing and data

At the technical level, another focus that has attracted much attention is the generalization of Scaling Law.

The first generation of Scaling Law guided model developers to find the optimal solution for model performance among parameter quantities, data sets and calculation amounts, triggering everyone's thinking about the allocation of computing power, data and other resources.

The Qubit Think Tank has observed that the expansion of parameters and calculations has driven the construction and development of my country's Wanka cluster and high-performance networks; at the same time, in the crisis of data exhaustion, the rational use of synthetic data has become a better choice .

In addition, OpenAI o1 is undoubtedly one of the models that have attracted much attention this year, and it reflects a significant improvement in reasoning capabilities. The new Scaling Law represented by o1 prompts large models to pursue higher reasoning capabilities.

A horizontal comparison of Apple Intelligence Foundation, Gemma 2, Llama 3.1, and Qwen2 training methods shows that the proportion of post-training is increasing, and imitation learning + reinforcement learning has become a typical AI development path paradigm.

AGI Exploration: Video generation ignites world models, spatial intelligence unifies virtuality and reality

In 2024, AI technology will continue to make breakthroughs in multiple directions, video generation, world models, embodied intelligence and spatial intelligence and other technologies have promoted human exploration of AGI.

In terms of video generation, the diffusion model has achieved remarkable results in multiple tasks and has become the mainstream technology path for video generation. In particular, the DiT (Diffusion Transformer) model attracts the most attention.

In the field of world models, researchers are committed to developing models that can simulate and understand the real world. The core lies in learning large amounts of data so that new behaviors and decision-making capabilities can naturally emerge from the model.

Inseparable from the world model is embodied intelligence. Since this year, embodied intelligence has gradually moved from concept to reality. Players have launched their first humanoid robots one after another, and at the same time they have begun to improve their dexterity, degree of freedom, control accuracy andMake efforts in sensing technology and continue to overcome technical problems.

Spatial intelligence is a concept closely related to both world model and embodied intelligence. Spatial intelligence refers to the ability of machines to perceive, reason and act in three-dimensional space and time. Its ambition is to combine the ability of spatial computing to control the virtual world with the ability of embodied intelligence to reach the real world.

Product perspective

AI application landscape: the first round of shuffling is over, focus 20 Track Five Scenarios

In order to better observe the current status of domestic products from the data dimension, Qubit Think Tank selected more than 400 representative products for research.

From the perspective of segmented tracks, these 400 products can be divided into 20 categories - AI intelligent assistant, AI companion, AI camera, AI writing, comprehensive suite, AI photo editing, AI video, AI education, AI music/sound effects, AI design, AI mapping, AI search, AI graphics, AI summary and AI translation, each track has produced representative products and then subdivided them, showing different development characteristics.

Among them, AI intelligent assistant is the most outstanding AI native product and the most intuitive embodiment of the technical strength of domestic large-model self-research manufacturers. At present, there has been a clear echelon division within the AI ​​intelligent assistant track, and Doubao has taken a sharp lead.

Although AI companionship has attracted widespread attention, its overall growth is currently sluggish, and there is still a considerable gap between top products such as Hoshino and Catbox and Killer APP.

AI search has become the focus of new business layout, including not only native AI searches such as Secret Tower AI search, but also AI-enhanced searches like Nano Search, Quark Browser, Zhihu Direct, and Xiaohongshu. DaVinci and other business AI searches.

If divided by specific usage scenarios, they can be divided into: full usage scenarios focusing on overall efficiency improvement, work efficiency improvement with optimal overall data performance, creative generation that is expected to have significant breakthroughs in 2025, Leisure, entertainment and daily life, etc., face severe compliance challenges.

AI application competition: The multi-field competition is more about operations than technology, and AI assistants must compete

In order to better restore the current status of domestic AI products, Qubits Think Tank focuses on user scale, new addition speed, user activity and user Data statistics were conducted from the four angles of stickiness.

At present, there is no product on the APP side or on the Web side that can compete with the phenomenon-level breakthroughs in the Internet era, and overall the difference between it and similar overseas products is more than 5 times.

On the APP side, there is currently no product that can produce outstanding performance in all dimensions, and the market lacks scenarios for the birth of killer products.

As of October 2024, a total of 56 products have historical downloads of over one million, 8 products have historical downloads of over 10 million, and the total historical downloads of Quark and Doubao have exceeded 100 million.

From a single monthIn terms of new additions, the monthly growth of Quark, Doubao and Kimi smart assistants can reach tens of millions, and 10 products can reach millions; in terms of DAU, Quark’s DAU exceeds 26 million, and Doubao, Kimi, Tiantian Skipping and Wen Xiaoyan’s DAU exceed one million ; In terms of user stickiness, the three-day retention rate of Quark and Daotao exceeds 30%.

On the Web side, all tracks other than the AI ​​intelligent assistant track are basically at a standstill. Tracks such as AI search, AI writing, and AI mapping have even seen data declines in leading products, or It is a situation of weak recovery after a decline.

In terms of user scale, there are 7 products with total monthly visits exceeding 10 million, including Quark, Tencent Documents, Baidu Library, Kimi Smart Assistant, Wen Xin Yi Yan, Doubao and Tong Yi.

In terms of user activity, a total of 3 products - Quark, Notion and Baidu Wenku have MAUs of more than 10 million, and 19 products have MAUs of more than 1 million. Only 14 products have an average of more than 5 visits per user per month, and 13 products have an average visit time of more than 10 minutes.

Based on data statistics, "Qubit Think Tank AI 100" nominated outstanding domestic AI products through two lists: comprehensive 100 and native 100.

AI application growth: AI + It can be divided into AI native products with AI as the underlying design logic, AI+X products that deeply embed AI functions in original Internet products, shell products based on external API micro-innovation, and multiple products /Model API is a centralized collection of station products.

From the data point of view, due to reasons such as closer integration with business processes and clear demand identification, the current overall data performance of AI+X products is significantly better than that of AI native products, and is based on office performance. Software and content platforms are key layout areas.

For office software, AI writing functions of varying degrees such as continuation, rewriting, and proposition writing, as well as AI summary functions for different themes such as essays and novels, have basically become standard.

Among them, the material library products and editor-type office software whose main business is to provide templates and reference content performed more prominently. Representative products are Baidu Library and WPS AI. Since the effects of AI generation will directly affect the core user experience of the product, such products place greater emphasis on the accuracy of specific functions.

In content platforms, AIGC mostly develops from three directions: AI search based on platform content, AI generation functions and templates used to drive UGC, and portal entry. A step-down content creation tool.

Based on this, Qubit Think Tank put forward three major suggestions for AI-native products: scene integration, simplified user experience, brand trust and promotion.

AI product trends: multi-modal launch, Agent sweeping everything, highly personalized is about to emerge

With large modelsThe ability to process image and video information is rapidly improving, and it is expected that more comprehensive multi-modal interactions will begin to appear in 2025. AI can collaborate through various perception channels such as the Internet of Things and specific information.

Multi-modal input and output make AI more interactive, more frequent, and more applicable to scenarios, significantly improving the overall level of AI products.

Agent, as an intelligent agent that integrates perception, analysis, decision-making and execution capabilities, can proactively provide suggestions, reminders and personalized execution capabilities based on the user's historical behavior and preferences to provide users with Provide highly personalized tasks. The initiative and automation of its interactions far exceed those of existing tools.

From the perspective of the development of both technology and supporting facilities, AI Agent will be widely used starting from 2025. Quantum Think Tank believes that AI Agent is expected to bring interactive methods, product forms and business models that are unique to the AI ​​2.0 era.

From personalized recommendations to directly generating personalized content, AIGC can significantly improve the degree of personalization of user experience, which will help the product further improve the user experience and improve user experience. It can reduce customer loyalty and migration costs, achieve differentiated pricing and further service value-added, which is of great significance to the differentiated competition of products.

Currently, highly personalization based on AIGC has been used in AI education (personalized question banks and teaching arrangements), AI companionship (AI personal assistants and virtual partners), AI marketing (personalized product recommendations, There has been significant progress in the field of personalized generation of marketing content. A number of AI smart assistants equipped on the hardware side have also begun to focus on highly personalized personal assistants.

Industry Perspective

AI intelligently changes thousands of industries: changing productivity with the left hand and changing productivity with the right hand Reshaping the Industry Ecosystem

In the past year, Qubit Think Tank has released a number of in-depth reports to continuously track the implementation and development potential of AI technology in thousands of industries.

Currently, AI presents two major scenarios: AI+ and AI native in industry applications.

In the AI+ scenario, AI mostly appears as a productivity tool and penetrates all aspects of the industry; in the AI ​​native scenario, the industry develops based on AI technology from the beginning.

Qubit Think Tank analyzed the implementation effects of AI in eight scenarios: intelligent driving, embodied intelligence, intelligent hardware, games, film and television, marketing, education, and medical care in the "2024 Top Ten AI Trends Report" and industry characteristics.

p>

In short, the transformation and penetration of AI in the industry deserves great attention, but it can only be distinguished in terms of sooner or later and degree, and there is no debate about whether or not.

AIIndustry penetration rate: The data foundation determines the initial speed, and user needs become the acceleration

In the report, Qubit Think Tank summarized the key to AI penetration into the industry into 3 types of scenarios and 9 major factors to decode the unchanging laws behind the development of the industry.

Take the eight representative industries shown in the figure below as an example:

From the AI ​​industry influence map, we can see that at the current stage, AI has penetrated various industries and caused changes. Three relatively clear ecological niches are presented:

The intelligent driving and embodied intelligence industries in the first echelon have close demand and strong association with AI technology, showing strong correlation.

The second echelon includes marketing, gaming industry, film and television industry and intelligent hardware. The first three use AI technology to reduce production costs, increase efficiency, and deeply integrate workflow; the smart hardware industry is expected to promote industry upgrading through AI technology.

The third echelon covers the basic industries of education and medical care. These industries are actively embracing AI technology with policy support and have higher requirements for security and controllability.

In general, the penetration and transformative power of AI technology in different industries are affected by a variety of factors, among which the industry's data base and user needs may become key factors.

AI Venture Capital: The Matthew Effect in investment and financing is obvious, and the frequency of the national team’s shots has increased

Looking back at 2024, looking at the world, AI is still the strongest money-drawing track.

According to statistics, the total amount of financing in the domestic AI industry has increased, but the number of incidents has decreased year-on-year, reflecting the more cautious and rational attitude of institutions; at the same time, the Matthew effect has become more and more obvious, and capital prefers hot spots and high maturity. track.

Among all the subdivided tracks, intelligent driving ranks first, with the number and total amount of investment events far exceeding other tracks, and many companies The successful IPO has injected great confidence and vitality into the market.

AI+education, AI+games, AI+medical and other tracks have also seen growth in total investment. Survey statistics show that institutions are more technically demanding, have stronger barriers, and are later to reach TPF (Technology- Product Fit) track showed stronger interest.

In terms of policy, due to the government’s long-term concern for AI technology itself and its implementation in various industries, and especially the active promotion of the development of AI-native industries, cities such as Beijing, Shanghai, and Wuhan have introduced a series of policies to attract AI Relevant talent gathering and business landing. At the same time, the national team’s frequent investments reflect policy encouragement and support.

Get the full report

Full report download link: https:/// /jkhbjkhb.feishu.cn/wiki/W5D7wuDcbiPXDLkaRLQcAJpOn8f?from=from_copylink