Standing at the end of the year, what trends are emerging from the rushing AI wave? And what kind of ambush does it leave for 2025?

3 Perspective Dimensions, 10 Annual Trends in Quantum Bit Think Tank's re-release today of theAI Top Ten Trends Report 2024.It is clearly and exhaustively presented in the

There is no doubt that we are now in an era deeply affected by the all-encompassing transformation of AI.

Distinguishing itself from other think tanks and research organizations, Quantum Bits Think Tank is based on thequantum lepton (particle physics)Long-term understanding of the field of artificial intelligence to grasp and deep precipitation, continue to track the field in the session of industry, academia and research of innovation, reshuffle, dynamics, combined with the depth of nearly one hundred startups, research institutes, investment institutions, from theskill,offerings,sectorThree dimensions outline the current state of AI and look forward to future trends.

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

Now, focus your lens on AI, the top 10 trends of the year, and get a first look together:

• Big Model Innovation: Architecture optimization accelerates emergence, convergence iteration is the trend
• Scaling Law Generalization: Reasoning power becomes the crown jewel, forcing changes in computing and data
• AGI ExploreVideo Generation Ignites World Modeling, Spatial Intelligence Unifies Virtual and Reality
• AI Application Landscape: The first round of reshuffling is over, focusing on 20 tracks and 5 scenarios.
• AI application competition: Multi-disciplinary race for operations over technology, AI assistants soldier on
• AI adoption growth: AI+X empowered products go big, native AI pop-ups are hard to come by
• AI Product TrendsMultimodal: Agents are sweeping across the country, and a high degree of personalization is on the horizon.
• AI Intelligence Changes Thousands of Industries: Transforming productivity with the left hand and reshaping the industry ecosystem with the right hand
• AI industry penetration: Data base determines initial speed, user needs become acceleration
• AI Venture Capital: The horse-trading effect in investment and financing is obvious, and the frequency of the national team's strikes has increased.

I. Technical perspectives

Big Model Innovation: Architecture Optimization Accelerates Emergence, Convergence Iteration is the Trend

2017 Attention Is All You Need paper published.Transformer ArchitectureTransformer has been introduced and gradually become the mainstream technology paradigm in the field of natural language processing. However, Transformer is not perfect, and there has been a voice in the industry, academia, and research community that new breakthroughs are needed in the architecture field to build a powerful and efficient new generation of basic big models.

Who will revolutionize or even subvert Transformer and replace it?

Since 2023, a large number of innovative large model architectures have emerged, attempting to solve the problem of Transformer's too-high computational overhead while retaining its advantages, and are expected to achieve breakthroughs in performance and efficiency, posing a strong challenge to Transformer's absolute dominance.

• recurrent-like neural network model(represented by RWKV)
• state-space model(represented by Mamba)
• Hierarchical Convolutional Models(represented by UniRepLKNet)
• Multi-scale modeling of retention mechanisms(represented by RetNet)
• Liquid neural network model(represented by LFM)
• ......

A variety of representative technology paths, in varying degrees to retain the advantages of the Transformer architecture on the basis of a combination of RNN, CNN and other ideas made by the innovative development, which also makes the large model architecture presents theIncreasing trend towards blendingMore innovative architectures are characterized by a "best of all worlds" approach.

Scaling Law Generalization: Reasoning Power Becomes the Crown Jewel, Forcing Changes in Computation and Data

On the technical side, another focus of much attention is theGeneralization of Scaling Law.

The first generation of Scaling Law guided model developers to find the optimal solution for model performance between the number of parameters, datasets, and computation, which triggered thinking about the allocation of resources such as computing power and data.

Quantum Bit Intelligence observed that the expansion of the number of participants and the amount of computation drives the construction and development of China's WANKA clusters and high-performance networks; at the same time, in the crisis of data exhaustion, the rational and good use of synthetic data becomes a better choice.

In addition, OpenAI o1 is undoubtedly one of the models in the spotlight this year, in which a significant increase in reasoning power is reflected.The new Scaling Law, represented by o1, motivates big models to pursue higher reasoning power.

A side-by-side comparison of Apple Intelligence Foundation, Gemma 2, Llama 3.1, and Qwen2 training methods shows that the weight of post-training is increasing.Imitation Learning + Reinforcement LearningBecome a typical AI development path paradigm.

AGI Explorations: Video Generation Ignites World Modeling, Spatial Intelligence Unifies Virtual and Reality

In 2024, AI technology continues to break through in multiple directions, with technologies such as video generation, world modeling, embodied intelligence, and spatial intelligence driving human exploration of AGI.

For video generation.diffusion modelIt has achieved remarkable results in multi-tasking and has become the dominant technology path in video⽣. In particularDiT(Diffusion Transformer) model has received the most attention.

In the field of world modeling, researchers work to develop models that can simulate and understand the real world, centered on learning large amounts of data so that the model is able toNatural emergence of new behaviors and decision-making capabilities.

Also inextricably linked to the world model areembodied intelligenceSince this year, embodied intelligence has gradually moved from concept to realization. Since this year, embodied intelligence has gradually moved from concept to realization, with players launching their own models of⼈形機⼈, and at the same time starting to make efforts in dexterity, control precision, and sensing technology, and continuing to overcome technical difficulties.

(indicates contrast)spatial intelligenceSpatial intelligence is a concept closely related to both world modeling and embodied intelligence. Spatial intelligence refers to the power of machines to perceive, reason, and perform in three-dimensional space and time, with the ambition of combining the power of spatial computing to manipulate virtual worlds with the power of embodied intelligence to reach the real world.

II. Product perspective

AI Application Landscape: The First Round of Shuffling Ends, Focusing on 20 Tracks and 5 Scenes

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

In terms of segmentation tracks, these 400 products can be specifically categorized into 20 categories - theAI Intelligent Assistance,AI accompaniment,AI camera,AI writing,Comprehensive kits,AI retouching,AI Video,AI education,AI Sound/Audio,AI Design,AI⽣图,AI search,AI Tutor,AI Summarycap (a poem)AI translationIn addition, each track has produced its own representative products and then subdivided them, and has shown different development characteristics.

Among them.AI Intelligent AssistanceIt is the most prominent AI product in the category, and is the most intuitive reflection of the technological strength of the country's large-scale self-developed research companies. At present, there is a clear division of echelons within the AI intelligent assistant track, with Doubao taking a breakthrough lead.

AI accompanimentWhile widely popular, overall growth is currently sluggish, with considerable distance still between Top products like Hoshino and Catbox and Killer APP.

AI searchThen it has become the focus of the new business layout, including both native AI search such as Secret Tower AI search, and AI-enhanced search similar to Nano Search, Quark Browser, and business AI search such as Zhihu Direct and Xiaohongshu Da Vinci.

In terms of specific usage scenarios, these can be categorized as follows: full-use scenarios with a focus on overall efficiency, work efficiency with optimal overall data performance, creative development with the prospect of a significant breakthrough by 2025, and recreation and daily life with serious compliance challenges.

AI application competition: multi-discipline race operation is greater than technology, AI assistant soldiers

In order to better restore the current situation of domestic AI products, Quantum Bit Think Tank from theuser scale,Added speed,active usercap (a poem)user adhesionStatistics are presented in four main areas.

At present, the APP terminal and Web terminal have not yet appeared to be comparable to the phenomenal circle-breaking products in the era of the Internet grid, and the overall view and overseas products of the same type are more than five times different.

On the APP side.No product has been able to come up with an all-dimensional bright performance, and the market lacks the scenario to birth a killer product.

As of October 2024, a total of 56 products had more than one million historical downloads, eight products had more than ten million historical downloads, and Quark and Doubao had more than one hundred million total historical downloads.

And from the point of view of single-month additions, the monthly growth of Quark, Doubao and Kimi intelligent assistants can reach ten million, and 10 products can reach one million; in terms of DAU, Quark's DAU exceeds 26 million, Doubao, Kimi, Daily Rope Skipping and Wen Xiaoyan's DAU exceeds one million; and in terms of user stickiness, Quark's and Chattering's three-day retention rate is more than 301 TP4T.

On the web side.All tracks outside of the AI smart-assisted manual track were largely stagnant, and tracks such as AI search, AI writing, and AI graphics even saw headline product data decline, or recover weakly after the decline.

In terms of user scale, a total of seven products with more than 10 million total monthly visits include Quark, Tencent Documents, Baidu Literature Library, Kimi Intelligent Assistant, Wenxin Yiyin, Doubao and Tongyi.

As for user activity, a total of 3 products - Quark, Notion and Baidu Literature - have more than 10 million MAUs, and 19 products have more than one million MAUs. Only 14 products have more than 5 visits per capita per month, and 13 products have an average visit duration of more than 10 minutes.

Based on the data statistics, "Quantum Bit Think Tank AI 100" nominated the outstanding AI products in China through two lists, the Comprehensive 100 and the Native 100.

AI application growth: AI+X-enabled products go big, native AI pop-ups hard to find

Currently, AI products can be categorized into those that use AI as their underlying design logicAI productsIn addition, it is also possible to use the AI function in the original connected grid product, which has been deeply embedded in the product.AI+X ProductsThe external API-based micro-innovationShell productsand the centralized cobbling together of multiple product/model APIsGathering station products.

From the data point of view, due to closer integration with business processes and clear identification of needs, the overall data performance of AI+X products in the pre-processing period was significantly better than that of AI native products, with office software and content platforms as the key areas of layout.

For office software, AI writing features such as language writing, rewriting, and proposition writing, as well as AI summarizing features for different topics such as essays and novels, are basically standard.

Among these, library products, which provide templates and reference content, and office software in the form of editors were the most prominent, with products such as Baidu File Library and WPS AI emphasizing the precision of specific functions, as AI creativity has a direct impact on the product's core user experience.

In content platforms, AIGC is mostly working together in three directions: AI search based on platform content, AI creation capabilities and templates for UGC, and content creation tools to further reduce the threshold.

Based on this, the Quantum Bit Think Tank has proposed an AI-native productScene integration,Streamline the user experience,Brand trust and promotionThree major recommendations.

AI Product Trends: Multi-modal Uploading, Agent Sweeping, High Personalization on the Horizon

With the rapid increase in the power of university models to process image and video information, it is expected that 2025 will see the beginning of a moresynthesisAI is able to collaborate on multimodal interactions through multiple sensing channels such as the Internet of Things and specific information.

multimodalThe inputs and outputs make AI more interactive, more frequent, and with a wider variety of scenarios, the overall level of AI products has increased significantly.

AgentAs an intelligence that combines sensing, analysis, decision-making, and execution power, it provides highly personalized tasks for users by proactively offering suggestions, reminders, and personalized execution power based on users' historical behaviors and preferences. The proactivity and automation of its interactions are far superior to those of existing tools.

From the perspective of both technological and supporting facilities development, from 2025 onwardsAI AgentAI Agent will soon be put into widespread use. According to Quantum Intelligence, AI Agent is expected to bring about interaction methods, product forms and business models unique to the AI 2.0 era.

From personalized recommendations to direct creation of personalized content, AIGC can make the user experience more efficient.personalizedThis will help the product further improve the user experience, differentiate pricing and further add value to services by increasing user loyalty and migration costs, which will be important for product differentiation.

So far, AIGC-based high-level personalization has made significant progress in the areas of AI education (personalized question banks and teaching arrangements), AI companionship (AI personal assistants and virtual partners), and AI marketing (personalized recommendation of goods and personalized growth of marketing content). A number of AI smart assistants on the hardware side have also begun to focus on highly personalized personal assistants.

III. Industry perspective

AI Smart Changes Thousands of Industries: Left Hand Changes Productivity, Right Hand Reshapes Industry Ecology

Over the past year, Quantum Bit Think Tank has released several in-depth reports to continuously track the landing situation and development potential of AI technology in thousands of industries.

Currently, AI in industry applications presentsAI+cap (a poem)AI nativeTwo major contexts.

In the AI+ scenario, AI is mostly characterized by theProductivity toolsroles emerge to permeate all segments of the industry; in AI-native contexts, the industry develops based on AI technology from the very beginning.

The Quantum Bit Think Tank analyzes AI's role in the "Top 10 AI Trends for 2024 Report" in theintelligent driving,embodied intelligence,Intelligent Hardware,play,movies and television,market,teach,medical care8 scenarios for landing results and industry characteristics.

In short, the change and penetration of AI on the industry deserves a high degree of attention, but there is only the sequence of early and late, the degree of severity, and there is no dispute over whether there is or not.

AI industry penetration: data base determines initial speed, user demand becomes acceleration

In the report, Quantum Bit Think TankThe key to penetrating the industry with AI is summarized into 3 types of scenarios and 9 factors, to decode the unchanging laws behind the industry's growth.

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

From the AI industry influence map, we can see that the current stage, the penetration of AI on various industries and the changes triggered by AI, showing a clearerThree ecological niches::

first echelon (of military)The Smart Driving and Embodied Intelligence industries in the middle have a close need and strong companionship for AI technology, showing strong correlation.

second echelonThis includes marketing, the gaming industry, the film and television industry, and intelligent hardware. The first three realize production cost reduction and efficiency and deep integration of workflow through AI technology; the intelligent hardware industry is expected to promote industry upgrading through AI technology.

third tier (e.g. military)It covers education and healthcare based industries. These industries are actively embracing AI technology with policy support and higher requirements for security and controllability.

Overall, the penetration and transformative power of AI technology in different industries is influenced by a number of factors, among which, the industry'sdata basecap (a poem)user needsor become a key factor.

AI Venture Capital: Investment and financing horse-trading effect is obvious, the frequency of the national team's hand improves

Looking back to 2024 and looking at the world.AI remains the strongest money-sucking track.

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

In the various segments of the track, theIntelligent driving takes the leadThe number of investment events and the total amount of money far exceeded that of other tracks, and the successful IPOs of a number of companies injected great confidence and vitality into the market.

AI+Education, AI+Gaming, AI+HealthcareSuch tracks have also seen an increase in the total amount of investment, and research statistics show that institutions show stronger interest in tracks with higher technical difficulty, stronger barriers, and reaching TPF (Technology-Product Fit) later.

On the policy front.Due to the government's long-term focus on AI technology itself and its landing in various industries, especially actively promoting the development of AI-native industries, Beijing, Shanghai, Wuhan and other cities have introduced a series of policies to attract AI-related talents to gather and enterprises to land. At the same time, the national team's frequent investment reflects the policy's encouragement and support.

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