On May 15, Jiazi Lightyear, a think tank in China's science and technology industry, held the "AI Creation Era - 2024 Jiazi Gravity X Technology Industry New Trend" conference at the Renaissance Hotel of Dongsheng Science and Technology Park in Zhongguancun, Beijing. Zhang Yijia, founder & CEO of Jiazi Lightyear, released the theme report "30 Judgments on the New Trend of AI in China in 2024". Details of the report are below.
Welcome to the era of the creation of Jiazi Gravity XAI.
1.AI one day, one year on earth
One day of AI, one year in the world. What would you say to describe the AI industry in the past year in one word?
1.2 A condensed version of a technical hype
Over the past year, we've gone through a condensed version of AI hype.
1.3 AI has entered the most intensive release period in history
Since the release of Sora in February, AI has entered the most intensive release period in history.
1.4 Artificial intelligence has a significant impact on the five major factors of production at the same time
The importance of AI lies in the fact that it has a significant impact on five factors of production at the same time, and the effects of these factors of production are interrelated. Labor creates technology, and technology requires data and capital.
1.4.1 Labor: Polarization and Equality, AI Rewrites the First Principles of Labor
First, let's look at the impact of AI on the workforce. There are two characteristics that stand out, one is polarization and the other is equality.
Polarization: People who use AI and those who don't, those who use AI have higher productivity. For example, employees who use Microsoft Copilot or Github Copilot spend 26 to 73 percent less time completing tasks than non-users.
Affirmative action: Among the groups that use AI, AI is gradually closing the gap between ordinary people and professionals. Both groups of counselors experienced improved performance after adopting AI, with high-skilled (bottom half) participants showing a 16.5% increase, while low-skilled (bottom half) participants showed a 43.0% improvement.
In addition to polarization and equality, AI makes the cost of acquiring knowledge infinitely close to zero, which means that the first principle of labor is changed.
1.4.2 Technology: AI writes into the DNA of all technologies and changes the aesthetics of R&D
Let's look at the impact of AI on technology. The most direct impact is that AI will be written into the DNA of all technologies - AI becomes the technology behind the technology, the tool behind the tools, for example, after using AI, humans have shown significant improvements in protein analysis, virus prediction, weather forecasting, and many other aspects.
This is also driving a collective change in the aesthetic orientation of technology: before it was driven by wisdom, now it is driven by wisdom + resources, from simple is beautiful to "rough calculation" is also beautiful, and the emergence of large models will smooth out the differences in many subdivision technologies, a bit like physics entering the era of large particle colliders.
AI, on the other hand, changes the logic of technological breakthroughs. In the past, technology research and development was a "causal to effect" model, where speculation was first made and then verified by experiments. In the future, technology research and development will become "cause and effect", and AI can emerge a large number of unknown technologies.
In addition, AI is changing scientific research, and it is also promoting cross-border technology and interdisciplinary integration.
1.4.3 Data: from unusable to usable, from finding out to creating
AI makes data elements truly useful. Gartner predicts that enterprise data utilization could increase by 400% by 2025 compared to 2022.
On the other hand, the researchers predicted the growth trend of the size of large model training datasets in the future. The results show that high-quality linguistic data will be used up by 2026. As a result, synthetic data is becoming increasingly important and has become one of the main sources of training data. In 2024, it is expected that 60% of the data used to train AI will be synthetic, and by 2030, the vast majority of the data used in AI will be synthesized by AI.
In terms of data elements, AI makes data go from unusable to usable, from finding out to creating.
1.4.4 Capital: Total investment in AI has declined, and the proportion of generative AI investment has risen rapidly
Then there is the impact on capital. Due to the impact of the capital environment, the total investment in AI has been declining year by year since 2021, and the proportion of investment in generative AI is increasing rapidly.
1.4.5 Land: The digital chain is elongated, the physical chain is shortened, and the information assets command the material assets
The impact of AI on land is more like a substitution role - the digital chain is lengthened, the physical chain is shortened, and the AI-driven world is a world in which information assets command material assets.
1.5 Overall grand consensus, partial non-consensus
The biggest feature of the past year is that a big consensus has been formed: AI is the future. But there are also many cognitions that have not converged, forming various controversies: open source, closed source, how to close the loop of AI commercialization? What are the key indicators of 100 billion parameters, long texts, and multimodal models? Where is the ceiling of the law of scale? Is Transformer the optimal solution? OpenAI, NVIDIA, Is there an Iron Throne? Is there a unique path between the world model and AGI? Will AI get out of control? Some specific issues that we will discuss in today's roundtable discussion will first outline a few salient features.
1.6 Face Up Games Compete with Me Too: Once You Have It, I'll Have It
First of all, AI has become a game of cards – once the first movers have run through and verified, the speed of the latecomers has accelerated. As a result, AI is a lot like a me-too competition: once you have it, I'll have it. Over the past year, the catch-up cycle for AI has been significantly shortening. GPT-3, text generation, 100 billion parameters, Sora, long text, the interval between the release time of the second and first place is getting shorter and shorter. Every time there is a press conference, the hero seating table may change.
1.7 First-mover advantage or late-mover advantage? Low-end subversion: second place is always the highest input-output ratio?
On the one hand, AI seems to have a first-mover advantage: the first mover can have a data flywheel; AI, on the other hand, seems to have a late-mover advantage: latecomers have a more cost-effective input-output ratio.
Sequoia Capital revealed that the AI industry spent $50 billion on Nvidia chips alone last year, but the revenue output was only $3 billion, with a 17:1 input-output ratio - is this number good or bad? On the bright side, as a comparison, SaaS took nearly 10 years to reach this level of revenue, and on the negative side, the business closed loop of AI has not yet been formed. OpenAI obviously likes the first-mover advantage, but many marketers like the first-mover advantage. Fu Sheng told me before that the essence of business is low-end subversion. Does this mean that second place will always have the highest input-output ratio? We can talk about this issue in a moment.
2. Open the era of AI creation
In the long river of scientific and technological development, we are standing at a special historical intersection. On the one hand, remarkable technological achievements are iterating at an unprecedented density, but on the other hand, we are also facing many unprecedented challenges, such as the bifurcation and controversy of technological paradigms, the shortage of computing power and electricity, the question of the unclear "money scene" of AI, and complex issues such as copyright, privacy, true and false information, and ethics.
In a report released in March this year, the Jiazi Light-Year Think Tank proposed the concept of "AI Creation Era". This is a new historical stage, in which the transformation of productivity and the change of production relations occur at the same time, and AI technology penetrates from the digital world to the physical world, gradually approaching and surpassing the boundaries of human production activities, forming a "second wisdom system" in addition to human wisdom. In the era of AI creation, we pay attention to how the technological paradigm converges, how the technological leap forward reshapes thousands of industries, and more importantly, the far-reaching impact of every decision on the future socio-economic structure.
2.1 Two dimensions of AI changing the world: AI subjectivity + AI mapping power
There are two dimensions to how AI can change the world. First, AI subjectivity: from human-led to AI-led, people's dominance has gradually been ceded, and AI has become subjective; Second, AI mapping power: AI's ability to map the physical world has gradually increased, and it has gradually achieved the catch-up and surpassing of human capabilities, from the brain, cerebellum to physical strength.
2.2 There are four stages in which AI will change the world: the era of AI production, the era of AI native, the era of AI creation, and the era of AI civilization
Starting from the two main lines of AI, Jiazi Lightyear Think Tank divides the process of AI changing the world into four stages:
- L1-AI production era: humans are the mainstay, AI is supplemented, and AI's ability to map the physical world is low. AI has triggered productivity changes in various industries, ushered in the "industrial revolution", and greatly improved production efficiency.
- L2-AI native era: AI is the main thing, supplemented by humans, and AI has low ability to map the physical world. The penetration rate of AI will be infinitely approached until it exceeds the boundaries of human production activities in the digital world.
- The Era of L3-AI: Humans are the mainstay, supplemented by AI, and AI has a high ability to map the physical world. AI is gradually infiltrating and approaching the boundaries of people's production activities and behaviors in the physical world.
- L4-AI civilization era: AI is the mainstay, supplemented by humans, AI has a high ability to map the physical world, and human civilization has entered the "twin era".
It is worth mentioning that since the Dartmouth Conference in 1956, the AI route has been divided into symbolism, connectionism, and behaviorism. L1 and L2 are mainly oriented towards symbolism and connectionism, while L3 is superimposed on behaviorism (embodied intelligence). In the era of L4AI civilization, it is a three-stream convergence of three isms.
2.3 What does AI mean at different stages?
What does AI mean at different stages? In the era of L1 AI production, AI means secondary productivity, and the key is to reduce costs and increase efficiency, and promote digital transformation, which is essentially a matter of efficiency and cost. In the L2 AI native era, AI means a second language, a new form of interaction and a content carrier. In the era of L3 AI, AI means the second subject other than humans, promoting the implementation of end-to-end intelligence, the combination of software and hardware, and the implementation of world models. In the era of L4 AI civilization, AI means the second civilization system.
2.4 Xinengbi's next step, the evaluation system for AI to change the world
In April last year, Jiazi Lightyear Think Tank put forward an evaluation index to evaluate the new generation of intelligence: "Xin-Energy Ratio", and recently, Jiazi Light-Year Think Tank has further improved the evaluation system on the basis of Xin-Energy Ratio, adding a new dimension in addition to information and energy: behavior.
Energy, information and behaviour are three basic and interrelated concepts in modern society and the natural world – the progress of science and technology is a reflection of the increased transformative capacity between the three.
Energy and Behavior: Measured by Productivity: In the traditional industrial era, it is the transformation of energy into behavior.
【Energy and Information: Measured by Credit-Energy Ratio】Modern information technology is very dependent on energy. In this transformation process, the process of AI influencing the world is in the L1-L2 stage.
[Information and Behavior: Measured by Information Yield] Information guides behavior. In this transformation process, the ability of AI to penetrate the physical world continues to increase, which is called L3.
With the continuous optimization of AI's ability to map the physical world, a dynamic equilibrium line of AI's influence on the development of the world will be built in practice. We are building a twin civilization in which AI and humans coexist, and this is L4. This basic framework allows us to analyze AI and its impact, opportunities, and sustainability at this moment.
3.30 judgments
From this framework, we can enter into 30 specific judgments.
3.1 L1: The era of AI production
[Judgment 1] The essence of the AI production era is still supply-side reform, which reduces costs and increases efficiency with generalization capabilities
As mentioned earlier, the core of the L1-AI production era is digital productivity, which still belongs to the category of digital transformation. The impact of artificial intelligence on the supply side will be much greater than that of the Internet, and the underlying logic is still to reduce costs and increase efficiency, supply side reform, and the core model is to B. L1 tends to focus on providing new supply capacity on top of old demand, rather than opening up new scenarios.
In the era of the Four Tigers, AI has been criticized for the fact that the investment in customized services is difficult to support the business closed loop of to B, "there are as many customers as there are models"; However, in the era of AI generation, large models have more generalized wisdom through pre-training, so that the investment in customization can be reduced.
[Judgment 2] There is no Iron Throne in the computing power rivers and lakes, and infrastructure and marketization have their own division of labor
In the era of AI production, computing power is the ballast stone of productivity. The biggest proposition here is to solve the contradiction between the supply and demand structure of computing power.
Computing power includes computing power producers, computing power dispatchers, computing power service providers, and computing power consumers. They have their own problems to solve.
First, computing power producers: It is necessary to deal with the structural contradiction between the supply and demand of computing power, and solve the problems of waste of computing resources and low-level repetitive construction. If the construction is carried out in a hurry, it may lead to waste of resources and low efficiency, for example, some intelligent computing centers cannot be used after completion, and the machines have to be turned off. Therefore, computing power producers need to continue to innovate technology, for example, Huang believes that computing power will be 1 million times in 10 years in the future, which is the flag of technological innovation.
Second, computing power scheduler: The demand for computing power is very diverse, so computing power should be interconnected to solve the problem of uneven distribution of computing resources and optimal scheduling.
Third, computing power service providers: to solve the problem of high threshold for the use of computing power. The use of computing power requires technical knowledge and operational skills, and some companies can't light it up, and some can't use it after lighting it; In addition, there are energy consumption and data security issues that need to be addressed.
Fourth, computing power consumers: pay more attention to the issue of cost-effectiveness.
Among these four levels, some are suitable for infrastructure construction, and some are suitable for marketization, and a healthy computing power ecology should be an ecology that performs its own duties. For example, computing power production can be optimized through large-scale infrastructure, while scheduling and operation services are more suitable for marketization.
To give an example, Xingfan Xingqi (Chengdu) Technology Co., Ltd. is specially designed to solve the ecological problem of domestic computing power, and deeply adapts to commonly used large and small models, development tool chains and a variety of domestic chips to ensure that computing resources are fully utilized, improve the inference performance of large models, unify the management of software and hardware and large models, and automatically provide full-process development and application services for large models with one entrance.
It represents a group of startups emerging in the industry – with a core focus on making computing power more usable: integrated delivery, low-cost construction, low barrier to use, and extreme performance.
The composition of the entire "computing power rivers and lakes" is extremely complex and diverse, and there is no "Iron Throne" that can dominate the overall situation - because only the supply of computing power is "all-inclusive" enough to meet the "strange shapes" of computing power demand.
【Judgment 3】AI computing power operators, let computing power really "use"
Here we need to emphasize: making computing power "use" is no less important than making computing power productive.
On the one hand, the diversity of computing power scenarios requires heterogeneous computing power, and on the other hand, there is an urgent need for capability openness and unified management under heterogeneous computing power, especially in emerging business fields such as autonomous driving and intelligent manufacturing, which require flexible and convenient resource matching.
From this figure, it can be seen that computing power scheduling is an indispensable link in the middle of the chain of computing power. This has led to the emergence of a new role: AI computing power operator. AI computing operators use the computing power scheduling platform to balance the supply and demand of computing power, lower the threshold, and improve utilization rate.
For example, be enlightened. As a representative enterprise of HPC+AI heterogeneous computing, Shansi Kaiwu has abundant self-sustaining and operating computing resources, is equipped with high-end GPUs, has advanced networking capabilities, and provides efficient, stable and innovative transmission and computing services through flexible resource allocation, and has completed four rounds of financing. The company has brought together a number of outstanding talents at home and abroad, has a leading domestic Vanka experience cluster networking team, and co-created the "AI Innovation Empowerment Joint Laboratory" with Sun Yat-sen University, relying on its high-performance computing platform to deepen the industrial application of artificial intelligence technology.
【Judgment 4】Let the large model be like a tiger with wings, communicate more (prompt word engineering), read more books (RAG), and practice more (model fine-tuning)
In order for a large model to be used by an enterprise, it needs to be integrated into the existing scenarios of the enterprise. There are three paths to the core: prompt word engineering, RAG (Retrieval Enhanced Generation), and model fine-tuning.
方法1：多沟通——提示词工程（Prompt Engineering）
In many cases, users can quickly construct efficient prompts to solve their own problems. However, when you find that the template you want to build is becoming more and more complex and still not up to the task, that's a sign that you need to introduce RAG or fine-tuning.
方法2：多读书——检索增强式内容生成（RAG：Retrieval-Augmented Generation）
At its core, RAG is about supplementing knowledge with large models. Once any large model is trained, it becomes a static file, and when you ask ChatGPT about its own internal regulations regarding annual leave, it can't answer accurately, and when you continue to talk, it forgets the previous information. When you need to provide more context to your model, you need to use RAG technology. RAG is particularly suitable for those tasks that require a lot of knowledge.
方法3：多练习——微调（Fine-tuning）
Fine-tuning is called fine-tuning because it doesn't start from scratch, but adjusts the model's behavior by continuing training based on a pre-trained base model. This process is very similar to what we call practice makes perfect and the process of drawing inferences. The amount of data used in this process is much smaller than the amount of data required by the pre-trained model, which is basically about 1% of the basic training amount.
These three approaches are not either/or, but work together. For enterprise technology managers, it is important to build a mechanism to enable continuous improvement within the enterprise and continuously approach more efficient, scalable, and economically viable solutions.
【Judgment 5】Synthetic data ≠ high-quality data, and the "self-improvement ability" of the model is the future highlight
As mentioned at the beginning, the use of synthetic data has increased dramatically, but synthetic data ≠ high-quality data. A study published in 2023 revealed that with only synthetic data, the quality of model output may gradually decline as the number of training algebras increases. For example, the image of the face generated on the right gradually shows a strange, hash-marked-like pattern, which severely affects the realism.
This begs a key question: can the model self-improve by generating better synthetic data than its training data? Can models "eat grass and milk"? If the model is able to generate higher quality data than the original training, then the upper limit of this iterative flywheel is opened. The upper limit of self-improvement of synthetic data and its practical feasibility are still issues that need to be discussed.
[Judgment 6] Open source is not equal to free, closed source is not equal to making money, behind it is a supply-side economic account
A while ago, Robin Li made a statement that "open source models will become more and more backward", which further caused the debate on the opening and closing of large models. Open source OR closed source, what exactly is the debate?
From 2019 to 2023, the number of open-source models is greater than the number of closed-source models. In terms of model capabilities, the open-source model Llama 3 released on April 19 has shown industry-leading level in a number of performance benchmarks, and Mistral, Grok, and DBRX in the open-source community have also recently shown the level of "the same generation" as GPT-4. Rapid iteration of technology is gradually reducing the performance difference of open-closed source models.
It is worth emphasizing that open source is not equal to free, closed source is not equal to making money, and now no matter whether it is open source or closed source, it is almost not profitable - the two are not opposites, behind it is a supply-side economic account, and economic sustainability is the essence of this debate.
[Judgment 7] Cloud is not equal to cheap, cloud is not equal to security, a demand-side economic account
On the one-year anniversary of Musk's acquisition of X (formerly Twitter), Team X said one thing: they optimized the way X uses cloud services and moved more workloads on-premises, a shift that has reduced X's monthly cloud costs by 60%. Why does X go to the cloud?
Most enterprises see cloud migration as the key to reducing costs and increasing efficiency. However, many enterprises spend an average of 15% more than budgeted on public cloud and waste up to 27% on IaaS and PaaS. This has prompted enterprises to look for more effective ways to manage their cloud resources.
According to an IBM report, 80% of enterprises have considered or are considering moving workloads already deployed on the public cloud back to private infrastructure. So, is cloud migration still a must-have?
There are many elements of identity determining the position of each family. Going to the cloud is not cheap, and going to the cloud is not the same as security. Behind the cloud and off the cloud is a comprehensive ledger involving finance, strategy, and technology, which enterprises need to comprehensively evaluate in order to make more informed decisions.
【Judgment 8】AI + enterprise management, starting from the construction of a super intelligent management assistant
I just talked about the economic accounts of the supply side and the demand side, so what is the specific delivery model in the L1 era and how does the value land?
We can look at Kingdee. Through Kingdee Cloud Sky AI Platform, AI management assistants and AI applications, enterprises can freely expand, customize, and assemble AI assistants according to their own scenarios and needs, and adapt to their own business needs. At the same time, enterprises can also simultaneously call Kingdee's self-developed large models, open source and third-party cloud vendor large models for training, fine-tuning and optimization, and finally build an integrated enterprise-level AI solution from data to intelligent decision-making.
It's like a super smart management assistant that makes the digital boss possible. The construction of digital intelligence in enterprises is entering the era of "digital boss".
3.2 L2：AI原生时代
3.2.1 Essence
【Judgment 9】 The core feature of AI is end-to-end, and AI is constantly approaching the "shortest path"
The key word in the AI-native era is the digital world, which is dominated by AI.
One of the core features of the AI-native era is end-to-end processing power, which enables AI systems to establish a direct mapping relationship directly from the original input to the final output, without having to go through multiple processing steps in between. In this characteristic, AI is constantly approaching the "shortest path", and software is constantly replacing services. The AI Agent is the embodiment of this end-to-end.
【Judgment 10】Changes in AI native and the underlying logic of the Internet: the future is generated
What are the core changes between AI native and Internet native? Huang said: The future is generated. What does he mean?
In the case of image storage, for example, in the past, the letter "A" was reduced to a collection of pixels, each of which existed in binary form; Now, instead of thinking of the letter "A" as a static image made up of pixels, we see it as a point in a multidimensional space that can generalize and recognize its features, regardless of font, size, or style; In the future, computers will not only understand this multi-dimensional point, but also actively understand the context and meaning of information, not only to see the surface of the image, but to gain insight into the concepts and situations behind it. This process is close to the human way of thinking.
Note that this change has a profound philosophical undertone, where the human brain constructs reality not by storing pixels, but by thinking about "concepts"—a very advanced form of wisdom that philosophers have been thinking about for thousands of years.
In the future, concepts are not static, but a dynamic and constantly evolving process. The future is not stored in advance and searched when it is used, the future is generated and responsive. For example, the biggest difference between the new generation of search and the old generation of search is the contextual awareness brought by long input.
3.2.2 Alternating
【Judgment 11】Slapping the traditional workflow, "prompt interactive" has become a new paradigm in the content industry
This year's little climax is AI-native products for video and music. Hollywood takes an average of 871 days to shoot a film, and the remake of "Terminator 2: Judgment Day", which was released in March this year, is the first feature-length film in human history to be made entirely by AI, and the entire film was produced in only three months; In the world of music production, Suno v3 is able to generate a great song in less than a minute. Workflows have been flattened, prompting interactivity to become a paradigm, which has a direct impact on the content industry.
【Judgment 12】The interaction revolution continues, Prompt is only a phased product, and the best UI is to forget the UI
Currently, Prompt plays an important role in interacting with AI systems, but it is not the best interaction. Why is there such a judgment? Because Prompt isn't simple enough, it's not natural enough. AI products should understand what you want by themselves, if they give a Prompt, they will understand the Prompt, and if they don't give a Prompt, they should refine the intent from your other forms of interaction. Long text may replace fine-tuning, hand-drawn may replace language, and Prompt is more like a phased product.
User-friendliness became the compass of the recent AI release. User-friendliness encompasses personalization, accessibility, but at its core, it's the interaction. The best UI is when you forget the UI, or even the interaction, and focus on the most natural intent and goal. The general direction of human-computer interaction is from process interaction to prompt interaction, and then to unconscious interaction - bidding farewell to the stage of "words do not reach meaning".
【Judgment 13】Real-time is the original soul of AI, replacing the finite nature of space with the infinity of time
AI-native ease of use pursues a minimalist design, with a maximum of two levels of menus, preferably no menus. How to solve complex tasks? Real-time.
There are two profound meanings behind real-time: first, from discrete to continuous, the finite nature of the menu is replaced by the infinity of real-time interaction; Second, from the finite to the infinite, the finite nature of space (graphical interactive interface) is replaced by the infinity of time.
GPT-4o runs much faster, making chatbot conversations much more responsive, with an average response time of 320 milliseconds to audio input, similar to the response time of humans.
Users can ask ChatGPT (powered by GPT-4o) a question and interrupt it as ChatGPT answers. OpenAI says the model provides "real-time" responsiveness.
Coincidentally, a few hours ago, Google released a general-purpose AI agent called Project Astra, an application that uses the viewfinder as the main interface. Google showed a person holding a mobile phone and pointing the camera at various places, and Astra was quick to respond and interact with humans in real time.
Overseas giants are starting to roll up ease of use, which is more like a domestic opportunity. For example, Yixin Technology is an innovator of AIGC real-time rendering application in China, and its first AI flash painting application combines AIGC with traditional image processing and creative design to achieve real-time interaction, real-time design, and real-time rendering, reshaping the whole process of design, greatly improving design efficiency and quality, allowing everyone to become a designer and a typical representative of new quality productivity tools. Today, you can also experience it directly in the exhibition area outside the site.
【Judgment 14】AI To C super product, the first battle is in traffic, and the second battle is in stickiness
From reducing costs and increasing efficiency to a greater industrial revolution, there is a very important watershed in to C. To C is the most provocative to people's imagination, once the technology to C, the next step will be asked: will there be a universal super app? At present, there are basically four categories of super products: personal chat and assistant, search, image and design, and office. However, they face a similar problem: at present, many of them are "daily throwaway" usage, and the monthly retention of top AI products (42%) is not as good as that of top Internet products (63%). Where does loyalty come from? Tools are not as good as emotions; Content is not as good as social.
【Judgment 15】AI native social network, your social object is not necessarily a person
The transition from content to connection is a natural one – AI can generate content, which inevitably affects socialization. Therefore, AI social networking is a track with strong user stickiness and high monetization potential.
【Judgment 16】AI Agent, from single intelligence to multi-body intelligence
The development of Multi-Agent Intelligence (MAI) is an important direction in the field of artificial intelligence, which involves the collaboration, communication and decision-making of multiple agents.
For example, based on the self-developed CarrotAI large model and the original patented digital life technology, Huizhi Intelligence gives the Agent the vitality of continuous learning, evolution and iteration, which can quickly build an exclusive digital employee team for enterprises and build a new paradigm of organizational collaboration and interaction of "employee + digital employee". The interactive objects of digital life will be expanded from monomers to many, and agents will enter the era of multi-body intelligence.
Why develop multibody intelligence?
Many-body intelligence can solve complex problems beyond the capabilities of a single agent. Even if some of the agents fail, the entire system can continue to function. In addition, many-body agents can learn and adapt to each other, make more efficient use of computing resources, and can mimic the behavior of human society. In fields such as biology, ecology, and physics, many-body intelligence can simulate natural phenomena and assist research.
3.2.3 Commercialization
【Judgment 17】AI builder, academic authority is transferred to industrial authority
In 2023, the industry is significantly ahead of academia in the number of famous models released, in stark contrast to a decade ago. At the same time, the flow of AI PhD talents to the industry is also accelerating, and developers have become the most important drivers of AI change.
[Judgment 18] "Use the production model cloud to calculate investment": small companies do small closed loops, and large companies do large closed loops
Users buy products, product purchase models, models buy clouds, cloud buy cards, and at the same time, large manufacturers also have to make investments. If you follow this link to settle accounts, you can see how AI commercialization closes the loop.
Big company, big closed loop. Alibaba Cloud has invested in almost all domestic AI large model unicorns, which is jokingly called "China's large model ETF". Judging from the 2024Q1 financial report, technology giants are the biggest beneficiaries of this wave of AI. For example, 7% of Microsoft's 31% year-on-year growth in cloud revenue is directly attributable to AI technology. According to Morgan Stanley's analysis, AI-driven will effectively promote Baidu's advertising conversion rate, and it is expected that by 2024, AI technology will bring about 3 billion advertising incremental revenue to Baidu.
Small companies small closed loops. The integration of production and model production provides a path for AI commercialization: Mobvoi, which practices "integration of production and model", has become the first AIGC stock in China. The biggest charm of model production and model combination is that it can achieve more thorough end-to-end training, and then form a "data flywheel" effect, and finally realize the update and iteration of models and products that are automatically driven by data. If a company only has a product and does not have some related technologies of the model, it will lose its core competitiveness, but if a company only makes model parameters and does not make products, the technology is likely to be the self-congratulatory of researchers.
3.3 L3：AI创生时代
In the era of AI creation, AI and the physical world are further integrated, gradually penetrating and approaching the boundaries of human production activities in the physical world. From AI for science to manufacturing, from humanoid robots to world models, AI will gradually break through the creation of human beings, and the world model will create a "second intelligence system" in addition to human intelligence. The digital chain drives the physical chain, which in turn influences the digital chain.
[Judgment 19] All things are integrated, and the five forms wrestle with the most powerful AI terminals
In order to achieve large-scale expansion, the focus of AI processing is shifting to the device side. If AI terminals are not popular, it will be difficult to rely on unified cloud AI alone. From last year to this year, almost all the releases of terminal manufacturers are related to AI, AI mobile phones, AIPC, XR and other portable wearable devices, smart cars, humanoid robots, the five forms are the core.
The event of the AI terminal is not only related to the form, but also related to the power consumption. It is not difficult to understand that terminal manufacturers have to make their own chips. It is not difficult to understand what it means for model manufacturers to win a terminal like Apple.
Judging from the latest developments, GPT-4o has aroused widespread attention from the outside world about how GPT-4o and Siri will be integrated, and how ChatGPT and Apple will jointly shape the next generation of AI mobile phones. Jim Fan, Senior Research Scientist at Nvidia, commented: Whoever wins Apple first wins. This will be an AI product with a billion users from the start, and for Apple, OpenAI is like the FSD of smartphones.
Coincidentally, Google has just announced that Gemini is "becoming the new AI assistant on Android". This year's highlight is: whether Google and Apple can play better in the combination of software and hardware.
For terminal manufacturers, this is a battle, in the short term, this means to seize the traffic entrance, driving the demand for replacement; In the long run, Device-as-a-Service, the terminal will gradually evolve from an object to a subject, interacting with people and the external space, which has surpassed the concept of consumer electronics and Internet super platforms.
[Judgment 20] FSD accelerates the landing from car to robot end-to-end
The application of FSD (Full Self-Driving) technology in the automotive field is currently a hot topic, and today, the concept is also moving towards robots. What happens in autonomous driving will happen in robotics. Cars and robots share many similarities: perception systems, decision-making algorithms, path planning, system integration, safety and redundancy, interactivity, adaptability, ethics, and responsibility. Of course, there may be some additional challenges associated with applying FSD technology to the field of robotics, such as the variety of sizes and shapes of robots, and closer interaction with humans.
【Judgment 21】From spatial computing to spatial intelligence, the core is to make virtual space-time conform to the laws of physics
Li Feifei founded a new AI company focused on "spatial intelligence", aiming to make AI perform advanced visual reasoning like humans. From Vision Pro to Li Feifei's entrepreneurship, spatial computing has become a trend. The spatiotemporal properties of the real space physical system follow the objective laws, and the program control of the spatiotemporal attributes of the spatial computing platform needs to consider the premise of objective physical laws. The core proposition of spatial computing is to make virtual space-time conform to the laws of physics. Why conform to the laws of physics? Only in this way can the ability to make predictions and insights into the relationships between objects, from spatial computing to spatial intelligence.
【Judgment 22】Embodied intelligence is the "king of expectations", and the barrel effect determines the speed of landing
Why are embodied intelligence, and humanoid robots so popular? Because we need a super terminal category to carry AI and various technologies to drive the growth of the upstream of the industry.
So, does embodied intelligence have to be more and more "human-like"? Are both feet necessary? Is humanoid a must? Not necessarily. Using metal structures to mimic the bones and muscles of organisms with very different molecular structures is a very sexy direction of biomimetic thinking, but it does not necessarily conform to the Occam's razor principle. At the heart of Occam's razor principle is – don't add entities unless you need to. From a pragmatist perspective, embodied intelligence doesn't have to be human; From the perspective of emotional companionship, humanoid robots don't necessarily need to be overly versatile.
A while ago, I hosted a fireside talk at the Zhongguancun Forum, and an audience member stood up and said that the child is abroad, he is forty or fifty years old, but he is already an empty nester, and he is willing to spend two or three hundred thousand yuan to buy a humanoid robot - it doesn't need to cook, it doesn't need to work, it just needs to look like a human, you can sit with him, and walk two steps at home. Functional and emotional appeals are different from the expectations of embodied intelligence.
But on the other hand, it is worth emphasizing that the development of embodied intelligence is limited by the barrel effect. Software can iterate exponentially, hardware is hard. When AI enters embodied intelligence, there are many technical dimensions, and the slowest iteration of technology will affect the implementation speed of embodied intelligence.
【Judgment 23】The AI scientific revolution, human beings take a step back, and AI takes a step forward
AI is driving a new wave of scientific revolution. For example, the just-released AlphaFold 3 uses a common technique for AI painting - denoising and diffusion model. In one case, AlphaFold 3 predicted how a protein and DNA double helix would hug each other, a prediction that was almost exactly exactly what scientists had painstakingly discovered in reality. Imagine AI doing scientific research like art: starting from a fuzzy "mud" of atoms, carving step by step, and finally presenting a clear molecular structure, even the three-dimensional position of each atom can be accurately given. AlphaFold 3 is better at predicting interactions between molecules than any tool available today — and who would have imagined that drug discovery and art creation could be the same thing in AI's intelligence system.
【Judgment 24】The current AI is very "INTJ", and the world model still has multiple possibilities
The example just now also shows that in the AI intelligence system, the classification of disciplines is blurring. We need to further ask, what are the characteristics of the AI intelligence system? Kidding aside, today's AI is more like the INTJ in the MBTI test: I need energy to interact with people; N: More focus on abstraction than realism; T: More rational than emotional; J: Prefer to make decisions and draw conclusions rather than be flexible.
When people are debating the AI world model, there is a different understanding of intelligence itself. Physical Rule Models, Statistical Models, Surrogate Models, Mixed Reality, Digital Twins and Simulations, System Dynamics Models, Causal Models, Simulated Evolutionary Models, Cognitive Architecture Models...... The choice of technical paradigm behind this is also closely related.
Since Google published its seminal paper in 2017, Transformer architectures have become the dominant paradigm, however, at the edge of the AI research landscape, some teams are working to develop a new generation of AI architectures that are superior to Transformer in different ways. From the perspective of natural evolution, the criterion for measuring the success of a species' evolution is the number of copies of its DNA helix in the world. In the same way, the answer to the question of where the world model will go is from the final application.
3.4 L4: The era of AI civilization
[Judgment 25] Social role: AI has a specialization in the art industry, and the division of labor among humans is flattened
In the era of traditional industrial civilization, society has a clear division of labor, and different industries, different scenes, and different occupations are clearly distinguished. With the application of AI technology, the ability gap between ordinary people and professionals will be flattened, and the division of labor will be flattened.
【Judgment 26】Autologous psychology: The half-life of knowledge is shortened, and the sense of security needs new satisfaction
In 1900, human knowledge doubled roughly every century; In 1945, human knowledge doubled roughly every 25 years; In 1982, it doubled every 12-13 months; In 2020, the total amount of human knowledge doubled every 12 hours.
A century ago, half of what engineers learned when they earned their degrees took 35 years to be overturned or replaced. Now, the half-life of an engineering degree is between 2.5 and 5 years. As the cost of acquiring knowledge approaches zero infinitely, the half-life of knowledge is rapidly shortened.
With the flood of content brought about by generative technology and the limited bandwidth of human beings, where will rapidly outdated knowledge and large-scale information that cannot distinguish between truth and falsehood lead humanity? The civilization of AI has challenged the autologous psyche. Technology is fast, but human nature is slow. People need new security, and these security are increasingly not derived from knowledge.
【Judgment 27】Human-machine collaboration: from centaur mode to cyborg mode
AI has different effects on different people, and white-collar jobs may be at risk of being directly replaced by AI, while blue-collar and green-collar jobs may be more affected by AI-assisted and augmented.
There are two modes of human-AI collaboration: the first, the Centaurs (centaur model): the worker decides which tasks are completed by the AI and which are done by humans, and the AI and the human worker have the second mode of task completion, and the Cyborg model (cyborg model): the human completely integrates their workflow with the AI and continuously interacts with the technology, and the human and the AI's capabilities form a unified system that works
The proportion of people in the digital chain is increasing, and the proportion of AI in the physical chain is increasing. With this process, the cyborg model will become more and more mainstream.
[Judgment 28] economic weight: from R&D to production services, to production to R&D services
AI will change the weight of the economy. In the traditional industrial era, R&D is for production; In the era of digital economy, production serves R&D. For example, Apple has hundreds of parts suppliers around the world, does Apple serve these manufacturers, or do these manufacturers serve Apple? There is no doubt that production serves R&D - Apple's R&D creates soft value, while producers and assemblers only cash on these values, with the former creating 80% of the value and the latter creating only 20% of the hard value, and this "28 phenomenon" has become the norm in the distribution of economic value. Information wealth will inevitably control the creation of global material wealth.
【Judgment 29】In the form of culture, AI promotes the reshaping of human cognition and opens a new round of renaissance
People with the same cognitive concept are gradually forming a new cultural circle, and further promoting the AI version of the renaissance. intergenerational conflicts, discourse segregation, and the exchange of cultures and subcultures...... Many of the things humans do at this moment could become intangible cultural heritage of the future.
【Judgment 30】Twin civilizations: the boundaries are blurred, and humans and AI are "each other"
AI will be deeply involved in the physical and spiritual worlds of human beings, human intelligence and AI will influence and evolve together, and human civilization will enter the "twin era".
4. Anti-speculation
4.1 Does AI really create new demand?
There's a question I've been thinking about lately about AI. There is no doubt about AI's supply-side reform, but what about aggregate demand? If a technology is much more powerful to the production side than it is to the demand side, there is a possibility that the price of related products will fall, and as a result, the relevant GDP will become smaller. The impact of technology on the economy is very complex. Perhaps AI cannot completely avoid the mistakes of the Internet bubble back then, and "more roads and fewer cars" cannot truly complete the closed loop. So it's very likely that we'll experience local economic pains.
4.2 Did humans tame AI, or did AI tame humans?
There is a famous statement in "A Brief History of Mankind": it is not man who tames wheat, but wheat who tames man.
Humans lived by gathering and hunting for 2.5 million years until the Agricultural Revolution about 10,000 years ago. In just 1,000 years, wheat suddenly spread all over the world, and at the same time started a population explosion. It seems to be the story of individuals getting smarter and wiser, people domesticating sheep, growing wheat, growing at sunrise and resting at sunrise, and humans busy sowing, watering, weeding, and shepherding sheep...... But Yuval Harari said something very extreme, saying that the true essence of the agricultural revolution was to allow more people to live in worse conditions.
He said there was no evidence that humans were getting smarter. Far from bringing about a new era of easy living, the agricultural revolution made the peasants live harder and less satisfying than the gatherers. The total amount of food for humans has increased, but farmers have to work harder than gatherers, and the diet may be even worse.
Today, despite all the advanced technology of human beings, more than 90% of the source of food calories is still domesticated plants. The word "domestication" comes from the Latin word meaning "house". But now it's not wheat that's locked up in the house, it's Homo sapiens. The story seems familiar today – from the beginning of the information age, we are increasingly "trapped in the system". Humans have changed the message and have been changed by the message. Further, it is not known whether humans tamed AI, or AI tamed humans.
"From animals to gods" is not necessarily the pursuit of human beings, but it is more like the development direction of AI under human rule. Some people say, then you might as well stop - but last year, we also mentioned the three laws of scientific and technological development, and the development of technology does not depend on the will of individuals. In a free market, maximizing efficiency is a process that cannot be artificially stopped. This leads us to ask whether the nature of life is different from AI.
4.3 Life feeds on negative entropy
In physics, entropy is a measure of how disordered a system is. The Austrian physicist Schrödinger first proposed that life feeds on negative entropy. Life depends on the ingestion of negative entropy from the external environment to sustain and develop. The process of metabolism is the elimination of entropy that the organism has to produce when it is alive through the exchange of "in" and "out".
Life is far from thermodynamic equilibrium. It ingests the energy of the higher forms from the environment, sustains and develops life, and discharges the energy of the lower forms to the environment.
Life feeds on negative entropy. What about AI? The impact of AI on entropy is a complex issue. On the one hand, AI systems can identify patterns and predict trends, thereby reducing data chaos and increasing the orderliness of information. On the other hand, the changes in social structure caused by the development of AI, such as the disappearance of jobs, will increase the complexity of the social system in stages, especially the "black box" model introduces new uncertainties. In addition, AI systems consume energy during operation and increase thermodynamic entropy, especially when the "signal-to-energy ratio" is low.
In short, AI systems reduce some entropy and increase others at different spatiotemporal scales, shifting chaos from one system to another. Of course, from a longer-term perspective, AI will form a new order and structure, but it will also test human social governance capabilities from time to time in the process.
AI adds complexity, and more effort is needed to grasp this complexity. The wisdom generated by GPUs often consumes more GPUs to counterbalance, which puts forward new requirements for energy supply...... This logic will continue to be deduced.
4.4 Prologue to Twin Civilizations: Will AI Be the Moon in the Water?
In the face of the endless progress of the AI, I occasionally think of a story about a monkey climbing a tree to reach the moon...... Oh yes, we're getting closer; Oh no, we're always down to ...... The hardest path to take is perhaps a shortcut, either to heaven or to the Great Pit. Speed is not the yardstick of civilizational progress. AI is our way of fishing for the moon, and maybe it's not the moon.
Above, thanks to the Jiazi Lightyear team, today's big report is here, thank you.