Most of the New Year's predictions have been confirmed, especially last year's hit rate of more than 90%. 2025 Continue.

16 Possibilities for 2024 says that we are at the beginning of four cycles at once. The overall characteristics of 2025 can be summarized by the six points of "deterministic tsunami, multi-domain tipping point, AI's own paradigm shift, all things intelligent, polarization is reflexive, and there is an urgent need for swarm intelligence". Today, I publish my New Year prediction made after continuous pre-research: "50 possibilities in 2025", respectively from five perspectives, such as ecology, business, technology, products, enterprises. They are predictions of what may happen, not a list of what has happened. For reference only.

It is very certain that the application of large models will enter a period of substantial growth in 2025, and more importantly: "certainty" itself has entered a tipping point. Determine that the level is locally available or even highly available, determine that the main technical issues in question are not fundamental obstacles, determine that the big model is not a bubble bet, determine that the AGI Holy Grail will be won, determine that the computing power and energy are actually sustainable, determine that billions of people across thousands of industries will be widely used, determine that AI2.0 will go from text data to the real world, Determine that the flywheel of investment - technology - product - market - application can not only close the loop but also efficiently cycle, and determine that this is a trillion-dollar value industry... All these certainties are making things happen in a new chemical reaction, all parties to strengthen, the storm sense and shock wave of AI2.0 change will inevitably accelerate in 2025, an unprecedented huge energy burst and impact process is just around the corner. If the Internet that started in China 30 years ago is a technology tide, AI2.0 in 2025 is a technology tsunami. Its speed, strength and breadth are far more than the Internet that was cultivated for more than a decade before it began to get on the right track.

Ecological surface: all-round ecological competition, global digital infrastructure, competitive innovation, future and internal and external vitality

Artificial intelligence has entered an all-round ecological competition from 10 levels, and the 8 elements of the new IT end-to-end cloud intelligence have become the infrastructure of globalization and an important cornerstone of the globalization 4.0 order. New IT G8 firms bear the brunt. See my articles "The Six Two-Body Reglobalization" and "The New IT-G8."

The combination of super intelligence and super platform, the competition threshold to tens of billions of dollars, niche advantages and intelligent advantages are mutually enhanced, prompting the rapid polarization of intelligent forces, the difference between the head and the long tail is far greater than the Internet, the relationship between enterprises and the country, society and geopolitics has changed, and the centralism of science and technology has caused alarm.

The problem of post-process weakness and lack of internal force in long innovation may emerge, and the confidence must be enhanced at the same time, 2025 must pay attention to sustainability, and the essence of AI development comes from ecological capabilities and ecological phenomena rather than single point breakthroughs. Ecological innovation as an innovation view into the vision. See my special articles "Root Problem - Source Cognition -6 basis Points: How AI China can Develop sustainably" and "From single-minded Refinement to Full Ecological Competition".

The economic boom and the contraction of primary market investment are deeply transmitted to the industry, and the field of science and innovation needs to be vigilant about ecological faults and negative cycles, and the closed loop of capital finance needs to be opened up, and the closed loop of capital is an indispensable part of the closed loop of business.

The AI creator economy across more than 20 fields has risen in an all-round way, forming a clear contrast with the environmental background color, and becoming a hot spring of vitality stirring China's economy. See my special article "Creators Economic Rise: Intelligent people stir up China's economy, AI gives birth to New vitality for development".

The employment impact is rapidly approaching, and AI is determined to be the first scientific and technological revolution in history that has caused more unemployment than employment creation, and the pain of the software industry and programmers is rapidly approaching other people. Whether ecological imbalances emerge from the impact needs to be closely watched. See my special article "Intelligent Revolution Fast Social Transformation Slow: Human wisdom cannot be lost to the Time Gap".

AI, robots, and automobiles have become the three targets of a new round of China threat theory, and they are also the focus of feeding, and the geopolitical struggle around these three areas is the most intense, and the supply of high-computing automotive SoC may be limited.

The United States may shift from decoupling to ecological inhibition. By changing the supply chain and the relationship between supply and demand, the industrial chain and science and technology ecology cannot form a virtuous positive cycle, thus limiting the upgrading and development ability of China's science and technology industry. In addition to imports, chip exports may come under pressure. With the formation of the double system, the critical point of the positive circulation of the industrial market and the bifurcation point of the big diversion arrive at the same time.

While encouraging development, from the traditional perspectives of taxation, anti-internal volume, anti-unfair competition, anti-abuse of ecological dominance and data privacy, governance may increase leverage, regulate the industrial market and discipline enterprise behavior, and reset the relationship between the industry and other industries. See my special articles "Inside the Truth: Why the platform is so keen on price wars that it moves toward only refunds, but merchants are trapped in traffic" and "What BYD Wants". In addition, black box AI is not a good user experience, and from a regulatory perspective, white box AI is one of the common orientations.

Intelligent reshaping of the sea, AI2.0 enables the new sea, not only improve the scientific and technological content and innovation vitality, and help say goodbye to the traditional expansion paradigm of resource cost price is particularly large, and reposition the new power of the sea in the global industrial chain division of labor and supply chain position, but the thinking concept and cultural spirit reshaping is relatively lagging, and the way to go out in the local survival and development still needs to be explored.

Format: AI is breaking out from the mobile end, Scaling Law continues, and polarization is behind "super"

All parties have completed the confirmation of confidence in AI2.0, the large model has achieved high availability in many types of scenarios, and the investment and application markets have entered the critical point of accelerated expansion. When the theory of slowdown, bottleneck and bubble is greater than ever before, the confidence, determination and strength of the head enterprise are greater than ever before. See my articles "It is very certain that the application of large models will enter a period of substantial growth in 2025" and "6 points of consensus and Trends in the controversy of large Models".

The maturity of real-time voice natural interactive experience encourages mobile phones to take off ahead of PCS, and hundreds of millions of users use it frequently, with significant changes in DAUs - time -UP value.

AI To real, To B Angle fine-tuning, combining RAG and knowledge base, etc., has significantly improved the effect of open source model, significantly reduced the degree of illusion, and put the industry vertical scene into scale and practical.

Some closed-source large-model enterprises will also launch open source models, the open source intensity of the open source model is stepped up, the open source intensity of the high-level model is contracted, and the open source force multiplier is made To the B market, PC and mobile terminals, and the open source vitality is not reduced.

The computing arms race has stabilized but will not significantly slow down, and the amount of computing infrastructure investment of international leading enterprises is more than 10 billion dollars. One hundred thousand cards to hundreds of thousands of cards, and even a million card architecture is also in preparation. Domestic computing power has become the main body of domestic procurement, structural excess and shortage of computing power exist at the same time, and the overall utilization rate of computing power needs to be improved.

Scaling Law has neither failed nor slowed down. The computing power consumption and cost of unit reasoning of LLM unit training have decreased significantly and continuously, but the demand for data and computing power for deep reasoning such as multi-modal, visual understanding, spatial intelligence training and reasoning, and thinking company commander thinking is increasing sharply. The reality percepation-based intelligence of everything will lead to exponential growth of computing power and data. Data depletion has always been a false proposition, leaving synthetic data behind, perception will open up a new world of data. See my articles "No, Scaling Law Not Only Works for a long Time, but the World Is Just Opening Up" and "Is the AI Revolution Running out of Steam and Is Computing Power starting to Overflow?"

With the rapid increase of market concentration, the global market is mainly concentrated in Google, OpenAI, Meta, Anthropic, xAI, and five, and Apple will accelerate the development and iteration of its own end cloud model; The Chinese market concentration process is relatively moderate, Baidu, byte, Tencent, Ali advantages to maintain, Huawei will speed up, the moon's dark side, wisdom spectrum, magic square, IFFY, MiniMax, the wall, zero and one things may seek differentiated development, some enterprises may seek ecological alliance or restructuring cooperation, Baichuan and other continue to turn to the vertical field of professional models. Xiaomi, ideal, glory, etc., may become a new force for change. CV background AI1.0 enterprises cut into AI2.0, the ecological impact and visibility is not high.

At the same time as the super import takes shape, the Apple tax is also getting renewed attention. The globalization tool of multinational companies to penetrate national market barriers may be blunted because of the resistance of partners such as Apps and developers in the ecosystem to the super entrance, and the cornerstone of the mobile ecology through the Store tax model may be shaken. When the application scenario is fully integrated into the super portal, the platform and Apps may begin to conceive of a dynamic and real-time innovative settlement relationship. But 2025 is a bit earlier. See my article "Super Portals Begin to Emerge," "WDC Judge Rules Google violated Antitrust Laws, Not Final."

Solid-state batteries, intelligent driving SoC and other related issues will make all parties realize that there is no moat problem in China's new energy and intelligent automobile industry. See my special article "China's electric vehicles are not without core technology but without moats". The future industrial market leadership mainly comes from the four elements of energy, intelligence, form and ecology. All four elements are still in the iterative process. If you don't get dimensionality up fast you get dimensionality down. A few years later, the use of lithium iron phosphate, ternary lithium batteries and second-hand new energy vehicles without high-level intelligent driving systems will continue to depreciate significantly, which will be a footnote to this problem. Three transformations in the automobile market are accelerating: consumption power/cost performance transformation (consumption needs to upgrade in the spending downgrade), product form/product power transformation (energy across form and level), and brand mentality transformation (new forces in car manufacturing and recognition of domestic cars are rising). It also focuses on the inversion of the scene between the agent and the cockpit. The overall situation of domestic cars is optimistic.

The computing power base point of sapiens has evolved from 1000TOPS to 10000TOS in the long term, and only a few people have reached 1000TOPS at present, but the relationship between personal intelligent device ecology and future computing power has been vaguely visible. See my special article "How much Computing Power Can One Person Have in the Future?" . Everyone will have their own IC, which is the intellectual brain, and the intellectual brain will be everyone's own CP. Mobile phones, computers, intelligent agents, AI assistants, etc., are all part of the intelligent brain. The intellectual brain is the basis for people to become intelligent people, and it is also the standard for everyone to become a superbody individual.

Technicalities: Principle generates the Force, AI paradigm shifts, AGI comes countless times, and no AI Agent is a true agent in 2025

The upgrading of the underlying principle of the large model has become the focus of the head core players, and the improvement of the model cost efficiency is the focus of the running enterprise. The improvement of model prime energy is a linear observation Angle, and behind the change of model principle is the constant reconstruction of the relationship between intelligence and the world, as well as the paradigm shift of intelligence itself, and the principle generates the force. From the perspective of long-term evolution, the LLM grand language model is the first paradigm of the base model, the multimodal foundation model MFM is the second paradigm, the physical world and science model is the third paradigm, the integrated reality model is the fourth paradigm, the autonomous intelligence model is the fifth paradigm, and the real world model is the sixth paradigm. See my articles "Ilya Seems to be stuck in LLM", "Principle Change: Is Pulsed Neural Network a Direction", "The Force Comes from Principle: 6 Characteristics of AI Image Principle Change", and "Model Principle is the" First Principle of AI ".

Aggressive and powerful Chinese large model providers will pay close attention to follow up KAN, SNN, Mamba, TTT, Lory, CLLM, LANISTR, RL4VLM, VLA, LCM, MLLMs, LWM and other principles at different levels and changes in different aspects of technical architecture. Instead of quantization, distillation and other computing power and cost reduction as a breakthrough highlight, see my special article "Four quadrants of intelligent development", "Infield and Outfield, internal volume and external volume".

This year's leading frontier breakthrough will come from the early exploration of LWM in spatial intelligence, VLA in intelligent driving, visual understanding in embodied intelligence, and fusion perception in SICAS agents. Perceptual understanding, intelligent interaction and behavioral ability are the directions of intelligent evolution. See my articles "From Question Answering Intelligence to Behavioral Intelligence, Pool is a New Angle for Generating the future" and "Reality perception is coming, spatial intelligence has a native sensing basis".

The AI Agent that responds to needs, integrates processes and achieves tasks from the perspective of each end user is bound to be hot in 2025. The AI Agent that is fine-tuned, modular, personalized and naturally interactive is the last meter of all AI2.0 products and services, and is also the intelligent radius of each user radiating the whole network. It is the key to deep integration with other applications, especially Internet services, and the formation of super portals by OTT competitors. In the medium and long term, AI Agent is at the strategic crossroads of digital and reality. Major platforms will inevitably invest heavily in AI Agent in 2025, rapidly evolve products, and forcefully occupy the market.

But you won't see a single real Agent in 2025, even though from the beginning agents and AI agents have often been conflated, and almost all AI agents are considered agents. At present, the agent is a nominal pseudo-concept, product packaging and marketing hype means far more than the actual ability itself, but the future agent is not a concept, but the mainstream existence of the most prominent intelligent form. The Agent will be the complete form, or even the ultimate form, of almost all intelligent devices, including AI agents, digital robots, physical robots, and even all intelligent devices in the intelligent scene of everything. The intelligence of all things actually points to the intelligence of all things. At present, almost all AI agents do not conform to the basic characteristics of agents and the behavioral capability model of SICAS.

Infinite Agent is the bridge between AI Agent and agent, which is based on the general basic model as the base of brain ability, based on the understanding of the physical world and the digital world, and based on each user's personalized scenario and needs, to carry out continuous interaction and complete tasks. However, the defects of the infinite agent in real-time perception and the integration of the scene bottom layer still make it quite far from the goal of becoming a digital agent.

Despite the continuous "new actions" that have a strong impact on public opinion, robot 2.0 is still on the eve of a breakthrough, and the basic elements and capabilities of embodied intelligence are still in the process of formation, especially humanoid robots in 2025, but the total volume is limited. Limbs developed and cerebellar ability improved significantly, but the brain ability still needs to be deeply evolved, LLM local high availability and visual understanding will bring short-term gains, and the long-term breakthrough of embodied intelligence is driven by the agent SICAS behavioral ability model, especially the progress of real-time perception and intelligent interaction.

There is not actually a uniform, time-graduated AGI moment, AGI comes in an infinite number of times, there are an infinite number of AGI moments. See my article The AGI Process. AGI moments are different for different abilities, angles, modes, tasks, disciplines, industries, and agents. Waiting for a time-only goal, a general-intelligence grand model, a unique AGI moment, is tantamount to waiting for the rabbit. The year 2025, in terms of the level of availability that can be achieved by different applications, the evaluation of non-known novel tasks with a score of 95 or more and better than human performance in the field can be called the AGI moment in this field. Mathematics and programming may be the first to achieve this goal.

In terms of professional models and scientific models, significant progress will be made in 2025, and the in-depth expansion of different AI4S models in the fields of finance, healthcare, education, law and so on is a fruitful year. See my article "The Paradigm Shift in Science Is Really Happening." Compared with artificial intelligence, biological computing, brain-computer interface, and quantum computing continue to have new highlights, but the overall progress is actually relatively slow.

From an energy perspective, the United States restarts nuclear energy while developing small nuclear power plants, and China partially considers the reasonable allocation and on-site consumption of structural excess wind and solar power resources, but the cost efficiency of energy storage peak regulation is difficult to solve.

Product side: Brain power drives machines such as cars to intelligent agents, mobile phones to super entrances, people to super individuals, new hardware such as AI Glass emerges, network connectivity drives scene changes, and Web3.0 organizational knots

The assembly rate of L2 and above intelligent driving assistance system in new energy listed new cars in the Chinese market may reach 80%, and the total proportion of new energy vehicles such as pure electric-range extension and hybrid in new cars may reach 60%-70% at the annual peak, but L3 capability has become the focus of annual development and the key consideration for users in car purchase. AEB response range and success rate is not high enough short board problems need to be solved, and deep integration with intelligent driving system is inevitable; All AEB L3's that do not actually meet the vehicle specification level standard are pseudo-L3. See my special article "Users and car companies need cognitive alignment".

AI Glass2025 will be on fire, the 100-mirror war is beginning to emerge, and sales have reached the early level of smart watches. The AI Glass with simple non-AR display has a high price and a high sales volume, and the AI+AR AIR Glass with AR display has a high price and a low sales volume. High-cost AIR Glass such as Apple Atlas and Meta Orion are difficult to market normally. The vast majority of products seen in 2025 will have the feel of a transitional form, but it is basically enough to form a marketable usable product. Add the Ray-Ban update for simple display and the AVP update based on M4 or M5 chip, and the experience will be significantly improved. See my articles "20 Signs of AIR Glass in 2025" and "AIR Glass is the main direction of smart glasses" and "AIxAR=AIR".

The change in the user experience and sense of acquisition of the traditional species of mobile phone, 2025 May be a relatively significant year. From the latest actions of Qualcomm, Honor, Huawei, Vivo, Mediatek, Xiaomi, including ARM and other enterprises and Apple-related rumors, 2025 will be a big year for smart phones, and there will be many hot wars. AI is obviously the first battlefield, super entrance is the first battlefield of AI, and the "mouth control" of natural voice interaction is the first battlefield of super entrance. At present, most of them are just similar to the new Siri rather than the initial experience of at least OpenAI's AVM level that fully integrates Apps, and the experience of Doubao is better. AI operating system, end-side intelligence, NPU or AI computing power in the sense of multiple heterogeneity, is the foundation, and SoC is moving to the PC level. Strong intelligence detonates at the C end from mobile devices such as mobile phones. The battle of mobile phone AI starts from the fight for AI core capabilities, and the future will be won by the fight for ecological capabilities. But next year's hot war, not only fighting AI, image, cross-screen, scene, folding screen, especially three folding is also the focus. In my opinion, there is no multiple composite image Raw, no 15+ video dynamic range, no Log or video Raw, no professional mode in place, no high quality 4K120P, no telephoto equivalent to 10x optical zoom, can not be regarded as the true flagship of the image. The shape of mobile phones will gradually change, and users will even have a new species of "refresh" feeling on mobile phones.

The great expansion of AI2.0 intelligent scenario comes from the direct satellite Internet connection of mobile phones without modification, and 2025 is the first year of mobile phone satellite Internet (not the first year of satellite Internet). See the Satellite Internet chapter in my special articles "The First year of Mobile Phone Satellite Internet", "Satellite Internet Global Speed up, Where is China's Progress" and "Black Technology" book. From satellite SMS, satellite phone to satellite Internet access, mobile phones have jumped three times in three years. Starlink is not the only one that has a layout for this, but Chinese companies such as Huawei are still in the testing or test preparation stage, and they have initially reached the networking standard of satellite Internet access requiring pizza antennas. Mobile satellite Internet constellation networking is imperative. The application scenario between the smart terminal and the satellite explosion, satellite navigation, satellite short message, satellite phone can not do, and smart phones can directly connect to the satellite Internet.

Smart car SoC computing power to 2000-3000TOPS, 1000TOPS is only the end-to-end to VLA when the real available L3 computing power starting point. Reusing car computing power to individuals and homes as a computing base will be as important as its mobile energy hub function. See my special article "Robot Quartet: Musk-Tesla- Market - Ecology on Four Rhythms." The end of PC box and box Mini is computing power graining. Mobile SoC computing power is already making its way to pocket-sized PCS. The AI CP (not the AI PC) in your pocket keeps approaching reality. But users don't run around the world in a box, the Mini is still a desktop scene, and is relatively medium performance. The future form of personal computing will be "computing power mobile portable + computing power multiplexing network + cloud computing power", on the basis of personal computing power network and family computing power network, thousands of TOPS-level SOCs of intelligent driving cars will become the available computing power of individuals and families.

The AIization of OS and the evolution of AI to OS have become two paths for AI OS to move in the same direction. AI OS powered mobile phones and PCS bring unprecedented experiences. In this regard, personal focus on Honor and iPhone, the latter is currently experiencing more than 50 points, and it is very possible to increase to 60 or 70 by the end of 2025. See my articles "Is It Worth Getting Started with the iPhone16 for Apple Intelligence?" and "6 Certainties and 6 Uncertainties about Apple Intelligence."

200B is more than 200 billion parameters of the large model can be run in a single consumer AI PC (quantitative degree is not too much), but in 2025 May be limited to the top equipped M4 Ultra chip Mac Studio can do this. See my special articles "Within 100B open source large model personal availability", "123 billion parameter large model Upper body".

The mainstream position of ASIC architecture chips in the inference market was further confirmed, and AI2.0 was confident to confirm that the rapid expansion of AI applications drove the significant growth of investment related to the demand for inference computing power was on the one hand, and the significant increase in the scale of demand significantly diluted the cost of ASIC chip iteration was on the other hand. The latter makes the shortcomings of R & D replacement cost as high as computing power efficiency no longer a restricting factor. This is the fundamental reason why Broadcom and other ASIC architecture service providers are concerned. The leading enterprises with AI computing center, cloud computing platform, data center and other businesses have self-developed or commissioned research and development of their own ASIC chips, TPU, MTIA, Maia, Inferentia2, etc., and those without are also on the way to prepare.

The development of Web3.0 will become a symbiotic phenomenon of AI2.0, which can be understood as the network of elements, production relations and organization of the intelligent world, blockchain is the basic agreement in the infrastructure, and the future is a contract society. See my special article "Bitcoin on $100,000 witness history, 6 points of view". If a local flow interface from a cryptocurrency perspective is not possible, some kind of special zone or experimental area from a Web3.0 or digital finance perspective may need to be explored early. This aspect is not the general direction is unnecessary, but there are too many mouse droppings dedicated to cutting leeks.

The benchmark direction of the evolution of large models in 2025 May mainly depend on o3, LCM and LWM. In particular, the deep reasoning of o3, which is longer and further than the thinking chain of o1, although it requires tens of times more computing power, it also requires a longer waiting time. The o1 Pro has a waiting time of several minutes. See my special article Strawberry Meaning: When AI Starts to Think. The big jump in o3's level in the ARC-AGI test could come from three improvements: One is to build an understanding that is more holistic and dynamically adaptable than the chain of thought (which in turn supports and reinforces the chain of thought), the other is the ability to generate task programs for each task and dynamically adapt them (not just the chain of thought), and the third is the ability to think and supervise thinking, and the mechanism and process are more complex (rather than Mike Knoop's so-called search). The three abilities are superposed, and the ability to solve new problems that have not been learned has made significant progress, but it is still far from reaching the threshold that can be achieved by solving traditional problems, but the architectural direction that opens up, its significance and potential are AlexNet level. In this regard, plane thinking will be regarded as the evolution of the chain of thought, and three-dimensional thinking will be regarded as the emergence of a higher dimensional overall thinking ability than the chain of thought. The evolution of the brain ability of the large model base is a process of the gradual formation of a "point line surface". Point is Token prediction, line is thought chain, surface is cortical calculation, body is the whole thinking. And the element in the brain, from Token to Patch or other representations, is what the DiT-ViT-LCM-LWM people are doing. Its essence is to get out of LLM, out of the box of language symbol knowledge system, out of the limitation of human knowledge - experience - method - thinking - perception, and return to reality itself.

Enterprise side: Through new principles, new forms, new thinking and new problems, see what the world is competing for and what it needs from the underlying logic

New Principle paradigm: Annual focus on enterprises, starting with Li Fei-Fei's World Labs and its LWM Big World model. While not a true model of the world, she once again leads a cutting-edge exploration in the sense of "making the future happen," on a different level than Hinton. World Labs' exploration of spatial intelligence should not just be understood as 3D video generation, but as moving beyond the second paradigm of the base model to explore the early prototype of the third paradigm in understanding the physical world. Understanding the world starts with visual understanding and space-time intelligence. Compared with traditional physics engines, moving to an invisible shape can greatly expand the generalization boundary, and in the long run, it is more efficient and even less computationally powerful than current practices such as graphics workstation physics engine modeling and particle rendering. In terms of understanding the 4D space-time world, DiT and ViT go further than the LLM of pure Transformer, Veo goes further than Sora, and LWM goes further than MLLMs. Such continuous iteration, out of the language symbol knowledge system, deep feeling of the physical world, virtual and reality integration, it is possible to emerge a physical engine that is not a physical engine. If humans can explain equations and calculate models and then AI can do the same, then the meaning of intelligence emergence is lost, and the possibility of emergence is lost. There is the world before perception and data and then understanding, and the underlying law is also based on understanding to find some underlying law, which is also an important feature of the new paradigm of AI4S science. See my articles "Breaking the presets into the discovery and creation of new paradigms is the essence of AI" and "World Simulator is the End of AGI, 12 Situation Prediction".

Focusing on Google, the company most likely to surpass OpenAI in an important area by 2025, Veo2 has actually surpassed Sora. However, search will be the biggest impact of AI in 2025, and the search giant will not be overturned.

Pay attention to Microsoft, because the combination of super intelligence and super platform is the most likely to further cause the platform abuse of dominant position, monopolistic perspective of enterprises in 2025.

Focus on byte, AI, algorithm, traffic ecology rather than TikTok perspective, byte may be the most likely to face greater internal and external pressure at the same time in 2025, Dou Bao can subvert the market position of competitors from the mobile end is also worth continuous observation.

The four-legged four-wheel robot dog product B2-W that brushes the screen of the whole network can already climb the wall, but we want to know what level its brain can do. I will attend CES2024, it is estimated that the human type, four-legged, family, service, logistics robots are the most eye-catching, but the action performance such as somersaults and taking things should not be the focus of this year. See the articles "CES2025 I focus on these 8 categories of product technology" and "Toward the Second Curve: The Future is gradually happening, Why You Still need to re-imagine".

Pay attention to ideal, Xiaopeng, NIO and Huawei, in the process of end-to-end, large model on the car, whose intelligent driving system will show higher prime energy, who will take the lead in making a real L3+, who will brag that he has taken the lead in achieving L4, who can raise the level of the large model to realistic understanding rather than just the dimension of driving visual understanding. For FSD V13 architectures like Transformer+BEV, L3 and even L3+ are already within reach, and everyone can do it with effort. See my special article "L3 ripe, this year, but there may be accidents beyond 12 signals". Therefore, a situation suddenly appeared at this stage, everyone has the opportunity, and it is not necessary to use Huawei. But to L4 may not be able to say, one is the infinite amplification of the possibility of the scene, regulations and society's severity of the safety factor is bound to be much higher than L3, qualified systems and can not pass the system, the safety factor may be a few points of difference. But not only is the technology still far away, so is the public policy. In 2025, the extent to which public roads are open to autonomous driving is bound to be limited. See my special article "Waymo accelerates into the" Singularity ", the pace of autonomous driving in China can not lag." The intelligent part of the car, will vertical integration or horizontal industrial stratification, is essentially at such a crossroads. In the end, there may be specialized skills, and an enterprise cannot do all of them well.

Pay attention to enterprises that dare to do new species of product form, pay attention to Chery Jinyun, Kia PV5, Xiaopeng land carrier these new species. However, the low-altitude economy is small, the resistance is larger than the determination, and the idea is more than the method, especially the new energy manned aircraft may face a difficult situation in the market.

Focus on Alibaba Cloud and its thousand questions rather than magic Square DeepSeek. Including Mistral and llama, individuals have always admired open source large models, but DeepSeek cost efficiency is like the development of robot limbs is not the direction of personal focus, with the data regulated by existing models to feed their own models, the level is impossible to surpass the existing models. The deeper reason I have done a special article analysis, the big model is now the double helix evolution of two curves interwoven. The upward curve, the pursuit of general intelligence to organize the ability of perception and understanding behavior, the brain ability in the sense of overall understanding, thinking and supervised thinking, the overall understanding including vision, reality models such as spatial intelligence, scientific models such as physics, and multi-modes are not a word. Chain of thought, cortical computation, and exploration of different model principles are the core of current frontier exploration. Downward curve, improve data quality, training sense of quantity efficiency ratio, computing power energy efficiency ratio, greatly reduce the cost of reasoning. In this field, there are shortcuts but no corners overtaking, there is a late starter cost advantage but no late starter lead, and AGI routes can only be opportunistic without speculation. The upward curve of the model thinking length, speed, overall degree, multi-modal, perceptual thinking and behavioral ability integration, etc., can only pull the demand for computing power to a new height. The downward curve makes a difference in cost efficiency, but its contribution to the holy grail of AGI is negligible. See "Why DeepSeek-V3's popularity should not be overrated", "I proposed a concept of" dose efficiency ratio ", "Under the four laws of large model development: Who will be more ahead of closed source and open source models in the future". I won't go into details here. The open source large model will also test the ecological internal force and the post-development force. It is hoped that the domestic open source large model can continue to keep up in multi-modal, voice interaction, visual understanding, spatial intelligence and other aspects.

Pay attention to xAI, and hope that there will be a miracle on the future, which can change the open source and closed source competitive landscape at the same time with open source, although the probability seems small. Attention to Nvidia, AI design AI chip stage has come, Nvidia challenger or did not appear? AMD and Broadcom are only flat from different angles, which can not challenge the status, and can not challenge the full SoC layout of NVIDIA from hardware to software, from training reasoning to intelligent cars and robots. However, the investment heat of AI chips continues, and Groq's challenge to Nvidia has no chance, which does not affect a dozen Cerebras. See my articles "The World needs the Next Nvidia", "There is a growing sense of challenger, Just not clear enough", "Long Innovation: The underlying logic of Nvidia's top, Let us see what."

Keep an eye on Meta and Apple, hopefully llama4 will be a surprise, Orion will have an open market version instead of just a handful of developers, and hopefully Apple Atlas will be far more amazing than AVP, which will be updated soon.

Towards 2025: More important than AI is human swarm intelligence, the New three Perspectives, five forces and ten relationships

Through the "50 Possibilities for a Smart World in 2025", we have conducted a panoramic observation and system prediction of AI2.0's ecology, formats, technologies, products, and enterprises. But there is something more important in this world than AI, that is, human collective intelligence - how humans face the certainty tsunami that has come, how to face the impact of the artificial intelligence technological revolution on various fields, and where is the collective intelligence of humans that is more important than artificial intelligence.

AI is definitely the first technological revolution in history that has created more jobs than jobs. When I forwarded this view shared in the circle of friends when I forwarded the "New Zhiyuan" article, Professor Shenyang of Tsinghua University saw it and wrote an AI comment with an instantaneous enhancement mode for one minute. Key issues are refined very accurately, especially the structural contradictions between fast and slow, and AI can see more clearly than people.

The problem is not the technological revolution, but the huge time lag between it and social transformation. Fast and slow alone may be enough to tip the table and not enough time to reason with each other. Has there ever been a revolution in the history of technology and society that was not thrilling? ! Therefore, development must have a sense of scientific and technological innovation, governance must have a sense of market economy, and business must have a sense of social ecology. If you only look at what is in front of you, history will only repeat itself, and there is no solution to this problem.

To the singularity is to some extent to the boiling point. The reshaping of globalization, the reconstruction of international relations, economic competition and downward challenges, geopolitical conflicts, ethnic and cultural contradictions, etc., and the ecological reversal of IT and AI, have become "heaters" that may lead to entropy increase. Human society is like a continuously heated energy body, and the temperature and individual energy levels in the energy ecology continue to rise, and individuals and groups are more active. Traditional organizations, order, rules and relationships are becoming increasingly unsuitable, and the effectiveness of the "social container" to carry and accommodate social activities is facing a test.

This change may require greater anticipation and psychological preparation. In the face of the tsunami of certainty, how to move steadily and far? Five forces affect the trend, one is the innovation and development force of intelligent technology, the second is the awareness (enlightenment) force of enterprises, the third is the constraint and driving force of international and domestic governance rules, the fourth is the gravity of demand, and the fifth is the pressure exerted by individuals and public opinion on intelligent and related enterprises for the protection of rights and interests. In the depth of the future, these five forces need to have a "new three views" of science and technology, innovative future views, scientific and technological values and intelligent development. The new Third View needs to be people-oriented. The development of human-centered intelligence needs to deal with the 10 major relationships in the AI era. 1. The relationship between development and norms:. The relationship between acceleration and alignment: 3. The relationship between concentration and distribution: 4. The relationship between data and privacy: 5. The relationship between open cooperation and ecological protection; 6. The relationship between competition and common development; 7. The relationship between intelligent technology and traditional industries; 8. The relationship between efficiency and equity; 9. The relationship between technology and politics; 10. The relationship between technology and ethics. See my articles "Hinton's Worry: What to Do About a Turbulent Future of Accelerating Change," "Swarm Intelligence Matters More than AI," and "9 Oppenheimer Moments of Superintelligence."

People-oriented group wisdom helps to realize intelligent universal wisdom, make AI a mutually beneficial and win-win technological change, and make this certainty tsunami a common development without entropy increase.

Introduction of the author

Hu Yanping, IT frontier innovation and intelligent technology industry research expert.

Leader of "Global Innovation Frontier Technology Map" and related research projects. He has served as editor-in-chief of Internet Weekly, Director of Exchange Development Center of Internet Society of China and other media and industry organizations, and founded DCCI Internet Data Center and future think tank. Futurelabs Future Lab and many other think tank chief expert or expert. Member of the Information Society 50 Forum. Author of Digital Blue Book and other books, co-author of Black Technology and other books.

For many years, he has been conducting integrated research on the frontier of science and technology, industrial economy, hardware and software product technology services, etc., to realize the organic combination of macro trend insight, mesoscopic research and analysis, and micro data models. He founded a number of widely used research models such as SICAS, proposed a number of new concepts such as the new IT fourth industry, and published several related articles.