Top 10 AI Events in 2025

"By 2026, AI may be smarter than the smartest humans. AI and robots will be like a supersonic tsunami, instantly sweeping away all jobs." This is the latest warning issued by Elon Musk to all of humanity. In just three years, AI has undergone rapid and radical changes.

If 2023 was the year of initial exploration of AI, when people were just having a taste of the novelty; 2024 was the year of confusion, when enterprises were still looking for suitable applications; then the just - passed 2025 was definitely a turning point in the history of AI.

AI is no longer just a mascot for chatting. It has grown "hands and feet", picked up tools, and even become your boss. On the model side, DeepSeek challenged GPT - 4 with a cost of less than $6 million, tearing off the veil of the money - burning myth in Silicon Valley. On the application side, Manus and Doubao started to take over your computers and phones, and AI Agents became real digital employees. In the physical world, Ubtech's robots entered Tesla's factory and began to screw bolts with tools of 0.1 - millimeter precision. These are not only technological iterations but also a comprehensive reshaping of production relations. Facing the passing 2025, we still have many questions. Who is swimming naked? Who is leading the race? Will your job still be there next year? What dividends are there in this transformation?

Today, by reviewing the top ten events in the global AI industry in 2025, we will take you through the real changes that took place in the AI industry this year.

1. Efficiency Revolution and China's Breakthrough in the First Half of the Year

DeepSeek - R1 shocked Silicon Valley and completely rewritten the rules of the money - burning game for large models.

The first bombshell at the beginning of the year came from China.

In January, DeepSeek released the open - source model DeepSeek - R1. This release immediately made the bigwigs in Silicon Valley sit up and take notice. Why? Because it broke the unspoken rule that "large models must burn a lot of money".

Before the emergence of DeepSeek, large models in Silicon Valley followed the path of closed - source dominance, frantically piling up parameters and computing power. The training cost of a single model was basically in the tens of millions to hundreds of millions of dollars. They formed high - capital barriers by throwing money around. Whoever was willing to spend money could take the lead.

What about DeepSeek R1? Its training cost was only $5.576 million. What does this mean? The training cost of GPT - 4 was about $100 million. That is to say, DeepSeek used only one - twentieth of its money to create a model with comparable capabilities. It's like a poor boy assembling a Ferrari with bicycle parts and finally outperforming the regular army. Behind this is not only extreme architectural innovation but also a counter - attack and breakthrough for Chinese AI.

In the past, Americans were used to the "brute - force creates miracles" model, thinking that AI was a game for the rich and a nuclear weapon. But DeepSeek turned AI into a submachine gun, cheap, useful, and accessible to everyone. Through the efficiency optimization of algorithms, it directly dispelled people's panic about computing power monopoly and instantly reduced the application threshold of AI by two orders of magnitude. From then on, large models were no longer offerings on the altar but clay in the hands of developers.

Manus initiated the Year of the Agent, and the human - machine collaboration model underwent a qualitative change.

In March, a Chinese startup suddenly launched the AI agent Manus. As soon as the news came out, the A - share AI agent index soared by more than 6% in a single day, and a large number of related concept stocks hit the daily limit. Why was the capital market so crazy? Because Manus is the world's first truly general AI Agent. This can be seen from its name, which means "using both hands and brain" in Latin.

So what makes Manus so amazing? In the past, AI was just about answering your questions, talking but not taking action. But Manus can break down tasks on its own, control the mouse on its own, and operate across different software on its own. For example, if you just tell it "Help me do a competitor analysis", it will open the browser to search for information on its own, filter useful information on its own, open Excel to fill in data on its own, and finally generate a chart and send an email to you on its own. Throughout the process, you just need to sit back and watch it work while sipping your coffee. This is another update to the human production mode.

AI has moved from "passive Q&A" to "active execution". In the future, what you hire will no longer be a real person or a SaaS software but a digital employee without emotions and working 24/7. For business owners, this is a magic weapon for cost reduction and efficiency improvement. But for white - collar workers who only do basic execution tasks, this is a real occupational crisis. Because Manus has proven that for any repetitive mental work based on rules, AI can do it faster and cheaper than humans.

This is a milestone in the evolution of AI from a "conversation tool" to a "digital employee", and the human - machine collaboration model has undergone a qualitative change.

Qwen3 reached the top of the open - source throne, forcing closed - source giants to re - examine their moats.

In April, Alibaba released the open - source large - model series Qwen3. It creatively adopted a hybrid reasoning architecture of "fast thinking + slow thinking". Simply put, for simple questions, it uses its "cerebellum" for a quick response, and for complex questions, it uses its "brain" for in - depth thinking. It has strong performance and saves computing power. This mechanism directly reduced the reasoning energy consumption of the model by 60%. The most important thing is that it is free and open - source. Both individuals and enterprises can deploy it locally and conduct secondary development, enabling small and medium - sized teams and developers to access top - notch capabilities. On various authoritative evaluation lists, Qwen3 not only dominated the open - source field but also surpassed the then - closed - source GPT - 4 Turbo in hardcore indicators such as code writing and mathematical reasoning.

At this point, Chinese open - source models have occupied half of the global community. We can confidently say that open - source AI has officially entered the Chinese era. We are no longer followers. In terms of efficiency optimization and application implementation, we are already leaders. This has also forced closed - source giants such as OpenAI to re - examine their moats. When free open - source models are better than paid closed - source models, does their business model still hold?

This is a leading position of Chinese open - source models in the global technological high - ground, forcing closed - source giants to re - examine their moats.

Llama 4 redefined "infinite memory", and large models overcame the barriers for B - side business.

Immediately afterwards, in May, Meta open - sourced the Llama 4 series. The most terrifying thing about this generation of models is not the parameters but the memory. It supports a context window of up to 10 million tokens. What does this mean? A copy of "Dream of the Red Chamber" is about 700,000 words. It can swallow a dozen copies of "Dream of the Red Chamber" at once, or directly read all the financial statements, technical documents, and meeting records of an enterprise over the past ten years and remember them all. At the same time, it adopted the MoE (Mixture of Experts) architecture. You can think of it as changing from a general practitioner to a group of specialist consultations. Whoever is good at a particular task takes over, which also directly tripled the reasoning efficiency.

The release of the Llama 4 series represents a qualitative change in B - side applications. In the past, AI had a poor memory and would forget what was said earlier during a conversation, making it difficult to handle complex enterprise business. Now, Llama 4 has almost infinite memory. When an AI can remember all the historical data, codes, and rules of your company, it is no longer a simple chatbot but a senior expert who can truly understand business logic. In this round, Meta has indeed won a point for the United States in the open - source field and completely opened the door to enterprise - level AI applications.

2. Hardcore Infrastructure and Ecosystem Implementation in the Second Half of the Year

Nvidia's "European Computing Power Expedition" sounded the clarion call for the global battle for computing power sovereignty.

In June, Jensen Huang announced a crazy plan at the GTC Paris conference: Nvidia would build more than 20 AI factories in seven countries, including Germany and France, all at once. The goal was clear: to directly increase Europe's AI computing power by ten times before the end of 2026. For this purpose, they even brought in Airbus to cooperate, providing customized computing power services specifically for aircraft design. This news directly drove Nvidia's stock price up by 24% in a single month, and its market value approached $3.5 trillion.

The global computing power has witnessed a new "land - grabbing" movement. In the past, computing power was concentrated in North America. Now, computing power must be localized. Why? First, for data security, no country wants its core data to flow out of its borders. Second, for low latency, the response speed of AI should be as fast as turning on a light switch. Of course, the most core reason is to compete for computing power sovereignty. In the face of uncertain geopolitics, computing power is the new oil. Not only Europe, but Saudi Arabia is also investing $5 billion in building data centers, and Japan and Singapore are frantically hoarding graphics cards. Whoever has computing power locally has the power to define the future.

The successful mass - production of the Ascend 920 chip officially announced that domestic computing power has moved from being usable to being good - to - use.

In July, Huawei officially announced the mass - production of the Ascend 920 chip. Many people may not know the value of this chip. It is an NPU (Neural Processing Unit) neural network processor specifically designed for AI computing. The Ascend 920 not only uses advanced manufacturing processes but also has a single - card computing power of 8P Flops. Its performance in handling specific AI tasks has reached 120% of that of Nvidia's H100. More importantly, it has achieved the ultimate in energy - efficiency ratio. Based on it, the cost of training the DeepSeek - R2 model was reduced by 35%.

Under the successive blockades of the United States, the mass - production of the Ascend 920 is not just a domestic substitution but also a strategic counter - attack for "de - Nvidia - ization". By building a full - stack ecosystem of "chip - framework - application", Huawei has completely unblocked the "meridians" of domestic AI. This means that China's intelligent computing centers will no longer be held back, and we have our own core computing power base. This is not only a technological breakthrough but also a hardcore guarantee for national competitiveness.

Ubtech's Walker S2 entered the factory to work, and a new species of productivity emerged for Made in China.

In October, Ubtech announced the mass - production of the humanoid robot Walker S2 and its batch delivery to Tesla and CATL factories. This time, the humanoid robot didn't just dance and do somersaults; it really entered the factory to work. The Walker S2 is equipped with an end - to - end large model, and its visual recognition and hand - operation coordination have reached an unprecedented level. The assembly accuracy has reached 0.1 millimeter. On the assembly line, it can, like an experienced master, accurately complete a series of tasks such as labeling, quality inspection, and handling. It solves the problems of those high - risk, repetitive, and high - precision - required positions, and it doesn't eat, sleep, or complain. It offers the best cost - performance ratio.

If last year was the year of the explosion of embodied intelligence, then 2025 was the year of implementation. When AI has a body, it steps out of the screen and starts to create value in the physical world. This means that China's powerful manufacturing industry chain will achieve the "integration of soul and body" with the most advanced AI brains. When robots can be as flexible as humans and as precise as machines, Made in China will truly move towards Intelligent Manufacturing in China. This is our strongest card to deal with the aging population and maintain the competitiveness of the manufacturing industry.

Google Gemini 3 broke the limit of multi - modality and opened a new era of in - depth reasoning.

In November, Google officially released Gemini 3, refreshing the record on the list with a high score of 1501. Its biggest breakthrough lies in "in - depth reasoning" and "multi - modality understanding". What is in - depth reasoning? In the past, AI was retrieval - based, but now it is thinking - based. It can not only understand complex financial statements but also directly generate visual reports with dynamic charts and data analysis. And multi - modality means that its skills have been enhanced. In the past, AI could only describe pictures, but now it can simultaneously process different forms of content such as text, pictures, voice, video, and tables, and connect the relationships between them.

From just chatting and drawing to now being able to think independently and produce results, the IQ of AI has reached the level of a human doctor or even higher. This also indicates that AI will make a large - scale foray into high - end professional fields such as scientific research, finance, and healthcare from the entertainment and media fields. In the future, the best assistant for scientists may no longer be graduate students but "AI scientists" like Gemini 3, which will accelerate the speed of human exploration of the unknown.

Doubao Mobile Assistant redefined smartphones, and the traditional APP ecosystem was completely subverted.

In December, ByteDance quietly released the Doubao Mobile Assistant and the Doubao phone in cooperation with Nubia. Priced at 3,499 yuan, it was sold out as soon as it hit the market. Why was it so popular? Because it completely subverted the traditional perception. With the Doubao phone, APPs may completely disappear in the future. What can it do? It can answer harassing calls for you and elegantly retort. It can sort out thousands of unread messages in a WeChat group and generate summaries. It can even help you grab red envelopes, tickets, and order takeaways. It's like a butler living in your phone with the highest authority.

If Manus in the first half of the year was a revolution on the computer side, then Doubao is a subversion on the mobile side. It announced the end of the "APP era". In the past, when using a phone, people had to adapt to APPs. In the future, APPs will adapt to people. In the future competition of smartphones, it will no longer be about how many pixels the camera has but about how well your AI assistant understands you and how capable it is. Hardware will become the carrier of AI, and AI will become the soul of the phone. This is a complete "dimension - reduction strike" for the Apple and Android ecosystems.

AI Agents got rid of information silos and entered the "HTTP moment" of the Internet of Everything.

The last major event at the end of the year was that tech giants such as Google, Microsoft, and IBM jointly established the AAIF Foundation and established the "AI Agent Interoperability Protocol".

This protocol is specifically designed to solve the problem of system fragmentation. Simply put, current AI agents operate independently, cannot cooperate with each other, and are easily tied to a single vendor. It's expensive and troublesome for enterprises to use them. It's like everyone speaking their own dialects. The signing of this protocol is to unify the "Mandarin" in the AI world.

This is like when the Internet established the HTTP protocol. The HTTP protocol allowed computers around the world to access each other through web pages, breaking down information silos and creating today's Internet. The AAIF protocol solves the problem of language barriers between different AIs. From now on, AI Agents will bid farewell to working alone and enter a new era of "Internet of Everything, collaborative combat". In the future, your AI assistant can command AI services around the world to work for you, asking Alibaba's AI to buy tickets for you, Microsoft's AI to write your schedule, and Google's AI to make travel plans for you. This is the real era of the intelligent Internet.

3. Three Predictions for 2026

After reading these ten events