At the 2026 Chongli Forum, 360 Group founder Zhou Hongyi shared his thoughts on the future of artificial intelligence. He began by saying, "I came to Taiwu Town with the same goal as everyone else—primarily to ski. Additionally, I hope to fulfill a dream here. I've been skiing for many years, but my skill level has remained quite average; I really want to master the technique of separating my upper and lower body movements."

Zhou Hongyi also remarked, "I'm happy to share some perspectives on AI with you all. Honestly, talking about trends is particularly easy because you can say almost anything, even with your eyes wide open. If it comes true a year later, you can claim your prediction was accurate; if it doesn't, everyone will have forgotten anyway. Today, Old Yu (Yu Minhong) even set some boundaries for me—don't let people call me an 'old fogey' or come across as too paternalistic. So, the views I share today aren't necessarily all correct, nor are they entirely my original ideas. I'm just throwing out some thoughts to stimulate discussion."

While everyone was chasing large language models in 2024, and 2025 sees these models transforming into AI agents, what will 2026 look like? Zhou Hongyi predicted, "I believe that by 2026, not just in China but worldwide, there will be at least 100 billion AI agents. In the AI sector, even companies valued at hundreds of billions will be considered small. Why do I say that? Because large models themselves are continuously evolving. Previously, they were basically just chatbots, incapable of doing real work. Only by upgrading them into agents can they be practically deployed across various industries."

He further elaborated, "I think a mistake many made in 2024 and 2025 was conflating training compute with inference compute. In reality, the demand for training compute is enormous, but that's the exclusive domain of a few large companies. For the vast majority of ordinary people and businesses, what's needed is inference compute. If you chat with a large model, even if you talk extensively—say 20,000 words a day—you might only consume tens of thousands of tokens per month. But if you truly want an AI agent to help you produce a short film or deliver a training session within your business operations, the computational power required could easily reach millions or even tens of millions of tokens."

Zhou Hongyi also noted that China's open-source ecosystem has promoted the flourishing development of AI agents not only domestically but globally. For instance, ASEAN countries are now using Chinese open-source models to build their national systems. He pointed out that many nations may worry about a future where AI competition is dominated only by China and the US, leaving other countries that cannot participate at risk of becoming "food on the table."

He shared insights from his recent trip to Silicon Valley, where he met with numerous AI researchers and product managers. He discovered a stark contrast in their views: research scientists told him that AI agents are trivial and unnecessary, believing that further model training and advancements will render agents obsolete. Conversely, everyone involved in engineering, application development, and product creation emphasized the critical importance of agents. Zhou commented, "I believe large models and AI agents need to evolve in tandem. It's unrealistic to hope solely for models to add more parameters or achieve algorithmic breakthroughs and suddenly transform into AGI (Artificial General Intelligence) through model evolution alone."

He addressed a common misconception, stating, "Many people mistakenly think that only those training large models or developing chips are the main players in AI. While that is undoubtedly important, the key to integrating AI into enterprises and industries lies in creating AI agents. I believe the real protagonists are the business and industry experts in vertical fields. Every organization possesses a wealth of tacit knowledge and unwritten rules—such as specific leadership expectations for a role or unique company procedures. This knowledge often resides in the minds of senior employees or executives, and many specialized technical skills are closely guarded corporate secrets. Using this knowledge for general training could lead to the loss of proprietary advantages. The best application is to train AI agents with it. Whoever can convert this tacit knowledge into model learning content will capture the core benefits of AI development."

He concluded with a personal example: "Recently, I spent a month working intensively on writing an AI agent myself, aiming to create one that could read online novels and directly output web short dramas. In the process, I realized that even the most skilled engineers cannot script the camera movements, emotional expressions, or character actions required by the agent. These elements need the expertise of professional cinematographers and directors. So, I want to correct a viewpoint: besides those involved in core technology, practitioners from various business fields are actually the main drivers of AI implementation."