In - depth interpretation of AGI - Next 2026: 40 important judgments on differentiation, new paradigms, Agents, and the global AI competition

The just - concluded AGI - Next 2026 was extremely information - dense and visionary.

At this event initiated by the Beijing Key Laboratory of Foundation Models at Tsinghua University and Zhipu AI, in addition to Academician Zhang Bo and Academician Yang Qiang, two representatives from the academic community, Professor Tang Jie from Zhipu AI, Yang Zhilin from Dark Side of the Moon, Lin Junyang, the technical leader of Alibaba's Qwen, and Yao Shunyu, the chief AI scientist at Tencent, the core players in China's large - model field gathered together. Li Guangmi, the founder and CEO of Shixiang, also participated as the panel host.

Undoubtedly, the Chinese have become an important force in AGI. The open - source models developed by Chinese teams are well - deservedly in the global Tier 1. We believe that this position will become even more solid in 2026, and we are also looking forward to seeing more breakthrough explorations by overseas Chinese in the field of AGI in 2026.

There are already many full - text transcripts of the event. This article is a summary and extraction of the core viewpoints from all the keynote speeches and panel discussions at the event by "Overseas Unicorns". We also highly recommend that you read the full transcript to more comprehensively experience the thinking and insights of the smartest minds in the AI field.

• The differentiation of models has become an obvious trend. The reasons behind the differentiation are diverse. There are differences in the requirements of To B and To C scenarios. It is a strategic bet after careful consideration of the model competition landscape, and also a natural strategic choice of different AI labs.

• In the To B field, there will be a growing differentiation between strong and weak models. In To C scenarios, the bottleneck of model tasks is often not that the model is not large enough, but the lack of context and environment.

• Autonomous learning is a new paradigm with strong consensus. In 2026, almost everyone will invest in this direction.

• Scaling will continue. It is the result of the progress of technology, data, and taste. The exploration of cutting - edge intelligence will not stop due to potential risks.

• The model is the agent, and the agent is the product.

• The difference in computing power between Chinese and American AI is not only in absolute quantity but also in structure. The computing power in the United States is 1 - 2 orders of magnitude larger than that in China. More importantly, a significant part of it is invested in the exploration of next - generation key technologies.

……

Keyword 01

Differentiation

1. Based on observations in both China and the United States, Yao Shunyu believes that there are two aspects of differentiation in the AI field today: 1) Differentiation occurs between To C and To B; 2) There is also a growing differentiation between the two approaches of "vertical integration" and "stratification of models and applications".

2. For To C, most users do not need highly intelligent models most of the time, and they do not have a strong perception of the improvement of model intelligence. The bottleneck in To C scenarios is often not that the model is not large enough, but the lack of context and environment.

Yao Shunyu gave an example. For instance, when asking "What should I eat today?", even the most powerful model may have difficulty giving the most accurate response. The bottleneck here is that it does not know the user's current feelings, the weather environment, and other related demands.

3. The solution to this problem lies in real personalized data. For example, rather than blindly pursuing a more powerful pre - trained model, making good use of context such as WeChat chat records under the premise of compliance can bring greater value to users.

4. Under the To B logic, users (enterprises) are willing to pay a premium for "the most powerful model". Therefore, within the To B market, there will also be a differentiation, that is, the gap between strong and weak models will become more and more obvious.

For example, a powerful model like Opus 4.5 can answer 8 - 9 out of 10 tasks correctly, while a weaker model can only answer 5 - 6 correctly. Even though the latter is cheaper, enterprises still need to spend a lot of effort on monitoring without knowing "which 5 are wrong". So, they are more motivated to choose a strong model from the beginning.

5. In To C scenarios, vertical integration of the model as an all - in - one solution is feasible. The model and the product can be strongly coupled for close - loop iteration. However, in To B (productivity applications), since it involves many production links, application companies have enough opportunities to optimize around the environment and tasks. Conversely, it is difficult for application companies to train models independently, and task delivery depends on the improvement of model pre - training capabilities. Therefore, there is a stratification between model companies and applications.

6. Lin Junyang from Qwen observed that the differentiation of models is not a pre - set roadmap but more of a result of natural evolution. This natural evolution usually stems from high - frequency communication with customers. For example, Anthropic's foray into finance was an opportunity discovered through high - frequency communication with customers.

Note from Shixiang: The differences in the vertical healthcare solutions launched by ChatGPT and Claude also fully reflect the To C and To B genes of these two companies. The former provides health data interpretation for C - end users, while Claude Health takes the approach of connecting to medical systems. In the long run, OpenAI will be the next Google, and Claude will obviously become the Microsoft of the AI era.

7. "Differentiation" is also related to the timing of model competition. Zhipu's bet on coding was also based on its judgment of the model competition landscape at that time. Professor Tang Jie mentioned that after the emergence of DeepSeek, the team judged that "the battle of chatbots replacing search" was basically over. After internal discussion, the Zhipu team finally decided to bet on coding.

Keyword 02

New Paradigm

Autonomous Learning

8. First, scaling will continue. However, in terms of scaling investment, Professor Tang Jie believes that two different directions need to be distinguished.

• Scaling the known path: By continuously increasing data and computing power, continuously exploring the upper limit of capabilities. However, in essence, it is also a "lazy" approach.

• Scaling the unknown path: That is, finding new paradigms that have not been clearly defined. Allowing the AI system to define the reward function, interaction methods, and even training tasks for scaling.

Note from Shixiang: Currently, the AI community does not have a unified concept definition for the new paradigm. Autonomous learning, active learning, continual learning, and self - learning essentially express the same expectation, that is, to improve the model's autonomous learning ability so that it can continuously enhance intelligence without human intervention.

9. Yang Zhilin summarized the Scaling Law as a perspective of converting energy into intelligence. Its core lies in efficiently approaching the upper limit of intelligence. The model embodies values and taste. Scaling is the progress of technology, data, and aesthetics. The exploration of cutting - edge intelligence will not stop due to potential risks.

10. The goal of autonomous learning is to enable the model to have self - reflection and self - learning abilities. Through continuous self - assessment and self - criticism, the model can gradually distinguish which behaviors are effective and which paths still have room for improvement.

11. Yao Shunyu believes that the emergence of the new paradigm is not a "sudden change point" in the future but a "gradual change" process that is currently underway. In fact, he had already seen some signs in 2025.

For example, Cursor's Auto - complete model learns from the latest user data every few hours. ChatGPT uses user data to fit the chatting style, which is also a form of self - learning. Claude Code even wrote 95% of its own project code. From a certain perspective, AI has already shown signs of helping itself improve.

12. The biggest bottleneck of the new paradigm is actually imagination. More specifically, if it is announced in 2027 that the new paradigm has been realized, what tasks can we use to prove that this paradigm has been achieved? Should it be a profitable trading system? Or should it solve unsolved scientific problems? That is to say, when we think about the new paradigm at present, we should first be able to imagine what it looks like.

13. Lin Junyang believes that from a more practical perspective, the potential of RL has not been fully explored, and there is still much potential to be tapped. There are two dimensions for the next - generation paradigm: first, autonomous learning; second, AI having stronger initiative. Currently, humans help AI start, but in the future, the model may no longer need human prompts, and the environment itself can prompt it.

14. Active learning will bring serious security challenges. The risk lies not in "saying what should not be said" but in "doing what should not be done". Active learning is definitely an important paradigm, but it must be guided in the right direction.

15. Autonomous learning can be reflected in personalization, but it will be difficult to measure whether it has "improved". In a recommendation system, click - through rate can be used as an indicator. However, when AI covers all aspects of life, the evaluation indicators become extremely vague.

16. The issue of continual learning involves a time concept, that is, the model is in a continuous learning process. However, for long - term tasks with multiple agents in series, once the ability of an agent is not 100%, its ability often declines exponentially as the task progresses. In the human learning mechanism, sleep is used to "clean up noise". Perhaps AI also needs to explore similar noise - cleaning and new computing models.

17. Professor Tang Jie proposed the concept of "Intelligence Efficiency". That is, future paradigms should not only focus on scaling but also on "how much intelligence increment can be obtained with how much resource input". This is the key to solving the cost bottleneck. The significance of the new paradigm lies in how to achieve the same or even more intelligence improvement with fewer paradigms.

18. The development path of large models has always drawn on the cognitive learning process of the human brain, gradually moving into tasks such as knowledge compression, reasoning, mathematics, coding, and other abstract deductions. In terms of 1) multimodality, 2) memory and continual learning, and 3) reflection and self - awareness, humans significantly outperform current models, and these points may be new breakthrough directions.

19. In 2020, Zhipu drew a structural diagram of an AI system referring to human cognition, which has three modules: System 1, System 2, and self - learning. The introduction of self - learning is mainly based on the following reasons, corresponding to 3 types of scaling:

Native Multimodality

20. The native multimodal model is similar to human "sensory integration". Human sensory integration can collect visual information on one hand and also collect auditory and tactile information at the same time. The brain then integrates these sensory information to perceive an object. Currently, the sensory integration ability of models is not sufficient.

21. Multimodal sensory integration is one of Zhipu's key directions this year. Only when the model has this ability can AI perform long - chain and long - term tasks in a real - world work environment, such as continuous collaboration on devices like mobile phones and computers.

22. Multimodality is also something that Qwen will continue to work on. Lin Junyang believes that if one wants to create a truly intelligent product, it should naturally be multimodal. However, there is also a debate on whether multimodality can drive intelligence.

23. From the perspective of the model providing more productivity and better assisting humans, developing multimodal capabilities such as vision and speech is a natural choice.

24. Video is a more general form of expression. Images can be regarded as single - frame videos. Understanding long videos is a very interesting task.

Keyword 03

Agent

25. Coding is the inevitable path to becoming an agent. In Zhipu's practice, it was found that although GLM - 4.5 had high scores in benchmarks, it could not write a "Plants vs. Zombies" game. Only after introducing RLVR and a large amount of real - world programming environment training did GLM - 4.7 solve this problem.

26. The model is the product. For an agent to perform complex tasks, the requirements for the model are quite high. The model is the agent itself, and the agent is the product itself. If they are integrated, then developing a basic model today is actually developing a product.

27. The differentiation between To B and To C models is also reflected in agents:

• The indicators of To C products are sometimes not related to model intelligence, and may even be the opposite.

• To B agents do not even need to make many innovations. As long as the model's intelligence improves, the ability to solve real - world tasks increases, and more value is created.

28. The use of agents in productivity scenarios is just beginning. In addition to model improvement, the environment and deployment are equally important and are the keys to the agent creating value. Even if the model does not improve any further, simply deploying the existing models in various companies can bring a 10 - fold or even 100 - fold increase in revenue. However, currently, the impact of AI on GDP is still far less than 1%.

29. Education is very important. The gap between people is widening. It is not that AI will replace human jobs, but that people who can use these tools will replace those who cannot.

30. Lin Junyang from Qwen believes that future agents will become "managed". Users will no longer need to interact frequently as they do now. Instead, they can set a general goal, and the agent will run independently in the background for a long time until the task is completed.

31. For an agent to achieve this, it also depends on the self - evolution and active learning mentioned earlier, because the requirements for the model in this regard are quite high. Under this logic, we can say that "the model is the agent, and the agent is the product".

32. In the process of developing a general agent, long - tail tasks are more worthy of attention. Users can feel the value and charm of AI because a long - tail task has been solved. The so - called AGI today essentially aims to solve long - tail problems.

33. Whether to develop a general agent is a matter of personal opinion. If one is confident in being a "wrapper expert", they can do it. However, if the wrapper does not have more information than the model company, then the general agent presents an opportunity for the "model - as - product" concept. For model companies, many engineering problems can perhaps be solved by simply "burning some computing resources".

34. Professor Yang Qiang divided the development of agents into four quadrants from two dimensions:

• Goal definition: Is it defined by humans or automatically defined?

• Task planning: That is, the intermediate actions. Are they defined by humans or automatically defined by AI?

Currently, we are still at a very early stage: the goal is defined by humans, and the planning is also done by humans. However, in the future, a large - scale model will observe human work, especially by making use of human process data. Eventually, both the goal and the planning can be defined by the large - scale model. Therefore, the agent should be a native system inherent in the large - scale model.

35. Several important issues that will determine the future trend of agents:

• Can agents truly solve human tasks? Can this create value? How much value can it create?

• How high is the cost of agents? On the one