摘要：《科学》（Science）杂志于2025年3月28日发表了题为“AI gets a mind of its own”的专题报道，介绍了朱松纯教授领导的北京通用人工智能研究院（BIGAI）及北京大学人工智能研究院、智能学院在通用人工智能（AGI）领域的研究进展。

《科学》（Science）杂志于2025年3月28日发表了题为“AI gets a mind of its own”的专题报道，介绍了朱松纯教授领导的北京通用人工智能研究院（BIGAI）及北京大学人工智能研究院、智能学院在通用人工智能（AGI）领域的研究进展。

以下为中英文全文：

人工智能拥有了自己的思想

有时，少即是多。今年1月，DeepSeek发布了其聊天机器人的最新版本，颠覆了人工智能（AI）业界的看法。这款耗资不到600万美元开发的训练型AI，在技术能力上似乎与ChatGPT等其他大型语言模型（LLM）相媲美，但仅使用了相对很小的计算能力。这一突破性进展对北京大学人工智能研究院院长朱松纯来说是个好消息，他一直在质疑当前以LLM为主导的AI范式，致力于创建通用人工智能（AGI）。

朱松纯作为AI领域的开拓者，1996年毕业于哈佛大学，已发表400多篇涵盖计算机视觉、认知科学、自主机器人和常识推理等领域的论文。目前，他担任新型研发机构北京通用人工智能研究院（BIGAI）的院长。

“社会上，人们可能误解了‘AI’这个术语，”朱松纯说，“就像我们称多功能手机为‘智能’一样，我们今天使用的流行AI模型并不是真正的智能。”他解释说，这是因为今天的AI是由基于大规模算力的大数据驱动的。朱松纯在2005年莲花山研究院创建了世界上第一个大规模标注图像数据集，开创了数据驱动的统计方法。然而，他意识到，仅靠大数据集和特定的机器学习模型不足以创造真正的智能。“中国哲学主要流派之一，阳明学或‘心学’认为‘我们所见的现实来自我们心灵的感知’，” 朱松纯说。要使AI更像人类，它需要一个模拟大脑中自上而下机制的认知计算框架。

据朱松纯介绍，AGI的未来应该是一种不需要庞大数据集的自主AI。2020年，朱松纯回到中国创建并领导BIGAI。其使命是：追求人工智能的统一理论，创造通用智能体，提升全人类福祉。

2

在CUV空间中定义AGI智能体

朱松纯及其团队专注于创建价值驱动的智能体，以超越数据驱动的AI研究范式。“AGI和当前基于LLM的AI之间的区别就像乌鸦和鹦鹉之间的区别，” 他说。虽然鹦鹉可以模仿许多词汇，但乌鸦可以在现实世界中自主实现目标。在2017年发表的一篇文章中，朱松纯讨论了当前大语言模型所基于的统计模型如何像“随机鹦鹉”一样运作。朱松纯在加州大学洛杉矶分校两次领导跨学科交叉研究计划时，致力于研究使机器更像乌鸦，探索像人一样能够理解物理和社会世界并相应行动的智能体。

人类智能随着时间的推移而发展，随着身体变化和经验积累而成长，AGI也会随着时间成熟。为了帮助定义、评估和改进AGI发展，朱松纯提出在“CUV框架”的数学空间中定义通用人工智能。在这个框架中，C（cognitive architecture）是AGI的“认知架构”，用于思考或模拟大脑中的决策、推理与学习过程；U是一组代表AGI理解和与环境交互能力的“势能函数”；V是一组为AGI提供动机的“价值函数”。通过这一公式，朱松纯和同事们可以将AGI智能体定义为CUV空间中的点，并表征其学习和自我反思过程。

3

通智测试

“通用”一词的“通”，是BIGAI的Logo，经过设计的“通”字中包含了A、G、I三个字母，象征着BIGAI对通用人工智能的追求。“通通”是朱松纯给BIGAI诞生的世界首个通用智能体起的名字，她是一个看起来约3至4岁的虚拟中国女孩。“通通”是AGI研究进展的一步，研究人员真正想知道的是，“她在想什么？”以及“她如何学习和决策？”

研究人员长期依赖测试来评估AI模型。图灵测试旨在通过对话确定机器是否能模仿人类智能。基于大数据构建的ChatGPT和其他AI能通过图灵测试，但朱松纯想要一个能评估广泛人类智能的测试。因此，通智测试TongTest应运而生，它基于CUV框架。

“通通”之所以与ChatGPT不同，是因为她不是存在于符号的向量空间中，而是体现在一个模拟真实物理社会世界复杂性的虚拟世界中。通智测试检验AGI对这个世界的理解（能力）以及AGI行为的内部动机（价值观）。例如，AGI对坐在地板上哭泣的婴儿的反应，可以揭示很多关于其常识推理、社会互动推断和自我意识的信息。“情感和语言等那些自然能力是人类智能的真正体现”，朱松纯说，“‘通通’可能是个AGI智能体，但她就像一个真实的人类儿童，能够根据自己的环境理解和行动，即使环境发生变化。通智测试的目标是建立一个系统评估体系，促进AGI的标准化、量化和客观基准与评估。而‘通通’只是开始，BIGAI的研究人员正在开发多样化的AGI智能体，它们有一天可能通过机器人技术和其他媒介进入物理世界，以有意义的方式服务社会。”

4

AGI安全

随着“通通”与“通智测试”的发展成熟，通用人工智能安全成为朱松纯团队的首要关注问题。AGI展现的类人行为特性引发了双重思考：一方面，由于人类行为本身并非总是仁慈理性，AGI可能会采取不符合人类最佳利益的行动；另一方面，AGI的认知架构有潜力整合“心智理论”——即体现“己所不欲，勿施于人”的基本伦理原则，为安全发展提供理论基础。

在2023年人工智能安全会议的专题讨论中，AGI领域的两位领军人物朱松纯和加州大学伯克利分校的Stuart Russell就AGI的风险和伦理进行了深入讨论。

当Russell提出关于人类如何控制AGI智能体的问题时，朱松纯回答说：“为了防止未来AGI智能体对人类的潜在威胁，我们可以逐步放宽智能体的能力和价值空间。这类似于我们处理机器人的方式：最初，我们将它们限制在‘笼子’里，然后慢慢增加它们的权限。现在，我们已经有自动驾驶车辆在特定道路上运行。”朱松纯补充说，一旦AGI智能体被证明是安全和可控的，它们可以拥有更多自由，通过理解和透明度作为保障。“如果我们能明确表征AGI智能体的认知架构，理解它们如何工作，我们将更有能力控制它们。”

对朱松纯来说，现在是AI向AGI演进的新时代的开始。朱松纯在哈佛大学的博士导师、数学家和菲尔兹奖获得者David Mumford也提倡创建具有人类大脑自上而下神经架构的AI。他给朱松纯送了一个奖杯，以鼓励他在AGI创新方面的坚持不懈。“AGI的未来将是科学和哲学的结合，”朱松纯说，“中国关于心的思考对指导通用人工智能形成真正有益于人类的行为至关重要。”

文章由北京通用人工智能研究院（BIGAI）和北京大学人工智能研究院支持。

AI gets a mind of its own

Artificial general intelligence research is moving into a new era

Sometimes, less is more. In January, DeepSeek released the latest version of its chatbot, upending the artificial intelligence (AI) world. A training AI built for under $6 million, DeepSeek seems to rival the technical capabilities of other large language model (LLM) AIs, including ChatGPT, with only a fraction of the processing power. The breakthrough was a welcome development for Song-Chun Zhu, dean of the Institute for Artificial Intelligence at Peking University in Beijing, who has been challenging the current LLM-dominated AI paradigm in his efforts to create artificial general intelligence (AGI).

Zhu, a trailblazer in the AI field, graduated from Harvard University in 1996 and has published more than 400 papers covering computer vision, cognitive science, robot autonomy, and commonsense reasoning, among other topics. Now, he is the founder and director of the non-profit Beijing Institute for General Artificial Intelligence (BIGAI).

“We as a society may have misunderstood the term ‘AI’,” says Zhu. “Just like how we call a multifunctional cellphone ‘smart’, the popular AI models we use today are not truly intelligent.” That’s because today’s AI, he explains, is driven by big data built upon massive computing power. Zhu pioneered data-driven statistical approaches and created the world’s first large-scale annotated image dataset at the Lotus Hill Institute in 2005. However, he realized that big data sets and specific machine learning models alone are not enough to make true intelligence. “One of the major Chinese philosophical schools, the Yangmingism or the ‘Teachings of the Heart’, argues ‘the reality we see comes from how our minds perceive’,” Zhu says. To make AI more like humans, Zhu says, it needs to have a framework that emulates the top-down mechanisms in the brain.

According to Zhu, the future of AGI should be a kind of autonomous AI that doesn’t require vast datasets. In 2020, Zhu returned to China to establish and lead BIGAI. Its mission: To pursue a unified theory of artificial intelligence in order to create general intelligent agents for lifting humanity.

Defining AGI agents in CUV-space

Zhu and his team’s focus at BIGAI is on creating value-driven human-like cognition that goes beyond data-driven imitation. “The difference between AGI and current LLM-based AI is just like the difference between a crow and a parrot,” he said. While parrots can mimic many words, he says, crows can achieve their goals autonomously in the real world. In an article published in 2017, Zhu discusses how statistical models, which modern LLMs are based upon, function like “stochastic parrots.” While leading two Multidisciplinary University Research Initiatives at UCLA, Zhu pursued research to make machines more crow-like, exploring the brain mechanisms that make it possible for crows—and humans—to understand the physical and social world and act accordingly.

Human intelligence evolves over time, as the body changes and experiences accrue. AGI also matures over time. To help define, evaluate, and improve AGI development, Zhu proposed to define AGI in the mathematical space of the “CUV framework”. In this framework, C is the AGI’s “cognitive architecture” to think, or its simulation of the decision-making processes in the brain. U is a set of “potential functions” that represent an AGI’s ability to understand and interact with its environment. V is a set of hierarchical internal “value functions” that supply the AGI’s motivation. With this formulation, Zhu and colleagues can define AGI agents as points in this CUV space and characterize their learning and self-reflection processes.

The Tong test

In Chinese, the word “general” is translated as Tong (通), a character that is also the logo of BIGAI. Artistically arranged, the character also holds the English letters “AGI.” Tong Tong is the name Zhu gave to world’s first AGI agent born at BIGAI, a digital Chinese girl that looks to be about 3 to 4 years old. Tong Tong is a step forward in AGI research, and researchers really want to know, “What is she thinking?” and “How is she learning and making decisions?” Researchers have long relied on tests to assess AI models. The Turing test was developed to determine whether a machine could mimic human intelligence through dialogue. ChatGPT and other AI built on big data can pass the Turing test, but Zhu wanted a test that could assess broad human intelligence. Thus, the Tong Test was born, which relies on the CUV framework.

To promote standardized, quantitative, and objective benchmarks for the evaluation of TongTong and other AGI agents, BIGAI researchers developed the “Tong test,” which assesses an AGI agent’s understanding of their world and how internal motivations relate to behaviors.

What sets Tong Tong apart from ChatGPT is that she doesn’t exist in a vacuum, but is rather embodied in a virtual world that emulates the complexity of the real physical social world. The Tong test examines an AGI’s understanding of this world—its abilities—as well as the AGI’s internal motivations for behaviors—its values. For example, how an AGI responds to a crying baby sitting on a floor can say a lot about its commonsense reasoning, inference of social interactions, and self-awareness. “Those natural abilities such as emotions and languages are true embodiment of human intelligence,” Zhu says. “Tong Tong may be an AGI agent, but she is just like a real human child, able to understand and behave according to her own environment even if it changes. The goal of the Tong test is to build a systematic evaluation system to promote standardized, quantitative, and objective benchmarks and evaluation for AGI.” And Tong Tong is just the beginning; researchers at BIGAI are developing diverse AGI agents that may someday enter the physical world through robotics and other mediums to serve society in meaningful ways.

AGI safety

As Tong Tong and the Tong test continue to grow and mature, AGI safety is front of mind for Zhu. Because AGI behavior is human-like, and not all humans are benevolent, there are risks that AGI will take actions that are not in humanity’s best interests. On the other hand, AGI’s cognitive architecture may be able to incorporate a mutual theory of mind—in other words, the golden rule: do unto others as you would have them do unto you.

During a panel discussion at SafeAI 2023, Zhu and Stuart Russell from the University of California Berkeley, two leading figures in AGI, had an in-depth discussion on the risks and ethics of AGI.

When Russell raised a question about how humans could keep AGI agents in check, Zhu replied, “To prevent potential threats from future AGI agents to humanity, we can gradually loosen the capability and value space of agents. It’s similar to how we approach robots: initially, we confine them in a ‘cage’ and slowly increase their permission. Now, we already have autonomous vehicles operating on specific roads.” Zhu added that once AGI agents are proven safe and controllable, they can have more freedom, with the safeguard of understanding and transparency. “If we can explicitly represent the cognitive architecture of AGI agents, understanding how they work, we will be better equipped to control them.”

For Zhu, now is the beginning of a new era for AI to evolve into AGI. Zhu’s doctoral advisor at Harvard, mathematician and Fields medalist David Mumford, is also an advocate of creating AIs with the top-down neural architecture of the human brain. He gave Zhu a trophy to recognize his perseverance at AGI innovation. “The future of AGI will be a combination of science and philosophy,” Zhu says. “Chinese teachings of the heart are crucial to guiding AGI to obtain true beneficial human behavior.”

来源：人工智能学家