Introduction

The 20th Central Committee of the Communist Party of China has proposed the comprehensive implementation of the “AI+” initiative to lead research paradigm shifts, integrate AI with industrial development, cultural construction, and social governance, and seize the high ground in AI applications. This strategic deployment aims to deeply integrate AI with economic and social development, reshape production and lifestyle paradigms, and promote revolutionary changes in productivity and production relations. What exactly is artificial intelligence, how does it impact our lives, and what does the future hold?

Interview Guests

• Zhang Haining, Dean of the Software College, Nankai University
• Li Yan, Director of the Institute of Technology and Cybersecurity, China Institute of Modern International Relations
• Nie Hua, Chairman and Senior Engineer, China Controlled Information Industry Co., Ltd.
• Chen Keliang, Associate Dean of the School of Artificial Intelligence, Beijing University of Posts and Telecommunications

Understanding Artificial Intelligence

Reporter: In recent years, AI has quietly permeated every corner of social production and life, profoundly changing human ways of living and working. So, what is AI? How do we understand concepts like narrow AI and general AI? What do breakthroughs in large AI models, represented by ChatGPT and DeepSeek, mean?

Zhang Haining: Artificial Intelligence (AI) simulates, extends, and enhances human intelligence. It aims to enable machines to think and solve problems like humans by learning, analyzing, and training on large datasets through algorithms and models. The original intention of AI was to accurately simulate every aspect of human learning and intelligence. For a long time, our achievements have mainly been in narrow AI (ANI), which excels in specific tasks (like Go or facial recognition) but lacks capability outside those domains. General AI (AGI), on the other hand, requires machines to possess human-like cognitive abilities, enabling cross-domain learning, reasoning, and decision-making while effectively collaborating with humans or other intelligent agents. Although large models like ChatGPT and DeepSeek show promise for AGI, challenges such as adaptability, continuous learning, and interaction with the external world remain. Achieving general intelligence is a gradual process, not a sudden event triggered by a single technological invention. Currently, large models are primarily probabilistic prediction models that, while knowledgeable, lack perception of the physical world and specific problem-solving skills.

Li Yan: From a technical perspective, the core of large language models is “learning language patterns” rather than “understanding world patterns.” They master the logic of language expression through vast amounts of text data, enabling them to generate fluent content but lacking real-world entity perception and causal reasoning capabilities. For example, ChatGPT is mainly based on pre-training, using large datasets for model tuning, akin to a “drill practice” approach, and does not possess autonomous intelligence. Current top models exhibit significant shortcomings in areas like long-term memory storage and multi-modal integration. Most conversational AI systems lack long-term memory, starting anew with each interaction, making it impossible to remember user preferences and build a rapport. Language can carry human knowledge and logic, but general AI also requires integration of visual, auditory, and tactile modalities, along with autonomous decision-making and emotional understanding. Yann LeCun’s concept of Advanced Machine Intelligence (AMI) emphasizes that machines should understand the physical world, possess persistent memory, and have reasoning and planning abilities, rather than merely extending language processing capabilities. Thus, large language models are merely a milestone in the journey towards general AI, not the endpoint.

Chen Keliang: In 1950, Alan Turing, the father of computer science, began humanity’s exploration of intelligent machines with the question, “Can machines think?” The term “artificial intelligence” was officially named at the Dartmouth Conference in 1956. The first chatbot, Eliza, was born at MIT in 1966, achieving simple human-computer dialogue. However, due to weak data, computing power, and algorithmic theories, AI development stagnated for a time. After decades of development, Geoffrey Hinton and his team at the University of Toronto proposed the concept of “deep learning” in 2006, ushering in a new era for AI. In 2012, the emergence of large annotated datasets and powerful parallel computing capabilities from GPUs led to the rise of deep learning technologies like AlexNet, allowing AI to “learn to learn” from vast amounts of data. In 2016, Google’s DeepMind team developed AlphaGo, which defeated world champion Lee Sedol in Go, showcasing a move beyond human strategies. Since 2022, pre-trained large model technology has propelled AI to unprecedented heights, with models like ChatGPT and DeepSeek capable of understanding language, generating content, and performing complex logical reasoning. However, all AI we see today is still narrow AI, designed for specific tasks, akin to specialists in various fields, while general AI remains a theoretical exploration. Language is a crucial carrier for building general AI, but not the only element. Single-modal models are limited to one information dimension, and even multi-modal models that integrate text and images have yet to break through the boundaries of narrow AI, unable to achieve general reasoning across scenarios like humans.

Key Elements of AI Development

Reporter: What are the core elements driving AI development? What are the future directions and paths?

Zhang Haining: If we compare training an excellent large model to cooking a grand feast, data, computing power, and algorithms are indispensable core elements. Data is the “top-quality ingredient.” In the era of large models, the quality of data is far more important than quantity; we need high-quality data that is rigorously cleaned and rich in human logic and industry wisdom, rather than chaotic raw data from the internet. Computing power is the “intense heat.” Both model training and inference require thousands of GPU chips forming a massive computing cluster to provide astonishing computational power, determining whether we can distill vast knowledge in a limited time. Algorithms are the “key recipes.” They dictate how we construct neural network architectures (like the Transformer architecture, which is based on self-attention mechanisms) and how to efficiently utilize computing power and data. Excellent algorithms can significantly enhance learning efficiency, training smarter models with fewer resources.

Regarding future directions and paths, I believe that deep reasoning models represented by OpenAI’s GPT-5.2 and Google’s Gemini 3.0 have made qualitative leaps in handling complex tasks, driving AI development. In the future, we will continue to expand this deep reasoning capability, transitioning from simple dialogue to action, building intelligent agents capable of long-term complex work with autonomous planning and execution abilities, making AI a super assistant for humanity. Secondly, we will develop embodied intelligence, integrating the brains of large models into robots to perceive the environment, make decisions, and solve real-world problems. Finally, a paradigm revolution in AI algorithms and underlying architectures is needed. Current large model architectures are gradually hitting bottlenecks in data and computing efficiency, with internet data nearly exhausted and training costs remaining high. Fundamental research needs to explore new algorithms and model architectures, such as developing more efficient learning algorithms to reduce data requirements and training costs, and drawing inspiration from the workings of the human brain to achieve brain-like intelligence at the foundational level.

Li Yan: I believe the future of AI will evolve along the path of “technical deepening + ecological collaboration”: first, multi-modal integration will continue to deepen, breaking down barriers between text, images, audio, etc., to achieve more natural human-machine interaction; second, model efficiency and lightweight solutions will run parallel, reducing training and inference costs while maintaining performance, effectively overcoming energy challenges and promoting AI proliferation; third, enhancing explainability and controllability will address the “algorithmic black box” problem, increasing transparency and trustworthiness, and solving issues like “hallucinations” and controlling “emergent” phenomena; fourth, achieving iteration through applications, feeding back technical iterations through practical applications, accumulating data from real scenarios to upgrade models from “usable” to “well-used.”

Key Factors for AI Breakthroughs

Reporter: What are the key factors driving a new round of AI breakthroughs? What advantages and shortcomings does our country have in these areas? How can we respond to opportunities and challenges?

Zhang Haining: As the scale law of foundational models faces diminishing marginal returns, the key to driving a new round of AI breakthroughs lies in application implementation and architectural innovation. On one hand, we need to deeply integrate general foundational large models with real industry scenarios, utilizing private data from sectors like industry, healthcare, and finance to address specific, high-value industry pain points that general large models cannot solve, transforming AI from a technical toy into a new productive force. On the other hand, we need to explore the architecture of large models, breaking through the bottlenecks of existing Transformer architectures in long text processing and reasoning efficiency, seeking new architectures (like sparse attention mechanisms or state space models) that utilize data more effectively and reduce inference costs, incorporating genuine mathematical reasoning methods into large models to generate stronger intelligence and lower computing costs, achieving truly inclusive AI.

Nie Hua: It is well-known that globally, AI development is entering a critical stage of transitioning from speed expansion to systematic competition, where the spillover effects of technological innovation and the shaping of institutional rules are becoming increasingly prominent. Developed countries possess varying degrees of first-mover advantages in large model research and development, AI chip design, high-performance computing platforms, and industrial applications. Coupled with export controls on key devices and software, as well as competition for ecological norms and standards, AI has evolved from a purely technical issue into a systemic strategic game. In this international context, our country faces a complex situation of both opportunities and challenges. On one hand, we have abundant data resources, a complete industrial system, vast application scenarios, and a huge market space, showcasing unique advantages in the AI wave. For instance, the ultra-large application scenarios and massive data foundation provide natural experimental grounds for model training and algorithm validation, playing a decisive role in transforming technology from laboratories to industries. In recent years, our country has made significant progress in algorithms, systems engineering, cloud infrastructure, and terminal adaptation, gradually forming a complete innovation chain from fundamental research to engineering realization, providing solid conditions for building an autonomous and competitive AI ecosystem. On the other hand, our foundational AI capabilities still face “bottleneck” risks, particularly in high-end AI chips constrained by advanced processes, high-efficiency GPUs facing industrial chain limitations, and foundational capabilities like electronic design automation (EDA) tools, intellectual property (IP) licensing, and advanced manufacturing equipment not being fully autonomous and controllable. Some critical foundational software, such as firmware, operating systems, databases, and middleware, still heavily rely on overseas systems. In the open-source ecosystem, while our contributions are increasing, the foundational frameworks and major standards are still dominated by Western countries, necessitating strengthening our voice in ecological rules and technological routes. Internationally, negative factors such as export controls on AI, data barriers, and supply chain restrictions are increasing, making the external environment for our AI development more severe.

Chen Keliang: In the face of the rapid evolution of new-generation AI technologies, we must seize historic opportunities, adhere to self-reliance, emphasize application orientation, and continuously promote the healthy and orderly development of AI towards beneficial, safe, and equitable directions. This will ensure that AI, as a strategic technology leading the future, injects a continuous and powerful impetus for technological innovation to drive high-quality development and advance Chinese-style modernization. To address opportunities and challenges, we should adopt a strategy of “bridging gaps and strengthening advantages”: optimizing computing efficiency through technological innovations like “in-storage computing” to break through the bottlenecks of storage-computing separation architecture; enhancing foundational algorithm research to build collaborative innovation platforms among government, industry, academia, and research; leveraging data advantages to establish a compliant data element market, while leading industrial ecological construction with standards, for example, by seizing international discourse power through standards for intelligent agent internet.

AI Application Scenarios

Reporter: How do we understand AI application scenarios? How can we prevent a disconnect between technology and actual needs and better promote the implementation of AI applications?

Li Yan: The biggest characteristic of our country’s AI industry development is “application-oriented,” which helps form an enhanced loop of “scenario application - data accumulation - technology iteration”. By focusing on solving practical problems, we accumulate real data through applications, allowing AI to become smarter and solve more complex issues. For example, the DeepSeek-R1 inference model has been efficiently trained at low cost and has been implemented in multiple industry scenarios, showcasing the practical advantages of AI applications in our country. To avoid a disconnect between technology and needs, we must establish a closed-loop mechanism of “demand-driven + iterative optimization”: first, clarify the direction of technological breakthroughs based on industry pain points, avoiding blindly pursuing technological leadership or conceptual innovations, ensuring product design aligns with actual scenarios. For instance, in AI applications in manufacturing, we should focus on core needs like cost reduction and efficiency improvement rather than pursuing flashy technological effects. Second, we should adopt an agile iteration model, quickly pilot the minimum viable product (MVP), and continuously optimize functions based on user feedback to achieve dynamic matching between “technology supply” and “demand satisfaction.” Third, we should establish scenario-based evaluation standards, moving away from single technical indicators to include practicality, cost-effectiveness, and usability in evaluation dimensions, guiding technological development to focus on application implementation.

Nie Hua: If core technology is the “root” of the AI industry, then a thriving ecosystem is its branches and leaves. Only with deep roots can leaves flourish, and a solid foundation leads to abundant branches. Breakthroughs in core technologies ultimately rely on a prosperous ecosystem for support, and competition in the AI industry is essentially a competition of ecosystems. Only by adhering to an application-oriented approach and maintaining an open-source collaborative philosophy can we build a high-quality, sustainable, and self-reliant AI ecosystem. This involves promoting national-level open-source large model communities, industry-level intelligent platforms, and full-chain collaboration among AI chips, frameworks, machines, models, and applications, as well as building localized application systems in key industries. For example, in the industrial sector, we should promote the application of industrial robots and humanoid robots in factories, prioritizing their implementation in welding, assembly, and logistics scenarios; in critical sectors like finance, we should encourage the use of domestically produced, controllable computing devices and software systems for large-scale verification in real high-concurrency and high-load scenarios. This kind of “battle-tested” practical examination helps form a virtuous cycle of “application driving technological innovation, and technology empowering industrial upgrading.” We should actively participate in the formulation of AI standards, leading or participating in the establishment of industry standards and testing specifications, transforming our technological accumulation in high reliability and high-performance computing into industry standards, enhancing the market recognition and international discourse power of domestic products, and providing institutional guarantees for the healthy development of the industry.

Ethical Challenges of AI

Reporter: With the deep application of AI, a series of ethical dilemmas involving safety, privacy, fairness, and responsibility have emerged. How do we understand and respond to the safety risks brought by AI development, prevent technological misuse, and ensure that intelligence serves humanity?

Zhang Haining: Technology itself is neutral, but its application can pose ethical challenges. The main risks include algorithmic bias (social biases in training data), intellectual property disputes (copyright ownership of generated content), and trust crises brought by deep fakes. To prevent technological misuse and ensure that intelligence serves humanity, we need to build a governance system that balances technology and institutions. On the technical side, we should strengthen research on value alignment, equipping AI with safeguards through reinforcement learning to ensure its outputs align with human moral standards, and develop technologies capable of identifying AI-generated content (like digital watermarks). On the institutional side, we should establish a tiered and categorized regulatory mechanism, implementing strict access controls for high-risk areas like autonomous driving, medical diagnostics, and content generation. More importantly, every practitioner should maintain a sense of responsibility, integrating social responsibility into the entire lifecycle of algorithm design, data processing, and training inference, ensuring that AI remains a tool serving human welfare rather than an uncontrollable black box.

Nie Hua: From deep fakes to algorithmic discrimination, from data privacy breaches to potential loss of control risks, if these issues are not handled properly, they could not only hinder the healthy development of technology but also impact national security and social stability. Therefore, we must adhere to the value concept of “people-centered, intelligence for good” and strengthen ethical and moral defenses. We should also achieve “internal safety” through “self-controllability,” mastering the initiative in governance. Ultimately, the strength of governance capabilities depends on the solidity of the technological foundation. If our AI industry is built on foreign foundational hardware and software, then so-called safety governance is mere talk. Therefore, we must transform the results of “strengthening chains and bridging gaps” into governance effectiveness, promoting the use of domestically produced, controllable AI chips, computing platforms, and foundational software. By mastering underlying code, firmware logic, and system architecture, we can achieve a deep integration of safety capabilities and computing systems. Finally, we should actively participate in global AI governance, promoting an open, cooperative, and shared global governance framework. In the face of common challenges, countries should strengthen deep cooperation in strategic alignment, policy communication, risk prevention, and standard formulation, making AI a public good that benefits all humanity.