1 Introduction
With the continuous maturation of deep learning technologies, the ongoing accumulation of big data resources, and the explosive growth in computational performance, the field of Artificial Intelligence (AI) science and technology has, after nearly seven decades of development and evolution, gradually transitioned from computational intelligence and perceptual intelligence into the “new frontier” of cognitive intelligence. Since 2023, LLM-driven general intelligence has become the focal point of global AI research, signaling a transformative phase in the development of artificial intelligence(
Bubeck et al., 2023). Unlike previous specialized intelligence systems that were limited to specific tasks or domains, the exploration, and practice of LLM technology in general intelligence concentrate on fundamental human cognitive and mental capabilities, such as language proficiency, common sense and reasoning abilities, spatial understanding, and knowledge processing (
Brown et al., 2020;
Lehr et al., 2024;
Petroni et al., 2019;
Wake et al., 2024). Furthermore, LLMs are comprehensively competing with top human performance across fields such as Science, Technology, Engineering, and Mathematics (STEM) (
Hendrycks et al., 2020;
Romera-Paredes et al., 2024;
Singhal et al., 2023;
Vemprala et al., 2024;
Wei et al., 2022).
AI systems with general intelligence demonstrate potential for enhancing the fundamental cognitive and mental capabilities of human thinking and intellect. Their problem-solving approaches are not based on “first principles” (logocentrism) but rather on empirical learning from observational data (
Fodor and Pylyshyn, 1988). These systems possess the ability to extract intuitive knowledge from vast amounts of data, enabling them to provide flexible and effective solutions when confronted with complex and ever-changing new scenarios. Fig.1 illustrates the primary research framework and technical characteristics for LLMs, highlighting the synergy between the technical layer, characteristic layer, and application layer. It comprehensively demonstrates the capability of LLMs, from foundational technical optimization to multi-domain task support, as well as their potential and adaptability in complex scenarios. Regarding LLMs, we have established a foundational understanding of their functionalities through various practical experiences, achieving a “knowing that” level of comprehension. However, attaining a “knowing why” understanding remains exceedingly difficult, thereby making it impossible to achieve a “knowing it necessarily” level of certainty. The reliability issues associated with LLMs significantly constrain their use across diverse contexts. Therefore, exploring and contemplating the technical characteristics and intelligent features of large language models can help us further understand this new species of general intelligence technology. This, in turn, provides a cognitive foundation for the transformation of decision-making paradigms in the era of large models (
Burton et al., 2024;
Duan et al., 2019;
Meng, 2024). The technical characteristics, intelligent features, and thinking patterns of LLMs are illustrated in Fig.1, and will be discussed in detail below.
2 Technical characteristics
At the end of 2022, OpenAI released a revolutionary large language model technology product—ChatGPT. ChatGPT employs the Transformer architecture based on the self-attention mechanism (
Vaswani, 2017), enabling deep neural networks to process large-scale, multi-source heterogeneous text data in parallel and effectively capture long-range dependencies. The training and learning process of LLMs can be primarily divided into three stages:
Stage I: Pre-training on High-Quality Corpus Data: The first stage focuses on collecting and processing vast amounts of unlabeled data (
Kaplan et al., 2020). This data can originate from various sources on the internet, including texts, news, blogs, and forums. After cleaning and processing, the data are fed into the large language model with the objective of enabling the model to learn statistical patterns and semantic information about the language.
Stage II: Instruction Fine-Tuning Based on Pre-training: The second stage involves further training the model for specific tasks by implementing instruction fine-tuning (
Devlin and Toutanova, 2019). Through fine-tuning, the model can adapt to different application scenarios and task requirements.
Stage III: Enhancing Model Performance Using Human Evaluation Feedback: The third stage utilizes Reinforcement Learning from Human Feedback (RLHF) (
Christiano et al., 2017) to optimize the model’s performance. This approach further enhances the model’s generative capabilities and alignment with human expectations by incorporating feedback from human evaluations.
Overall, LLMs exhibit four key technical characteristics: ① Enhanced Reasoning Abilities: LLMs achieve the generation of fluent and natural language by continuously enhancing their reasoning capabilities, enabling them to perform deductive reasoning with knowledge and inductive reasoning with common sense. ② Comprehensive Knowledge Representation: These models efficiently express nearly all human knowledge about the world. They integrate historical and contemporary knowledge from various cultures and eras into their knowledge-based neural systems, achieving comprehensive coverage of general knowledge across all domains. ③ Robust Generalization Capabilities: LLMs possess strong generalization abilities, facilitating intelligent transfer and emergence. This enables a single model to provide generalized solutions for a wide range of tasks. ④ Human-like Interaction, Intent Recognition, and Logical Reasoning: These models exhibit human-like interaction capabilities, intent recognition, and logical reasoning abilities. Their interactive expressions conform to human natural language habits, ensuring alignment with human values.
Today, large language model technology is evolving into a universal intelligence system, capable of efficiently generating natural language and aligning with human values, thanks to its powerful chain-of-thought reasoning abilities, extensive knowledge integration, robust generalization capabilities, and smooth human-machine interaction.
3 Intelligent features
LLMs have not only achieved leapfrog improvements in technical performance—such as common sense knowledge, deep reasoning, strong generalization, and natural and fluent human–machine interactions—but they also represent the most sophisticated imitation and efficient expression of human thinking and mental activities to date. Researchers from Stanford University investigated the cognitive processes of GPT-3 (
Tamkin et al., 2021). The study indicates that the cognition of LLMs is a form of In-context Learning (ICL), an intelligent paradigm that allows language models to learn tasks by being provided with only a few examples in the form of case demonstrations. ICL is a type of analogical reasoning where contextual learning adjusts the model based on “example-based norms,” effectively enabling the model to dynamically adapt to new tasks where the input distribution significantly differs from the training distribution (
Dong et al., 2024;
Xie and Min, 2022).
In the interdisciplinary field of artificial intelligence and linguistics, a groundbreaking study has garnered widespread international attention. The 2024 Annual Best Paper Award from the Association for Computational Linguistics (ACL) was awarded to this research (
Kallini et al., 2024). This paper, through a series of ingenious experiments, systematically tested and refuted a hypothesis regarding “impossible languages” proposed by Noam Chomsky, one of the most authoritative linguists globally (
Chomsky et al., 2023). The research demonstrates that LLMs struggle to learn “impossible languages,” a new finding that supports Geoffrey Hinton’s viewpoint—that linguistic intelligence does not stem from innate human endowments, and the essence of intelligence is learning (
LeCun et al., 2015). Geoffrey Hinton likens the learning process of neural networks to the evolutionary process in biology; both animals and artificial intelligence systems can exhibit intelligent behavior through experiential learning. Language acquisition does not require a predefined symbolic system but is achieved through the extraction and integration of relevant features in data (
Hinton and Salakhutdinov, 2006;
Wiggers, 2018). Through exposure to massive data sets, neural networks are able to capture the statistical regularities of language, forming the basis for coherent linguistic expression.
Building on this, analogical reasoning—a core component of human cognition—has also been observed in neural systems. The perspective that neural networks exhibit intelligent behavior through experiential learning is supported by a substantial body of research (
Bengio et al., 2013;
Gui et al., 2024;
Hinton et al., 2006;
Marcus, 2003). Studies have shown that the distributed representation capabilities of neural networks can decompose complex concepts into multiple features and generate accurate language and knowledge by adjusting the weights of these features (
Hinton et al., 2006). By learning from massive data sets, neural networks can capture statistical regularities in language, thereby generating coherent linguistic expressions. This learning mechanism is akin to how organisms gradually optimize their behavioral and perceptual systems through evolution to adapt to their environments and make reasonable decisions. Gary Marcus’s research points out that analogical reasoning is at the core of human cognition, especially in complex tasks. Neural networks, through “analogies” learned from patterns in data, can make inferences in new environments, demonstrating high flexibility (
Marcus, 2003). Yoshua Bengio further explores how neural networks capture relationships between complex concepts by learning features layer by layer, similar to human analogical reasoning abilities (
Bengio et al., 2013). Yann LeCun proposed self-supervised learning as a core principle in the development of neural networks, emphasizing that systems automatically learn representations by being exposed to vast amounts of data without relying on explicit symbolic representations (
LeCun and Misra, 2021).
This empirical learning paradigm sharply contrasts with the logocentric view that intelligence must be grounded in first principles. Unlike traditional first-principles solvers, LLMs mimic intelligence and achieve chain-of-thought reasoning through “simulation” (empirical intuition). LLMs represent a form of digital intelligence that transcends individual identity, with mental activities that reflect characteristics similar to human thinking. They possess the ability to extract intuitive knowledge from vast amounts of data. Overall, the intelligent features of LLMs are manifested in three aspects:
① Memory-Driven Core Characteristics: This process is highly similar to the primitive ways of human cognition, where continuous observation of the external environment in real-world contexts leads to cognitive judgments and subsequent natural responses. LLMs develop implicit cognitive priors not through abstraction, but through saturation: massive exposure to heterogeneous data gives rise to a distributed expression encoded in high-dimensional vector spaces. This knowledge is not interpretable in symbolic terms but nonetheless enables the model to behave as if it “understands.” This marks a shift from rationalist deduction to empirical abstraction.
② Robust Data-Driven Learning Capabilities: The internal state of an LLM—the learned parameters—is not a passive storage but an active encoding of task-relevant distributions, accumulated through gradient-based optimization. During training, LLMs continuously optimize their parameters to better adapt to data distributions, with these neural network parameters encapsulating the patterns and regularities in the data. This process is analogous to how humans deepen memory and form experiences through continuous trial and error and repeated practice.
③ Strong Generalization Ability: LLMs generalize via analogy, interpolation, and context-sensitive pattern completion—drawing on distributed representations that encode similarity, co-occurrence, and abstract mappings. This allows LLMs to rapidly adapt to new prompts or unseen compositions, simulating reasoning without recursive rule application. Well-trained LLMs can quickly recognize new data that are similar to the training data and perform reasoning and judgment based on existing memory (i.e., learned patterns). This experience reuse mechanism allows LLMs to excel in handling similar tasks.
4 Thinking patterns
In 1950, Alan M. Turing published his groundbreaking classic paper “Computing Machinery and Intelligence” in the journal Mind (
Turing, 2009). He began by stating, “I propose to consider the question: Can machines think”? From Turing’s perspective, if a machine can sufficiently obscure its differences from humans in a simulation game (through the Turing Test), it must be acknowledged that the machine possesses human-like thinking and intellect. Turing argued that intelligence is not defined by the internal processes of thought but by the communication and output of these processes. Anticipating objections to his “anti-scientific commonsense” viewpoint, Turing listed nine counterarguments in his paper and systematically refuted each one. Notably, he addressed the “argument from consciousness”: the assertion that machines cannot have emotions, cannot feel happiness, sadness, or frustration, whereas intelligent humans clearly can, and therefore machines cannot be intelligent. Turing rebutted this by stating that if the only way to determine whether a person can think is to become that specific person, then meaningful communication about thought becomes difficult. We should not deny machine intelligence merely because machines lack emotions; the definition of intelligence should not be confined solely to emotional experiences. If we are unwilling to acknowledge that machines possess attributes of consciousness, we must also be unwilling to acknowledge that other humans possess consciousness. In its most extreme form, this viewpoint suggests that to affirm whether a machine can think, the only way is to become that machine and experience its thought processes firsthand. Similarly, to know whether someone is thinking, the only way is to become that particular person. Therefore, if we are unwilling to acknowledge that machines possess attributes of consciousness, we must also be unwilling to acknowledge that other humans possess consciousness.
For decades, the Turing Test has remained a focal point of controversy in the field of artificial intelligence. On one hand, due to humanity’s extremely limited understanding of the essence of its own thinking, it is challenging to provide a definitive intrinsic definition of intelligence and thought based on internal mechanisms. Consequently, the Turing Test has been employed as a means to guide and advance the development of artificial intelligence. As a standard for measuring machine intelligence, the Turing Test offers a relatively objective and operational evaluation framework. On the other hand, skepticism about whether machines that pass the Turing Test truly possess intelligence has never ceased. Although these machine systems have achieved human-level performance in input and output, and even surpassed humans in certain aspects, they remain oblivious to the intrinsic meanings and essences of the objects they process. This state of “knowing that without knowing why” has deeply cast doubt on the authenticity of machine intelligence. In 2023, GPT-4.0 made significant strides in natural language understanding, demonstrating higher levels of reasoning and language generation capabilities. Despite this remarkable progress, public skepticism regarding the “ignorance” of LLMs has not been entirely alleviated, particularly concerning their lack of genuine understanding, reasoning, and interpretability. For instance, Yann LeCun argues that the Turing Test fundamentally cannot measure intelligence and that LLMs do not truly comprehend the meanings of symbols (
Browning and LeCun, 2023). Bender contends that although LLMs perform exceptionally well in certain tasks, their performance may rely more on training data and algorithm optimization, representing nonlinear fitting of high-dimensional data rather than true intelligence (
Bender and Koller, 2020).
Currently, it remains challenging to provide a universally satisfying answer as to whether LLMs possess human-like thinking and intellect. However, as a technological novelty that has achieved breakthrough progress in the direction of general intelligence, LLMs exhibit astonishing language generation and comprehension capabilities. They can simulate and understand the complexity and diversity of human language, with linguistic intelligence that even surpasses humans in certain aspects. This not only significantly enhances economic efficiency and boosts technological productivity but also exerts profound transformative impacts on human social organization and lifestyles. In the past, all technological productivity played merely a tool role. Today, LLMs can confuse their differences from humans in “simulation games,” thereby attaining a certain level of subjectivity. A wise society characterized by human-machine collaboration (
Ren et al., 2023) and human- machine Symbiosis (
Lu et al., 2021) is emerging. It is imperative for us to contemplate and explore beyond their role as productivity tools: What does artificial intelligence, which “simulates, expands, and extends human intelligence,” truly mean for human society? What new opportunities and crises lie ahead?
5 Summary and perspectives
LLMs represent a novel technological species in the field of general intelligence. Through empirical learning and the integration of vast amounts of data, they exhibit human-like intelligence characteristics. Technologically, they possess general intelligence capabilities such as common sense, deep reasoning, strong generalization, and natural human-computer interaction. Their intelligence expression is achieved through direct acquisition of “object–object” and “event–event” relationships, eliminating the intermediate processes of conceptual abstraction and symbolic logic processing. In terms of thought, they exhibit human-like traits such as contextual understanding, analogy, and intuitive reasoning. It can be said that LLMs represent a global, holistic, and systemic expression of human cognition. Despite the limitations in innovation and interpretability due to their empirical learning-based thinking patterns, the emergence of LLMs marks a significant advancement in artificial intelligence—toward the realistic simulation and cognitive automation of human thought processes. Theoretically, LLMs can adapt to a wide range of complex, open-ended scenarios, leveraging their extensive contextual understanding and pattern recognition capabilities. This paradigm shift encourages a rethinking of traditional AI approaches, highlighting the potential for more flexible, adaptive, and context-aware systems.
PDF
(775KB)
