AI Theory / v0.1
Reinterpreting Intelligence: From the Structural Differences Between Humans and AI
A framework for understanding the structural difference between humans and current AI through the separation of knowledge coverage, structured reasoning, and constraint retention.
Reinterpreting Intelligence: From the Structural Differences Between Humans and AI
Abstract
The development of artificial intelligence is changing how we understand “intelligence.” In the past, knowledge, reasoning, learning, goal retention, real-world adaptation, and task completion were usually discussed together. Because these capabilities have long been bound to the same acting subject in humans, their internal differences were not easy to see clearly. The emergence of large language models has changed this. Current mainstream AI systems display extremely strong structured reasoning capability and knowledge coverage, but in tasks with long chains, many conditions, and strong state dependence, they still easily show goal drift, condition loss, boundary violation, and acceptance failure.
This article argues that to understand the difference between human intelligence and current AI intelligence, we should not look only at knowledge reserves or structure-generation capability. We should also introduce “constraint-retention capability” as an analytical dimension. Intelligence is not only the ability to generate structure; it is also the ability to keep that structure continuously constrained by goals, premises, boundaries, conditions, and feedback. In this sense, intelligence is a constrained structure-generation capability. The main bottleneck of AI intelligence today lies in its ability to maintain constraints.
Keywords
intelligence; large language models; structured reasoning; constraint retention; human cognition; AI capability boundaries
1. The problem: why intelligence needs to be reinterpreted
“Intelligence” is an old and complex concept. It is used in psychology, artificial intelligence, education, management, and social evaluation, but its meaning is not always the same. In psychometrics, intelligence is often associated with IQ, the g factor, and standardized tests. In artificial intelligence research, intelligence is more often understood as a system’s ability to achieve goals, learn strategies, and complete tasks in an environment.
This article does not attempt to replace these existing definitions. It also does not attempt to give intelligence a closed and final definition. It is concerned with a different question: in the comparison between humans and current AI, what internal structures of intelligence were previously mixed together but are now worth separating?
In past discussions, people often mixed together “knowing a lot,” “being able to reason,” “solving problems,” “learning,” “adapting to the environment,” and “completing tasks.” This understanding worked for a long time in human society because, in humans, these capabilities are usually bound to the same continuous acting subject. If a person can solve complex problems over time, that often also means they have stable goals, long-term memory, real-world responsibility, feedback correction, and social constraints.
AI changes this default premise.
Current large language models can generate complex frameworks, unfold multi-level analysis, list key variables, construct classification systems, propose solution paths, and organize scattered information into coherent expression in a very short time. In many scenarios, their structured expression capability already exceeds that of most ordinary people. But at the same time, when tasks become longer, constraints increase, and historical state becomes more complex, they may gradually forget goals, lose conditions, swap problems, or generate an answer that is formally complete but already deviates from the original requirements.
This makes clear a problem that was previously difficult to see separately: a system can be very good at generating structure without being equally good at continuously maintaining constraints.
Therefore, the real significance of AI is not only that it gives us a stronger tool. It also illuminates the internal structure of intelligence itself. It shows that “being smart” in the traditional sense contains at least three different components: knowledge coverage, structured reasoning, and constraint retention.
2. The AI boundary discussed in this article
To avoid conceptual overgeneralization, this article first needs to define the AI object it discusses.
The AI discussed here mainly refers to current mainstream AI systems centered on large language models, especially work systems based on autoregressive generative large language models and organized through external engineering systems such as context windows, prompts, tool calls, retrieval, external memory, task documents, test feedback, and human review.
Therefore, the AI capability boundary discussed here does not refer to all possible forms of artificial intelligence, nor is it a final judgment about future AI. Future AI may of course have stronger long-term memory, online learning, self-updating, task-state management, and constraint-retention mechanisms. What this article discusses is the clear asymmetry, under the current mainstream technical path, between the structured generation capability of large language models themselves and the constraint-retention capability assisted by external systems.
In other words, this article does not deny that external systems can strengthen AI’s ability to maintain constraints. On the contrary, the importance of external systems shows precisely that constraint retention is not a natural accessory of structure generation. It is a capability that must be separately identified, separately maintained, and separately engineered.
This is very important for understanding the boundary of AI capability today. Many debates about AI intelligence are confused because people do not distinguish between the model’s own capability, the model’s temporary capability inside a context, and the system capability formed after the model is organized by external engineering systems. A large language model’s single-turn question-answering performance, a retrieval-augmented question-answering tool, a programming assistant connected to a code repository and test system, and an engineering workflow jointly constrained by task documents, version control, test feedback, and human review are not objects at the same level.
This article discusses a structural problem common to these systems: AI can quickly generate structure, but whether that structure can continuously obey goals, boundaries, and acceptance criteria still needs to be examined separately.
3. Related research and the position of this article
The theoretical discussion in this article is related to several existing research directions, but it does not fully belong to any one of them.
First, regarding the definition of machine intelligence, Legg and Hutter proposed a definition of “universal intelligence” that attempts to formalize intelligence as an agent’s ability to achieve goals across a wide range of environments. This path emphasizes goals, environments, and adaptation, and has had an important influence on machine intelligence research. This article is adjacent to that direction, but it does not attempt to establish a mathematical measure of intelligence. Instead, it focuses on a more easily overlooked internal structure in intelligent performance: the difference between knowledge coverage, structured reasoning, and constraint retention.
Second, regarding human cognitive limitations, research on working memory has long pointed out that humans have clear limits in simultaneously maintaining and manipulating multiple pieces of information. Baddeley’s working memory model shows that humans are not reasoning machines with unlimited capacity. This article accepts that point and further argues that while humans are often weaker than AI in the speed, breadth, and formal completeness of structured reasoning, humans as acting subjects in continuous reality have a more natural foundation for maintaining constraints.
Third, regarding situated cognition and the extended mind, Suchman’s critique of the “planning model” shows that human action is not merely mechanical execution of a prior plan, but continuous adjustment within concrete situations. Clark and Chalmers’s “extended mind” also reminds us that human cognition is not completely sealed inside the head, but couples with notes, tools, environments, and social relationships. This article’s discussion of human constraint retention is close to that tradition: humans maintain constraints not only through abstract reasoning, but also through bodily situation, social responsibility, long-term memory, environmental feedback, and the consequences of action.
Fourth, regarding the reasoning capability of large language models, research such as Chain-of-Thought shows that generating intermediate reasoning steps can significantly improve large models’ performance on complex reasoning tasks. At the same time, research on long-context use also shows that a long context does not mean that a model can use all information inside it stably and evenly. Related work shows that when key information is located in the middle of a long context, model performance may decline significantly. This is related to the constraint-retention problem discussed here: increasing context capacity does not automatically produce an equivalent improvement in constraint retention.
Fifth, regarding external memory, tool use, and agent systems, directions such as RAG, ReAct, MemGPT, and Generative Agents all try, to varying degrees, to extend large language model capability through retrieval, tools, memory management, action feedback, and reflection mechanisms. These studies do not weaken the argument of this article. They instead demonstrate a fact: to make AI more reliable in complex tasks, states, memory, feedback, and control structures outside the model must be introduced.
Therefore, the contribution of this article is not to propose a new AI technical solution, nor to offer a new psychometric model. It attempts to provide an explanatory framework: using the distinction among “knowledge coverage, structured reasoning, and constraint retention” to reinterpret the structural difference between human intelligence and current AI intelligence.
4. Structured reasoning and constraint retention
Structured reasoning is the ability of a system to decompose a problem into levels, relationships, variables, causes, paths, and steps. It includes abstraction, classification, decomposition, comparison, induction, deduction, planning, and solution generation.
AI is extremely strong in this respect. Faced with a complex problem, it can quickly generate an analytical framework, build multi-level classifications, list key variables, organize an argumentative structure, and produce a seemingly complete plan. Relative to ordinary people, AI’s structured reasoning capability is close to excessive in many scenarios.
But structured reasoning is not the same as complete intelligence.
Constraint retention is the ability to continuously maintain goals, premises, boundaries, conditions, priorities, and acceptance criteria during reasoning, decision-making, and execution.
A goal is a constraint because the task must serve a purpose. A premise is a constraint because reasoning cannot be built on assumptions that have been excluded. A boundary is a constraint because some things cannot be done and some conditions cannot be violated. Priority is a constraint because multiple goals must be ordered. Acceptance criteria are also constraints because task completion cannot be judged only by whether the expression is complete, but by whether the result satisfies verifiable requirements.
In short tasks, these two capabilities are not easy to distinguish. The shorter the problem, the fewer the constraints, and the faster the feedback, the easier it is for AI to appear like a complete intelligent subject. As long as it unfolds the problem clearly enough, it can give a satisfying answer.
But in longer, more complex tasks that depend more heavily on continuous state, the number of constraints increases quickly. Goals, background, boundaries, historical state, staged decisions, external conditions, and evaluation criteria all become conditions that must be continuously maintained. At this point, what truly determines intelligent performance is no longer merely whether the system can generate structure, but whether the structure it generates can continuously obey goals, conditions, and real-world boundaries.
To make the concept clearer, the two capabilities can be further operationalized.
Structured reasoning capability is usually expressed as whether a system can identify problem structure, decompose variables, construct categories, establish causal relationships, generate paths, and propose plans.
Constraint-retention capability is expressed as whether a system can continuously maintain goals, obey premises, retain boundary conditions, track historical state, maintain priorities, accept acceptance criteria, and revise plans after feedback appears.
The two are related, but they are not the same. A system can generate a large amount of complex structure without ensuring that these structures always serve the original goal. Conversely, a system may generate structures more slowly, but still continuously hold key boundaries in real tasks.
This is precisely where the difference between humans and current AI is most worth analyzing.
5. Current AI’s structural advantage and structural weakness
Current AI’s strength first comes from its ability to generate structure.
It can turn vague problems into clear frameworks, organize scattered information into progressive arguments, break complex goals into multiple steps, and turn experiential judgments into explicit variables. This capability is highly valuable in writing, coding, research, teaching, solution design, and knowledge organization.
But precisely because AI is so good at generating structure, people can easily mistake structure generation for complete intelligence. As long as an answer has complete form, clear levels, and coherent logic, it is easily regarded as a sign of being “smart.”
Yet in complex tasks, the real problem is often not whether there is structure, but whether the structure remains constrained.
AI can generate a detailed plan for a project, but the plan may ignore real resource limits. It can generate a repair plan for a problem, but the plan may break existing boundaries. It can write an article according to a user request, but the article may gradually drift away from the original thesis. It can process multiple conditions at the same time, but when the number of conditions increases and the chain grows longer, some key constraints may gradually disappear.
This is not merely a “hallucination” problem. Hallucination usually refers to factual error or fabricated information. Constraint loss is deeper. It means the system lacks a stable mechanism for continuously maintaining task boundaries. Even if every step appears reasonable, the whole may already have deviated from the original goal.
This problem is related to how current mainstream large language models operate.
Large language models certainly have temporary computational states during a single inference process, such as context representations, attention computation, and intermediate activations. But these states usually are not automatically written back into model parameters after inference ends, nor do they naturally form stable, inheritable internal knowledge structures across tasks and sessions. Multi-turn dialogue appears continuous, but that continuity mainly comes from the external system putting historical messages back into the context. In other words, the model does not live in continuous time like a human. It rereads the historical material passed in during the next inference.
Once the context becomes too long, history is compressed, or key information is omitted, constraints that should continue to hold may weaken or even disappear.
A similar problem exists within a single inference. Constraints may temporarily exist in the current context and reasoning process, but the model itself does not naturally distinguish which conditions are merely background information, which must always be maintained, and which boundaries mean task failure once violated. As the reasoning chain grows longer, goals, premises, limitations, and acceptance criteria all compete for attention in the same context. Structure can continue to be generated, but constraints may not be maintained at the same time.
Therefore, the special structure of current AI is this: it has extremely strong structured reasoning capability, but lacks equally stable constraint-retention capability.
External records, task documents, retrieval systems, test feedback, human review, and rollbackable processes can of course help AI maintain constraints to some extent. But this precisely shows that constraint retention is not a natural accessory of structure generation. It is a capability that needs to be separately identified and separately maintained.
6. The opposite structure in humans
The difference between humans and current AI happens to present an opposite structure.
Human structured reasoning capability is usually limited. Ordinary people facing complex problems often find it difficult to maintain many layers of variables for a long time, and also find it difficult to unfold complex causal chains steadily. Human working memory is limited, attention becomes fatigued, emotions interfere with judgment, and insufficient experience also limits problem decomposition. Therefore, in the speed, breadth, and formal completeness of structure generation, humans are often much weaker than AI.
But relative to current mainstream large language models, humans usually have a more stable natural foundation for constraint retention.
This is because humans are not isolated reasoning machines. They are acting subjects situated in continuous reality. Humans have bodily situations, life continuity, long-term memory, responsibility relationships, social evaluation, and real consequences from task failure. Together, these factors form a natural constraint loop.
An engineer may not be able to list the perfect system architecture all at once, but they know not to break production, ignore failing tests, casually delete user data, or treat unverified assumptions as facts. A manager may not be able to formally describe all organizational variables, but they know that goals, resources, risks, time, and responsibility cannot be casually decoupled. A writer may not be able to generate a perfect structure in one attempt, but they know an article must always serve its core judgment and cannot rely only on conceptual stacking to maintain surface completeness.
This capability does not come entirely from abstract reasoning. It comes from human continuity as real-world actors. Humans bind tasks, responsibility, risk, and consequences together. Even when humans are weaker than AI in structured reasoning, they can often maintain key constraints more naturally in real tasks.
This is not to say that humans are always reliable. Humans also forget, have biases, become emotional, deceive themselves, avoid acceptance, and lose constraints in complex tasks. This article does not claim that humans are stronger than AI in all situations. Its claim is that humans and current AI have different intelligence structures. AI is strong in structure generation, while humans are strong in natural constraint loops. AI’s structural capability is more explicit, while human constraint retention is more embedded in real life.
This shows that there is a long-underestimated part of human intelligence: not merely thinking in complex ways, but continuously not losing the goal, not breaking boundaries, not forgetting conditions, and not avoiding acceptance.
7. Model corpora and human knowledge: static knowledge is not intelligence itself
When discussing AI intelligence, another easily confused object must be separated: knowledge and intelligence.
Model training corpora and existing human knowledge can be compared approximately. Neither is reasoning that is currently happening. Both are static structures that have been deposited before reasoning begins. More vividly, both can be understood approximately as a kind of “solidified connection”: human knowledge is deposited in long-term memory, experience, and neural connections, while the knowledge a model gains from training corpora is deposited in trained parameters and statistical structures.
This analogy does not require the two mechanisms to be identical. Human memory, experience, concepts, and skills are certainly not the same thing as language patterns and world knowledge inside model parameters. But functionally, both provide background material for reasoning. Structured reasoning and constraint retention both occur on top of these static forms of knowledge.
Therefore, the high intelligence AI appears to show in chat scenarios often mixes three sources: corpus coverage, structured reasoning, and constraint retention.
If a question happens to fall inside a domain highly covered by the model’s training corpus, the model may not need much complex reasoning to give a mature, fluent, expert-like answer. In that case, part of the “smartness” it displays comes from knowledge coverage rather than from an ongoing intelligent activity. Knowledge coverage is valuable, of course, but it is not the same as intelligence itself.
The same is true for humans. If a person is familiar with a domain and can quickly give an answer, that does not necessarily mean they are performing high-intensity reasoning. It may simply mean their existing knowledge is rich enough. In traditional education, “being learned” and “being smart” have often been mixed together. AI amplifies this confusion. Because models have corpus coverage far beyond that of any individual human, they can appear unusually smart in many chat-style question-answering scenarios.
But the real difference in intelligence is exposed where corpus coverage is insufficient.
Real problems often contain a large amount of information generated on site: current state, historical baggage, temporary constraints, role relationships, task goals, evaluation criteria, and changes in context. This information usually does not fully exist in the model’s training corpus. The model cannot rely only on language patterns solidified during training to solve the problem. It must reorganize information in the current task, form structure, maintain constraints, and accept feedback.
Here a key difference between AI and humans appears. AI has strong reasoning capability and can quickly generate new structured information inside the current task. But these newly generated structures usually cannot be continuously solidified by the model as static knowledge for later reasoning. A conversation, a reasoning process, a plan, or a correction experience will not naturally become stable internal knowledge available to the model in the next inference unless it is externally recorded and passed in again.
Humans are different. Humans can gradually solidify new knowledge during action. A failure, an adjustment, or the discovery of a boundary condition may enter long-term experience and become part of later judgment. Human learning is slower, and human structured reasoning is not as broad as AI’s, but humans can gradually deposit newly generated structures from the scene into stable knowledge and naturally call on them in later tasks.
Therefore, AI’s weakness is not merely that it “does not remember the conversation.” More precisely, it cannot, like humans, continuously solidify effective structures generated during reasoning into its own knowledge system. It can use these structures temporarily in the current task, but it does not naturally turn them into long-term inheritable cognitive assets.
This also explains why AI can easily appear highly intelligent in chat scenarios, while its boundaries are more easily exposed in real tasks. Chat questions are often covered by training corpora, and the model only needs to locally organize existing knowledge structures. Real tasks depend heavily on on-site knowledge, where corpus coverage is insufficient, and therefore test structured reasoning and constraint retention more directly.
So when evaluating AI intelligence, knowledge coverage must be separated from intelligent performance. The more sufficient corpus coverage is, the more the model looks like it is calling on static knowledge. The more complex on-site constraints are, and the more insufficient corpus coverage is, the more the model needs to display real intelligent capability.
8. Reinterpreting traditional judgments of intelligence through the new structure
If we use “structure generation” and “constraint retention” to observe intelligence again, many traditional judgments of intelligence can be explained more clearly.
Mathematical ability is often regarded as a typical expression of high intelligence. The reason is not only that mathematics requires abstract reasoning, but also that mathematics is a highly constrained reasoning system. A mathematical proof cannot merely pursue formal complexity. Every step must be constrained by definitions, axioms, premises, and logical rules. Many people can understand the general meaning of a conclusion but cannot complete a rigorous proof. The gap is often not that they have no structured thinking at all, but that they cannot continuously maintain formal constraints across multiple reasoning steps.
Engineering ability is similar. Excellent engineers do not merely propose complex solutions. More importantly, they can maintain constraints such as interfaces, performance, compatibility, testing, maintainability, and release risk in real systems. People with high engineering capability usually are not strong because they have the most ideas, but because they know which conditions cannot be broken inside a complex structure. By contrast, if someone can always propose grand plans but repeatedly ignores real boundaries, historical baggage, and acceptance criteria, their structured expression may be strong, but their engineering intelligence is not high.
Writing ability can be interpreted in the same way. A good theoretical article does not simply unfold more concepts. It continuously maintains theme, thesis, levels, tone, reader comprehension cost, and argumentative goal throughout the unfolding process. Many articles appear information-dense but leave no clear conclusion after reading. This is often caused by excessive structure generation and insufficient constraint retention. Truly good writing makes complex structure consistently serve one core judgment.
Management ability is also not merely communication ability or decision speed. Managers face complex systems with multiple goals, roles, resources, and risks. Excellent managers can continuously maintain goals, priorities, responsibility boundaries, and feedback mechanisms amid organizational noise. Weak managers may also create many rules, hold many meetings, and propose many slogans. But if these structures cannot stably serve organizational goals, they are only low-constraint structural expansion.
Test-taking ability reflects another combination. Standardized tests usually examine both structured reasoning and rule constraints. High-scoring students not only need to solve problems. They must also accurately identify conditions, scope, answer format, and grading criteria. Some people think actively but often misread questions. Some people are knowledgeable but frequently answer something other than what was asked. These failures are essentially insufficient constraint retention.
Business judgment can also be placed within this framework. A business judgment is not evaluated by whether its narrative is grand, but by whether it remains constrained by cash flow, competitive structure, user demand, organizational capability, time window, and risk boundaries. Many business stories are formally complete, but once detached from these constraints, they are only structured imagination.
From this, we can see that what is traditionally called “high intelligence” is rarely just strong structured reasoning. People who have been validated over time as highly intelligent are usually people who can continuously maintain constraints inside complex structures. They can not only unfold a problem, but also close it down. They can not only propose possibilities, but also hold boundaries. They can not only generate explanations, but also accept verification.
This framework also explains why some people “look smart” but perform unstably in real tasks. They may have strong structure-generation capability and can express, classify, and propose plans quickly, but they lack continuous constraint-retention capability. Once a problem enters long-term execution, real feedback, and complex responsibility relationships, they easily drift, lose focus, lose control, or avoid acceptance.
Therefore, intelligence should not be judged only by whether a person can say complex things. It should be judged by whether they can make complex content continuously serve goals, conditions, and real-world boundaries.
9. Theoretical boundaries and scope of application
This article proposes a theoretical explanatory framework. It is not a psychometric definition of intelligence, nor is it a final judgment about all forms of AI.
First, the intelligence discussed here is not the same as IQ, the g factor, or other psychometric concepts. This article focuses on a broader cognitive structure in complex problem handling: how a system forms structure, handles relationships, maintains goals, obeys boundaries, and completes tasks.
Second, the AI discussed here mainly refers to current mainstream AI systems centered on large language models, as well as the external engineering systems that organize these models’ work. This article does not deny that future AI may obtain stronger long-term memory, online learning, self-updating, and constraint-retention mechanisms. On the contrary, precisely because these mechanisms have not yet become stable internal capabilities of mainstream large language models, constraint retention needs to be raised separately.
Third, this article does not attempt to prove the intelligence difference between humans and AI completely through a single experiment. Constraint-retention capability in real complex tasks involves long-term state, on-site knowledge, responsibility relationships, feedback loops, acceptance criteria, and external systems. Compressing these variables into a short-task experiment often loses the complexity of the problem itself. Short tasks, standardized tasks, and local metrics certainly have research value, but they are not enough to fully present the structural difference discussed here.
Therefore, the contribution of this article is not to provide a new test score, but to offer an analytical framework: to distinguish knowledge coverage, structured reasoning, and constraint retention, and to use this framework to explain why current AI appears extremely smart in chat, demonstrations, and standardized tasks, while its boundaries are more easily exposed in long-chain, multi-condition, strongly state-dependent real tasks.
10. Conclusion: intelligence is constrained structure-generation capability
The comparison between AI and humans gives us an opportunity to reinterpret intelligence.
AI is strong in structured reasoning and has enormous corpus coverage. It can quickly unfold problems, generate frameworks, construct paths, and organize expression. Humans are strong in natural constraint retention and can continuously solidify new experience generated on site into knowledge. Humans live in continuous reality, with task sense, responsibility loops, long-term memory, and real-world feedback. Therefore, they can more easily maintain key goals and boundaries in real tasks.
This does not mean humans will always be more reliable than AI. Nor does it mean AI cannot acquire stronger constraint-retention capability. On the contrary, future AI may of course obtain more stable constraint-retention mechanisms through stronger models, longer memory, better tools, and external systems.
But today, this difference remains key to understanding the boundary of AI capability. AI’s surprising performance in chat, demonstrations, and many standardized tasks does not necessarily mean it already has high intelligence in the full sense. These performances often mix corpus coverage, structured reasoning, and limited constraint retention. The more a task enters long chains, many conditions, and strong feedback, the more the importance of constraint retention becomes visible.
Therefore, intelligence should be reinterpreted from the structural differences between humans and AI.
AI reveals several components of intelligence that have long been blended together: knowledge coverage, structured reasoning, and constraint retention. Future evaluations of AI intelligence should not look only at whether it can call on more knowledge or generate more structure. They should also look at whether, when on-site knowledge is insufficient and task constraints are complex, it can continuously form structure, maintain constraints, and turn effective experience into inheritable state.
Intelligence is a constrained structure-generation capability. The main bottleneck of AI intelligence today lies in its ability to maintain constraints.
References
Baddeley, A. D., & Hitch, G. (1974). Working Memory. In G. H. Bower (Ed.), The Psychology of Learning and Motivation.
Baddeley, A. (2012). Working Memory: Theories, Models, and Controversies. Annual Review of Psychology, 63, 1–29.
Clark, A., & Chalmers, D. (1998). The Extended Mind. Analysis, 58(1), 7–19.
Legg, S., & Hutter, M. (2007). Universal Intelligence: A Definition of Machine Intelligence. Minds and Machines, 17, 391–444.
Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W., Rocktäschel, T., Riedel, S., & Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 2020.
Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., & Liang, P. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics.
Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., & Gonzalez, J. E. (2023). MemGPT: Towards LLMs as Operating Systems.
Park, J. S., O’Brien, J. C., Cai, C. J., Morris, M. R., Liang, P., & Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. UIST 2023.
Suchman, L. A. (1987). Plans and Situated Actions: The Problem of Human-Machine Communication. Cambridge University Press.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention Is All You Need. NeurIPS 2017.
Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E., Le, Q., & Zhou, D. (2022). Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. NeurIPS 2022.
Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.