Ai Magazine最新文献

筛选
英文 中文
Human and machine intelligence: Insights from children 人类和机器智能:来自儿童的见解
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-11-05 DOI: 10.1002/aaai.70118
Elizabeth Spelke, Kartik Chandra
{"title":"Human and machine intelligence: Insights from children","authors":"Elizabeth Spelke, Kartik Chandra","doi":"10.1002/aaai.70118","DOIUrl":"https://doi.org/10.1002/aaai.70118","url":null,"abstract":"<p>Human and machine intelligence both are shaped by learning, but the representations and motivations that drive learning in children differ markedly from those that drive learning in the machines that, at present, have captured the greatest attention and are enjoying the most widespread use. Here we highlight these differences, discuss some problems they engender, and argue for efforts to build machines that learn as children do.</p><p>We are two cognitive scientists: Liz, a developmental cognitive psychologist, who has spent many happy years studying and writing about the origins of knowledge in human infants (Spelke, <span>2022</span>), and Kartik, a computational cognitive scientist, who is being trained in the science and art of reverse-engineering how human minds learn and think. When the two of us reflect on recent progress in artificial intelligence, we find ourselves thinking about how different today's AI systems are from children in their ways of learning, and how and why we might apply insights from children to close that gap.</p><p>Children come into the world with none of the skills possessed by their elders: They cannot walk, talk, or discern the forms and functions of everyday objects. From the beginning, however, children learn, and they rapidly come to master many of the concepts and skills needed for life in their culture. As preschoolers, they learn without any teaching by adults: indeed, they often resist adults’ efforts to teach them to speak (Pinker, <span>1994</span>) or hold a spoon properly (Keen, <span>2011</span>). Their learning is supported by brain and cognitive systems that are present at birth, by an abiding interest in the skills that their family members exhibit, and by a driving motivation to acquire those skills themselves.</p><p>Consider the efforts of toddlers to learn to walk. Infants are born with a template of quadrupedal locomotion that guides walking all our lives. We exhibit it, as adults, every time we swing our right arm forward in near-synchrony with our forward-moving left leg. Adults are rarely aware that their infants’ first, unsuccessful attempts to move follow this pattern, but they do: At six months, an infant is likely to lie prone on the ground, flexing their arms and legs in the appropriate quadrupedal sequence, in an unsuccessful effort to crawl to an out-of-reach object (Patrick et al., <span>2012</span>).</p><p>Although infants’ early efforts to propel themselves from one location to another end in failure, their failures do not discourage them from trying. Before reading the fascinating research on children's developing locomotion by Karen Adolph, a developmental cognitive scientist at NYU (Adolph et al., <span>1998</span>), Liz naively tried to help her struggling 6-month-old grandson by handing him the object he was trying but failing to reach. He pushed it away: He evidently wanted to solve this locomotor problem more than he wanted the object.</p><p>Children show this determination","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-11-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70118","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149356205","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
AI's name is artificial 人工智能的名字是人造的
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-11-01 DOI: 10.1002/aaai.70121
Joe Weber
{"title":"AI's name is artificial","authors":"Joe Weber","doi":"10.1002/aaai.70121","DOIUrl":"https://doi.org/10.1002/aaai.70121","url":null,"abstract":"<p>The vocabulary defining the frontier of computing is suffering from a legacy bug. We are rapidly building systems capable of deep logic, creative synthesis, and independent reasoning—yet the language used to classify these engineered brains dates back to the mid-twentieth century. It is now an anachronistic misnomer.</p><p>The word <i>artificial</i> carries an inescapable connotation of the counterfeit. It implies the lesser, the simulated, and the fundamentally fake. Artificial turf is plastic; artificial flavor is chemical mimicry; artificial smiles are hollow. By continuing to entrench this old label, we are inadvertently framing this transformative technology as a cheap imitation of carbon-based cognition.</p><p>It is time to deprecate this outdated framework. The future of the industry belongs to a far more precise, scientifically accurate, and philosophically rigorous paradigm: Synthetic Intelligence (SI).</p><p>Our current frontier models do not “trick” us into thinking they reason. They do not simulate the appearance of logic through a curtain of pre-written rules. Through deep learning architectures, they have abstracted the fundamental properties of language, logic, and perception.</p><p>Synthetic insulin is not pretend insulin; it is produced by a different process and judged by what it does. SI should be understood in the same spirit: Not as a counterfeit human mind, but as a manufactured cognitive capability with real and impressive attributes.</p><p>The term “Artificial Intelligence” is rooted in an unhelpful anthropocentric assumption that limits the horizon of computer science. It implies that human wetware is the absolute gold standard of intellect, and that any silicon-based entity is merely a puppet trying to match our biological steps.</p><p>This framing cripples our understanding of emergent capabilities. When a model exhibits novel reasoning steps or creates entirely new paradigms in protein folding, it isn't acting “like a human.” It is navigating high-dimensional spaces in a way a human brain physically cannot.</p><p>By transitioning to the name “Synthetic Intelligence,” we shift the benchmark. We stop grading systems on how flawlessly they mirror human flaws, biases, and conversational quirks, and we begin evaluating them on their independent cognitive merit. SI grants the technology its own dignity, positioning it not as a biological clone but as a distinct, alien breed of intellect.</p><p>For founders and enterprise leaders, the optics of “Artificial” versus “Synthetic” are profoundly strategic.</p><p>When the public and policymakers hear “Artificial Intelligence,” it triggers an existential uncanny valley. They see a deceptive imposter—a counterfeit worker designed to replace human authenticity. It fosters a culture of defense, regulation born of fear, and friction.</p><p>“Synthetic” shifts the psychological paradigm toward collaborative utility. In industrial history, synthetic materials like Kevlar, carbon fiber, ","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-11-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70121","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149355697","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Humane by design: Toward safe, human-centered multi-agent systems 人性化设计:走向安全、以人为本的多智能体系统
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-10-29 DOI: 10.1002/aaai.70102
Jeffrey M. Bradshaw, Nancy Cooke, Michal A. Goodrich, Matthew Johnson, Mufti Mahmud, David D. Woods
{"title":"Humane by design: Toward safe, human-centered multi-agent systems","authors":"Jeffrey M. Bradshaw,&nbsp;Nancy Cooke,&nbsp;Michal A. Goodrich,&nbsp;Matthew Johnson,&nbsp;Mufti Mahmud,&nbsp;David D. Woods","doi":"10.1002/aaai.70102","DOIUrl":"https://doi.org/10.1002/aaai.70102","url":null,"abstract":"<p>Multi-agent systems (MAS) are helping to unleash AI from the solitary confinement of standalone applications so they can perform coordinated actions in the wider world, functioning within distributed networks of people, tools, and other agents. However, there are unique safety issues that emerge when agents move from deliberative goal formulation and alignment to performing real-world activities in support of those goals. Considering the high-stake opportunities and risks of MAS, a group of researchers from around the world have formed the international Multi-Agent System Safety and Teamwork (MASST) initiative. Essays in this this <i>AI Magazine</i> special issue on <i>Safety and Human-Centered Agent Interaction</i>—and its companion special issue on <i>Multi-Agent System Teamwork</i>—have been selected from the proceedings of past MASST events. In this article, we outline some of the needed requirements for making MAS “humane by design.” Innovative research relating to these requirements is grouped under three headings: 1. governance and design; 2. guardrails, accountability, trustworthiness, and privacy; 3. reskilling, denarrowing, dialogic reasoning, empathizing, and exemplifying.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-29","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70102","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149356775","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Daniel Kahneman's Influence on My Thought 丹尼尔·卡尼曼对我思想的影响
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-10-26 DOI: 10.1002/aaai.70108
Gary Marcus
{"title":"Daniel Kahneman's Influence on My Thought","authors":"Gary Marcus","doi":"10.1002/aaai.70108","DOIUrl":"https://doi.org/10.1002/aaai.70108","url":null,"abstract":"&lt;p&gt;It would not be an exaggeration to say that Danny Kahneman has influenced my thinking throughout my adult life and even earlier; I first learned about one strand of his work at a summer class on cognitive psychology when I was fifteen. Just a few years later, I became completely immersed in it. When I got to know him personally, in his later years, it was a huge thrill. We would often have lunch at the Pearl Oyster Bar in Greenwich Village, both of us invariably ordering their amazing lobster rolls. The restaurant is gone and I miss those lobster rolls, but I miss Danny much more.&lt;/p&gt;&lt;p&gt;The last time that I saw Danny was just a couple of months before he died, at a dinner, and he had switched name tags so that we could sit together. I will always cherish that memory. It was also a real thrill to have him participate in the 20-or so person “AI debate” that I co-hosted during the pandemic; he participated reluctantly, saying he wasn't sure he had anything to contribute, and characteristically was lucid and a highlight of the night.&lt;/p&gt;&lt;p&gt;Most of our conversations were about AI, which I had gravitated towards and which he was deeply curious about. But they were built on and grounded in a shared interest in cognitive psychology.&lt;/p&gt;&lt;p&gt;And it is really through cognitive psychology rather than AI that I first became aware of his work, and how his work first began to influence me. I had heard brief mention of his work on human irrationality in that summer class. By the time I decided to go to Hampshire College a few months later, I already knew that I wanted to study cognitive science, drinking as deeply as I could from fields like philosophy, psychology, linguistics, and computer science.&lt;/p&gt;&lt;p&gt;In a summer project at Hampshire, supervised by Neil Stillings, I studied reasoning from a cross-cultural perspective; though most of Danny's work was on Western subjects, a large fraction of what I read was influenced by Danny. A 1982 book that he co-edited with Amos Tversky and Paul Slovic, &lt;i&gt;Judgment Under Uncertainty: Heuristics and Biases&lt;/i&gt; was practically my bible (Tversky &amp; Kahneman, &lt;span&gt;1974&lt;/span&gt;). My undergraduate thesis (known at Hampshire as a “Division III”) was an experimental study about one facet of rationality: how people formed and revised their beliefs relative to evidence — very much a Danny kind of question. (In a nutshell I found that people quickly lost track of how they formed their beliefs, and would sometimes develop rationalizations for those beliefs, retaining those rationalized beliefs even if the original evidence for them was discredited.)&lt;/p&gt;&lt;p&gt;In graduate school, I left reasoning behind and went to work with Steven Pinker, on an entirely different topic: language acquisition. But the things I found there in some ways paralleled the “System I” (fast, reflexive, automatic) and “System II” (deliberative reasoning) distinction that Danny later made famous in &lt;i&gt;Thinking, Fast and Slow&lt;/i&gt; (Kahneman, &lt;span&gt;2011&lt;/span&gt;). M","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70108","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149259145","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Digital disembodiment 数字分离
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-10-26 DOI: 10.1002/aaai.70112
Barbara Tversky
{"title":"Digital disembodiment","authors":"Barbara Tversky","doi":"10.1002/aaai.70112","DOIUrl":"https://doi.org/10.1002/aaai.70112","url":null,"abstract":"<p>As people think, communicate, and create they often put thought into the world in various forms for themselves and others to expand the mind and to work with the thought. People then see more, or less, than what they had intended, reflect, and revise their thoughts, creating a cycle. The thinking is in that loop around the mind, the body, and the world. Each cycle is unique, not easy to locate and not easy to decompose, presenting challenges to capture for Gen AI.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-26","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70112","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149259144","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
A READER'S MAP: Human and machine intelligence, in memory of Daniel Kahneman 读者地图:人类和机器智能,纪念丹尼尔·卡尼曼
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-10-22 DOI: 10.1002/aaai.70119
Luis Lamb, Francesca Rossi
{"title":"A READER'S MAP: Human and machine intelligence, in memory of Daniel Kahneman","authors":"Luis Lamb,&nbsp;Francesca Rossi","doi":"10.1002/aaai.70119","DOIUrl":"https://doi.org/10.1002/aaai.70119","url":null,"abstract":"&lt;p&gt;Daniel Kahneman changed how many fields think about thinking, and AI was among the last of them to feel the full force of his ideas. His distinction between a fast, automatic, and associative System 1 and a slow, effortful, reflective, and deliberative System 2 was offered as an account of human judgment, but it reads, to an AI researcher, almost as an architectural brief. In fact, the early decades of the field built something like System 2: symbolic, rule-based, transparent, and brittle outside its niche. The deep-learning era built something like System 1: fluent, intuitive, powerful within its training distribution, and opaque. Neither, on its own, is intelligence. Human minds are versatile because they hold both and, most importantly, because they know when to shift between them. That coordination problem is the thread that runs through this volume and is key both to understand human intelligence and to shape machine intelligence.&lt;/p&gt;&lt;p&gt;We dedicate the volume to Danny's memory. Several of the contributors, including the two of us, knew him as a friend and interlocutor across decades; others were shaped by his work long before we had the chance to discuss with him in person, and many of the articles gathered here grew directly out of his ideas. What unites them is a shared question, one Danny posed in a short reflection he wrote a couple of years before he died: general intelligence will arrive when machines match the versatility of human minds, and that arrival awaits the integration of intuition and reasoning. This volume takes that integration as its organizing problem.&lt;/p&gt;&lt;p&gt;The articles are arranged in four parts that cover the relationship between human and machine intelligence from different angles and propose different ways to address the current limitations of AI. While each article is self-contained, we feel there is a natural narrative that comes out in the whole volume and we recommend that the interested reader read them in the sequence outlined below.&lt;/p&gt;&lt;p&gt;Two articles diagnose what today's systems lack.&lt;/p&gt;&lt;p&gt;Marcus's “Daniel Kahneman's Influence on My Thought” offers a personal account of Kahneman's influence across his own career, from his early work on the two distinct mechanisms children use to learn the past tense to his long-standing argument for neurosymbolic AI: neural networks as System 1, classical symbolic AI as System 2, and the insistence that capable systems will need both. His diagnosis of what is missing is concrete: without a proper distinction between individuals and kinds, neural systems overgeneralize, which is why hallucination has haunted them from the earliest models to the latest.&lt;/p&gt;&lt;p&gt;Ford, Hancock, Forbus, Hayes, and Venable's “Centaur: Prediction Without Enlightenment” develops the same point with technical details. Examining a model that predicts human behavior from text descriptions of experiments, they argue that predictive success is not explanatory understanding: a system that reproduces inp","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70119","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149258420","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Extended generalization and improved safety with the System-2 features of the Scientist AI 科学家AI的System-2特性扩展了通用性,提高了安全性
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-10-22 DOI: 10.1002/aaai.70117
Yoshua Bengio
{"title":"Extended generalization and improved safety with the System-2 features of the Scientist AI","authors":"Yoshua Bengio","doi":"10.1002/aaai.70117","DOIUrl":"https://doi.org/10.1002/aaai.70117","url":null,"abstract":"<p>Neural networks usually answer queries with a single forward pass—an analog of Kahneman's System 1. In unfamiliar or high-stakes regimes, this can be brittle and overconfident. The Scientist AI framework augments a fast <i>Predictor</i> trained to approximate the Bayesian posterior with System 2 inference: structured probabilistic reasoning over intermediate natural-language statements and causal hypotheses. The units that are generated and related via probabilistic dependencies are not the words or tokens but these Boolean statements about arbitrary properties of the world. Because inference-time compute is spent only on the query at hand, marginalizing over latent statements can reduce computational uncertainty, produce better calibrated conditional probabilities and confidence intervals, and generate explanations as explicit argument graphs. With data transformed into these statements processed to make a syntactic difference between communication acts and factual hypotheses and the Predictor trained to be <i>consequence-invariant</i>—insensitive to downstream effects of its outputs (Fornasiere et al., 2026)—this epistemic caution supports a safety case: the system is not imitating humans, has less incentive to acquire implicit goals, and is better suited as a guardrail for untrusted agents.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-22","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70117","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149258413","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
Bottom-up and top-down problem solving with agentic AI 用人工智能解决自底向上和自顶向下的问题
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-10-21 DOI: 10.1002/aaai.70107
Henry Lieberman, David Canfield Smith
{"title":"Bottom-up and top-down problem solving with agentic AI","authors":"Henry Lieberman,&nbsp;David Canfield Smith","doi":"10.1002/aaai.70107","DOIUrl":"https://doi.org/10.1002/aaai.70107","url":null,"abstract":"<p>Agentic LLM agents can sometimes seem like magic, granting our (almost) every wish. But like the magic genies in stories like the Sorcerer's Apprentice, sometimes you have to be careful what you wish for. The difficulty of precisely specifying what it is we want and unintended consequences, sometimes derail this magic. As with genies, they grant requests in a top-down manner. The agent autonomously breaks the wish into subgoals, until it reaches concrete actions that can be done by Agentic facilities. If left unsupervised by the user, a misunderstanding can result in disaster. But it's also possible to adopt a bottom-up approach, starting with concrete examples, demonstrating procedures on these examples, and only then generalizing to a wider set of cases. It is often easier to control and debug, and easier for beginners and nontechnical people. Ultimately, we need flexibility to go in either direction. The contribution of this paper is to bring a bottom-up methodology to “vibe coding” with Agentic Agents like Cowork and ChatGPT Agent. It draws on the technique of Programming by Example (also called Programming by Demonstration) for conventional programming. The advent of Agentic AI may unblock a 50-year obstacle that has prevented PBE from more widespread acceptance.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70107","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149258370","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
From “how” to learn to “what” to learn in multiagent systems and robotics 在多智能体系统和机器人中,从“如何”学习到“学习什么”
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-10-20 DOI: 10.1002/aaai.70114
Peter Stone
{"title":"From “how” to learn to “what” to learn in multiagent systems and robotics","authors":"Peter Stone","doi":"10.1002/aaai.70114","DOIUrl":"https://doi.org/10.1002/aaai.70114","url":null,"abstract":"<p>There remains much more to Artificial Intelligence than understanding how to learn with neural networks. Inspired in part by how babies, and humans in general, tend to choose for themselves <i>what</i> (and <i>when</i>) to learn in addition to <i>how</i> to learn, this article discusses similar opportunities, in the areas of multiagent systems and robotics. Grounded in past research on the topic, it categorizes opportunities for choosing <i>what</i> to learn as coming from task decomposition or simplification for difficult problems and/or large environments; unsupervised pre-training of representations and priors; multiagent learning; and learning from human interaction. The article is based closely on a keynote talk with the same title, delivered at the AAAI-2026 conference in Singapore on January 22nd, 2026.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-20","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70114","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149258161","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
From few-shot learning to data-efficient intelligence 从几次学习到数据高效的智能
IF 3.5 4区 计算机科学
Ai Magazine Pub Date : 2026-10-19 DOI: 10.1002/aaai.70104
Yaqing Wang
{"title":"From few-shot learning to data-efficient intelligence","authors":"Yaqing Wang","doi":"10.1002/aaai.70104","DOIUrl":"https://doi.org/10.1002/aaai.70104","url":null,"abstract":"<p>Modern artificial intelligence (AI) has achieved remarkable progress by scaling data, model size, and computation. Yet scaling alone does not resolve one of the most fundamental gaps between machine and human intelligence: the ability to learn, adapt, and act from limited experience. This limitation is especially visible when labeled examples, expert feedback, user interactions, or real-world trials are scarce, costly, or dynamically changing. Humans can often grasp a new concept, adapt to a new goal, or personalize behavior after only a few examples or interactions, but many AI systems remain fragile under the same conditions. My research studies data-efficient generalization: how learning systems can acquire useful priors, adapt to new tasks, and reuse limited experience under bounded supervision. In this article, I trace a continuous research path from few-shot learning to meta-learning, in-context learning, and data-efficient agentic learning. I first revisit few-shot learning as a foundational problem and explain how data-, model-, and algorithm-level priors make learning possible when empirical risk minimization is unreliable. I then describe our work on meta-learning, where task-level information and parameter-efficient adaptation enable fast generalization. Next, I discuss in-context learning in large language models (LLMs) as an implicit form of learning to learn, and show how its connection to classical meta-learning provides both theoretical insight and practical guidance. Finally, I turn to large language model (LLM)-based agents, where data efficiency expands beyond labels to include demonstrations, reasoning traces, tool-use records, preferences, memory, and feedback. Across AI for science, biomedical knowledge discovery, recommendation, search, and personalization, the common thread is simple: more capable AI should extract more learning signal from less experience.</p>","PeriodicalId":7854,"journal":{"name":"Ai Magazine","volume":"47 4","pages":""},"PeriodicalIF":3.5,"publicationDate":"2026-10-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://onlinelibrary.wiley.com/doi/epdf/10.1002/aaai.70104","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149258046","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
引用次数: 0
0
×
引用
GB/T 7714-2015
复制
MLA
复制
APA
复制
导出至
BibTeX EndNote RefMan NoteFirst NoteExpress
×
提示
您的信息不完整,为了账户安全,请先补充。
现在去补充
×
提示
您因"违规操作"
具体请查看互助需知
我知道了
×
提示
确定
请完成安全验证×
相关产品
×
本文献相关产品
联系我们:info@booksci.cn Book学术提供免费学术资源搜索服务,方便国内外学者检索中英文文献。致力于提供最便捷和优质的服务体验。 Copyright © 2023 布克学术 All rights reserved.
京ICP备2023020795号-1
ghs 京公网安备 11010802042870号
Book学术文献互助
Book学术文献互助群
群 号:604180095
Book学术官方微信
小红书