The success of predictive large language models (PLLMs) like GPT3 and ChatGPT has created both enthusiasts and skeptics of their widespread practical applications, but this book argues that the larger significance of such models is contained in what they suggest about human cognition. To explore this potential, the book develops a thought experiment called the Prediction Room, a reference to John Searle's influential Chinese Room argument, in which a human agent processes language by following a set of opaque written rules without possessing an inherent understanding of the language. The book proposes a new Room modelthe Prediction Room with its resident Prediction Agentgeneralizing the working of large language models. Working through a wide range of topics in cognitive science, the book challenges the conclusion of Searle's thought experiment, that discredited contemporary artificial intelligences (AI), through the suggestion that the Prediction Room offers a means of exploring how new ideas in AI can provide productive alternatives to traditional understandings of human cognition. In considering the implications of this, the book reviews an array of topics and issues in cognitive science to uncover new ideas and reinforce older ideas about the mental mechanisms involved in both sides. The discussion of these topics in the book serves two purposes. First, it aims to stimulate new thinking about familiar topics like language acquisition or the nature and acquisition of concepts. Second, by contrasting human psychology with the form of artificial psychology these models exhibit, it uncovers how new directions in the development of these systems can be better explored.

Evaluates PLLMs using cognitive science to reflect on implications for human cognition Explores how PLLMs reinforce and suggest new ideas about mental mechanisms of cognition Utilizes cognitive model theories to test limits of understanding large language models

Autorentext

Clayton Lewis is Ermeritus Professor of Computer Science and Fellow of the Institute of Cognitive Science at the University of Colorado Boulder, and Fellow of the Hanse-Wissenschaftskolleg, Delmenhorst, Germany. He earned an A.B. in mathematics from Princeton University, an interdisciplinary M.S. in mathematics and linguistics from MIT, and a Ph.D. in experimental psychology from the University of Michigan. His work has contributed to user interface evaluatiion, programming language design, cognitive assistive technology, educational technology, and cognitive theory in causal attribution and learning. He has been honored by appointment to the ACM SIGCHI Academy, by the SIGCHI Social Impact Award, and by the ACM SIGACCESS Outstanding Contribution Award.

Inhalt

Introduction.- The Prediction Room: A Rough Outline of a Model of Cognition.- Part I: Problem Solving and Reasoning.- Einstellung.- Analogical Reasoning.- More Reasoning: Inference and Prediction.- Transfer Skills, Production Rules, and Prediction.- Qualitative Physics.- Situated Cognition.- Part II: Memory.- Recall from Long Term Memory.- Interference with Real World Knowledge.- Speed-Accuracy Tradeoffs.- Is Knowledge Represented in Propositions?.- Associationism.- Procedural and Declarative Knowledge.- Part III: Language.- Language.- Inner Speech.- Babies and Other Primates.- Gestures.- Part IV: Action.- Action and Identity.- Predictive Modeling and Active Inference.- Part V: Being Human and Being Artificial.- Embodiment and Grounding.- Concepts?.- Emotions.- Intuition.- Belief-Desire Psychology.- Part VI: Mechanisms and Interpretation.- In the Engine Room: Transformer Models.- In the Engine Room--The Brain.- Virtuality, Reading in, and Emergence.- Coda.