دانلود کتاب روانشناسی مصنوعی – یادگیری از قابلیت های غیرمنتظره مدل های زبان بزرگ

This book is about an old idea—that prediction is a key process in the mind or brain— looked at for a new reason. Programs called Large Language Models, or LLMs, and in particular predictive Large Language Models, have shown surprising capabilities, when compared to past forms of artificial intelligence. The models are called predictive, because they are trained to perform a very simple task: given a span of text, predict what text will come next. We’ll be discussing what this training enables the systems to do, and how they do it. Because these are real, implemented systems, they give us an expanded, concrete picture of what prediction can do. This book explores this concrete picture with the question in mind, what can the capabilities of Large Language Models suggest to us about how our human minds work? Along the way, we’ll discuss what’s known about how LLMs do what they do. But in our discussion of LLMs we’ll be focusing on the relationship between these systems and human psychology, not on what LLMs may or may not do for us, as they are applied in various settings. This neglect of applications covers their potential harms as well as their possible benefits. That’s not because these things aren’t important—they certainly are—but because what we want to obtain from our study is insight into human psychology. We’ll discuss some extensions to current systems to bring them more in line with human capabilities. For example, we’ll imagine that the extended system, that we’ll call the Prediction Room, can process not only text but also images and sounds. We’ll also imagine that it can take actions, beyond just emitting text. But, again, our motive isn’t finding ways to make LLMs “better”, or more powerful, but rather finding more insights, as what we can see in existing predictive models is extrapolated to more capable models. I’ve been drawn to this exploration by being struck by how much more of human capabilities have been approximated by these models, than by any previous efforts in artificial intelligence. For me that happened against a background of years of research on human cognition, some of which comes into this book. I’ve also done research in artificial intelligence, in former styles, including a long-ago spell of work on the subfield called “natural language processing”, the effort to enable computer systems to work with human language. Our project aimed at a system that could answer questions about a collection of documents, posed in ordinary English, rather than in an artificial query language. At the end of that project I took a vow never to work again on natural language processing. That’s because the ideas of the time (the late 1960s) seemed so plainly unequal to the challenge. The ideas of today are no longer unequal to it. Not only are these new systems much more capable than earlier ones, but also they work in a way that’s quite different from how many psychologists, including your author, have imagined. We’ve tended to think of the mind as populated with highly structured information, like rules of logic, or scripts that guide behavior, or lists of objects and associated attributes. A mental lexicon, for example, is imagined to be a list of words, with attributes like parts of speech associated with them. LLMs act as if they have these kinds of information—for example, they observe the rules of grammar in the text they produce. But it appears plausible that they actually do not have such information, in the way we have imagined would be needed. Certainly the people who have designed and implemented LLMs have not been thinking about these structures at all. As I worked to understand how it is that LLMs can do what they do, a further idea emerged, that added to my motivation. Taking LLMs seriously as models of human cognition suggested an answer to an otherwise baffling question. How could it happen that millions of well-educated, capable Americans came to subscribe to the bizarre beliefs called QAnon? We’ll explore that as part of our inquiry in the book. What audience is the book aimed at? I have hoped to reach readers who are curious about how the mind works—that’s the “psychology” in the title—and also readers who are curious about artificial intelligence, signalled by “artificial” in the title. I’ve not assumed that you, the reader, have specialist knowledge of either of these broad topics. Rather, I’ve aimed to make the discussion of the ideas self-contained. Of course, you may actually be a specialist in one of these topics, psychology or AI, or even both. If this is so I hope you’ll find that the book contributes some ideas that are new to you, while not dealing too barbarously with the material you know already. A few cautions, before we get started. Because, as we’ll see, LLMs can seem quite human-like, it’s just about impossible to talk about them without using terms like “knows”, or “thinks”, or “believes”. This terminology can be very misleading, however. Does an LLM really “know” anything? Does it “know” things, in a way that is similar to the way people “know” them? I’ll be arguing that the answer to those questions is actually “yes”, but maybe not for the expected reason. I think people “know” things in a way similar to the way LLMs “know” things, and not (as I’ve said) in the way we’ve tended to think people “know” things. But using “know” for an LLM can seem to take all that for granted. Our purpose isn’t to do that, but to think critically about it. One approach would be to use scare quotes everywhere for words like “know”, as I’ve been doing here. But that gets tedious. If you like, feel free to put in imaginary scare quotes as you read. More cautions, or perhaps apologies: I apologize that what you’re about to read isn’t very scholarly, and many of the arguments aren’t well supported. On the first point, the territory we’ll be surveying is the subject of an enormous literature, even the new parts. Much of it, especially the new parts, is quite technical. I wish I could read and understand it all, to share it with you, but I can’t. Many of the arguments, that are presented just as verbiage here, really should be based on mathematical or computational modeling. I’ve gone on record in noting the hazards of not doing that. Evolutionary biologists have a phrase for the kind of speculative argument you’ll see represented here: just so stories, about how something or other evolved. They’re just too easy to make up. I’ve done a little bit of modeling in what follows, and I’ll be using a lot of examples from implemented LLMs. But there’s a lot of speculation in the discussion as well. You’ll also likely feel that some of what you’re about to read suffers from confirmation bias, the tendency to look for evidence for some idea, and neglect evidence against it. Please chalk that up to enthusiasm: I think the ideas we’re discussing are important, and interesting, and even fun, and I’m sure I’ve let them pull me along a bit.