Critical issues with LLMsEnjeux critiques des GML

GIGO, hallucinations, bias, environment, hidden work, power and access, jobs, skills and learning, privacy, copyright, the information ecosystem, and knowledge and trust.GIGO, fabulations de l'IA, biais, environnement, travail invisible, pouvoir et accès, emploi, compétences et apprentissage, vie privée, droit d'auteur, écosystème informationnel, savoir et confiance.

What to expectÀ quoi s'attendre: This module is about what goes wrong with large language models (LLMs), in a single answer and beyond it. It explains where fabricated answers and biases come from and what models cost in energy and emissions. It also looks at the people and the data the models depend on, at who owns what goes in and comes out, and at what the models change for jobs, for learning and for the information everyone relies on.This module is about what goes wrong with large language models (LLMs), in a single answer and beyond it. It explains where fabricated answers and biases come from and what models cost in energy and emissions. It also looks at the people and the data the models depend on, at who owns what goes in and comes out, and at what the models change for jobs, for learning and for the information everyone relies on.
PrerequisitesPréalables: The LLM lifecycle module, in particular its chapter on training data: many of the problems described here start there. The modules on capabilities and on using LLMs are related and can be read in any order.The LLM lifecycle module, in particular its chapter on training data: many of the problems described here start there. The modules on capabilities and on using LLMs are related and can be read in any order.
StructureStructure: The first chapter is the principle the rest of the module keeps returning to: the quality of what goes in bounds the quality of what comes out, at four levels, from the training corpus down to the prompt you type. Hallucinations come second, and they are the exception: fabrications turn up whether the input was good or bad. Bias and environmental impact follow, the two problems that show in what a model says and in what it costs. Four chapters are about people: the hidden work behind the models, the power concentrated in a few companies and the unequal access to their tools, what the models change for jobs, and what they change for skills and learning. Two chapters are about text going into a model and coming out of it: privacy, and copyright and authorship. The last two look at the information everyone relies on: what generated content does to the web and to peer review, and what it does to what people know and can trust.The first chapter is the principle the rest of the module keeps returning to: the quality of what goes in bounds the quality of what comes out, at four levels, from the training corpus down to the prompt you type. Hallucinations come second, and they are the exception: fabrications turn up whether the input was good or bad. Bias and environmental impact follow, the two problems that show in what a model says and in what it costs. Four chapters are about people: the hidden work behind the models, the power concentrated in a few companies and the unequal access to their tools, what the models change for jobs, and what they change for skills and learning. Two chapters are about text going into a model and coming out of it: privacy, and copyright and authorship. The last two look at the information everyone relies on: what generated content does to the web and to peer review, and what it does to what people know and can trust.