lectures.yml

- title: "Introduction au RLHF"
  subtitle: "Comment chatGPT est éduqué"
  date: "01/09/24"
  paper_link: "https://arxiv.org/abs/1706.03741"
  paper_title: "Deep reinforcement learning from human preferences"
  past: true

- title: "Inference Time intervention using semantic vectors"
  subtitle: "Modifier les comportements des modèles de manière chiurgicale"
  date: "08/09/24"
  paper_link: "https://www.lesswrong.com/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction"
  paper_title: "Refusal in LLMs is mediated by a single direction"
  blog_link: "inference-time-intervention"
  past: true

- title: "Limits of Machine Unlearning"
  subtitle: "Pourquoi on ne peut jamais vraiment faire oublier des IA"
  date: "15/09/24"
  paper_link: "https://arxiv.org/abs/2406.17216"
  paper_title: "Machine Unlearning Fails to Remove Data Poisoning Attacks"
  past: true

- title: "Watermarking LLMs"
  subtitle: "Comment introduire une signature invisible dans les textes générés par IA"
  date: "22/09/24"
  paper_link: "https://arxiv.org/pdf/2301.10226"
  paper_title: "A Watermark for Large Language Models"
  blog_link: "watermarking-llm"
  past: true
  
- title: "Fundamentals limits of alignment in LLMs"
  subtitle: "Un résultat fondamental sur l'imperfection des IA"
  date: "6/10/24"
  paper_link: "https://arxiv.org/abs/2304.11082"
  paper_title: "Fundamentals limits of alignment in LLMs"
  past: true

- title: "Privacy backdoors & MIA"
  subtitle: "Récupérer des données confidentielles qui ont servi à entrainer le modèle"
  date: "13/10/24"
  paper_link: "https://arxiv.org/abs/2404.01231"
  paper_title: "Privacy Backdoors: Enhancing Membership Inference through Poisoning Pre-trained Models"
  past: true

- title: "Représentants algorithmiques"
  subtitle: "L'opportunité de donner nos droits de vote à des algorithmes"
  date: "20/10/24"
  paper_link: "https://dl.acm.org/doi/10.1145/3359283"
  paper_title: "WeBuildAI: Participatory Framework for Algorithmic Governance"
  past: true

- title: "Introspection in LLMs"
  subtitle: "En un sens, on peut finetune les LLMs actuels pour qu'ils soient capables d'introspection"
  date: "11/11/24"
  paper_link: "https://arxiv.org/pdf/2410.13787"
  paper_title: "Looking Inward: Language Models Can Learn About Themselves by Introspection"
  past: true

- title: "Data Ownership Verification with Data Poisoning"
  subtitle: "Empoisonner pour signer des données, plus puissant que le watermarking ?"
  paper_title: "DATA TAGGANTS: DATASET OWNERSHIP VERIFICATION VIA HARMLESS TARGETED DATA POISONING"
  paper_link: "https://arxiv.org/pdf/2410.09101"
  date: "24/11/24"
  past: true
  
- title: "Un framework pour penser la confiance en les IA"
  subtitle: "Comment formaliser l'idée de confiance en une IA pour pouvoir l'implémenter ?"
  paper_title: "Formalizing Trust in Artificial Intelligence: Prerequisites, Causes and Goals of Human Trust in AI"
  paper_link: "https://dl.acm.org/doi/pdf/10.1145/3442188.3445923"
  date: "01/12/24"
  past: true