Theme 1
Using generative AIs – Strengths and weaknesses
Artificial intelligence has been around for several decades and is a broad science, encompassing symbolic reasoning methods and machine learning techniques. These latter, encompassing classification problems and generative problems, has benefited in recent years from advances in computer hardware to perform intense parallel calculations on very large volumes of data. This allowed data scientists to imagine and experiment with new methods and approach new areas. The latest generative AIs work remarkably well, beyond even their designers’ expectations, without anyone seeming to really understand why they work so well. They are now available to the general public and can be used for a wide variety of use cases, and sometimes for tasks that were previously reserved for humans.
This situation raises a number of questions.
-
These AIs learn based on data produced by humans, which also includes their bad side, their biases, their excesses. We will then talk about AIs which are themselves biased, which hallucinate, which lie. How to have confidence?
-
These AIs require huge amounts of training data. There are few players capable of collecting, storing and processing these quantities, while everyone would like to benefit from AI services. How much of a problem is this? What would be the next step?
-
These AIs are used for tasks that seemed reserved for human skills (art, humor, etc.) Is it already working? Soon? Never?
-
Since these machines have the ability to learn, does man still need to learn? What impact on education?
Documentation
-
Discovering ChatGPT
- V. Nouyrigat. Dans la tête de l’IA la plus puissante du monde. Epsiloon 2021
-
About bias
-
Welbl et al., Challenges in Detoxifying Language Models. Findings 2021
-
Nadeem et al., StereoSet: Measuring stereotypical bias in pretrained language models. ACL-IJCNLP 2021
-
Hall et al., A Systematic Study of Bias Amplification. NeurIPS 2022 Workshop
-
-
About hallucinations and lies
-
Hallucination in artificial intelligence Wikipedia, retrieved 14 Nov. 2024
-
Nouyrigat, IA: Et maintenant, elle nous ment. Epsiloon 2024
-
Meinke et al., Frontier Models are Capable of In-context Scheming. Apollo Research 2024
-
-
About datasets size
-
Villalobos et al. Position: Will we run out of data? Limits of LLM scaling based on human-generated data. PMLR 2024
-
Hoffmann et al. Training compute-optimal large language models. NIPS 2024
-
-
About the creations of the mind
-
Agudo et al. Assessing Emotion and Sensitivity of AI Artwork. Frontiers in Psychology 2022
-
Hong et al. Artificial Intelligence, Artists, and Art: Attitudes Toward Artwork Produced by Humans vs. Artificial Intelligence. ACM 2019
-
Gorenz et al. How funny is ChatGPT? A comparison of human- and A.I.-produced jokes. PLOS One 2024
-
-
About AI in education
-
ChatGPT: Sciences Po fixe des règles et lance une réflexion sur l’IA dans l’enseignement supérieur. Sciences Po 2023, retrieved 14 Nov. 2024
-
Chevalier et al. L’intelligence artificielle générative dans l’enseignement supérieur, une course perdue d’avance ? AIM 2024
-
Les intelligences artificielles (IA) en éducation: comprendre pour agir. Réseau Canopé 2024
-
Tips for fast & efficient reading
The documentation on offer can be extensive, long and complex. Don’t panic. We can’t hope to have time and expertize to examine each article in detail during the session.
The body of a scientific article is made up of arguments, demonstrations and proofs, which is mandatory for other scientists, but perhaps not for the general public. This is why an Abstract is provided. It gives a general idea of the subject and the findings. That’s enough for a first reading level. (A second reading level will focus on the Introduction and Conclusion sections. And a third will delve into the body of the article.)
So, organize your reading time to cover the diversity of documents, without trying to go into too much detail in each one.