Last Wednesday, November 20th, as part of the activities of the Trustworthy AI Chair, we offered two hours covering the main themes related to LLMs, at ENSEIRB-MATMECA. You can find here the post announcing the program and the day’s announcement (you will find the details of the speakers and the day’s program).
It was a real success (almost full amphitheater), with a varied audience that was attentive and interested. The originality of the day was to cover three important aspects of LLMs to present, in just two hours, the fundamentals, uses, and dangers.
We had Nathanaël Fijalkow who presented the technical fundamentals of LLMs, that is, the underlying transformer-based architecture, through examples from his re-implementation of a small LLM.
Then, Erwann Simon focused on the next phase, when the LLM is a one-shot learning and has learned to connect language elements. It then needs to be specialized in a particular task, and this phase, called fine-tuning, is truly fundamental for using LLMs in a production environment.
Finally, it was a pleasure to hear Sonia Tatbi on the limitations, biases, and dangers of these tools. Her presentation highlighted the necessity for users to be trained on the limitations of these seemingly effective tools.
Thank you again to everyone for making these two hours a fascinating and comprehensive moment!
(The post announcing the program and the day’s announcement).