LLMs are part of three major scientific breakthroughs in just 10 years of Deep Learning applied to NLP: word embeddings, transformers with attention and prompt tuning. LLMs have shown surprising effectiveness in all tasks and have been adopted even in cross-cutting creative tasks such as the generation of images, code or music from textual descriptions. They seem to exhibit abilities which go beyond the tasks on which they were trained. Emergent abilities arise with scaling and are a phenomenon still subject to investigation. LLMs also raised concerns, since the release of GPT-2, the withdrawal of Galactica or the unexpected success of ChatGPT. We will discuss their benefits, their limitations and how they will evolve. In particular, if the resources to build them will continue to grow, this might lead to a discrimination between groups of researchers and solutions to avoid this must be considered.
Giuseppe Attardi is a retired professor of Computer Science at the University of Pisa. He also worked at MIT’s AI Lab, Sony Paris Research Laboratory, ICSI in Berkeley and Yahoo Research Barcelona. He developed Omega, a description logic; CMM, the garbage collector used in Java and DeSR, a neural dependency parser. He participated in the development of Arianna, the first Italian search engine and introduced the technique of categorisation by context for web pages. He is the founder or partner of a few startups, in Italy and Spain. He contributed to building the fiber optic networks of the University of Pisa and of GARR. He led a national campaign to facilitate and disseminate Internet access. He led the development of the GARR cloud platform. He contributed to the drafting of the Italian strategy on Artificial Intelligence and the birth of the first national PhD in Artificial Intelligence.
ZOOM & Password: 148148
PDF & VIDEO