The Center for Artificial Intelligence (C4AI) at the University of São Paulo hosted a Perspectives in AI seminar featuring Professor Claudio Pinhanez, an independent AI scientist and innovator. The talk, titled "The Non-Determinism of LLMs and Other Stability Issues of GenAI," was streamed live on April 29th and has garnered nearly 1,000 views.
Pinhanez discussed two critical factors that large language models (LLMs) lack for stability: local boundedness and determinism. He theoretically examined the issue of boundedness and its connection to prompt engineering, arguing that it makes achieving safety in AI systems exceedingly difficult.
Empirically, Pinhanez presented studies on small language models (2B–8B parameters) where the same questions were posed identically and with minor variations multiple times. The results revealed significant concerns about accuracy and reliability, highlighting the inherent instability of current generative AI technologies.
The seminar is part of C4AI's ongoing series exploring cutting-edge topics in artificial intelligence.