AsianFin -- Despite generative AI showing significant improvements in efficiency—such as answering complex questions and generating high-quality content—the technology is still in an experimental phase, and reliability remains the primary obstacle to its further development, says Mark Nitzberg, Executive Director of the Artificial Intelligence Laboratory at the University of California, Berkeley, and co-founder of Dark Matter AI.
He made the remarks at the 2024 T-EDGE Conference on December 7.
"While current models have made substantial progress in performance, we cannot ignore their inconsistency in critical scenarios. For example, a slight change in input can lead to a significant deviation in the model's output, and this uncertainty is unacceptable in high-risk fields like healthcare and transportation."
At the same time, Nitzberg highlighted the exponential efficiency gains generative AI has shown in various emerging sectors. He shared insights from a Swedish study that analyzed multiple industries, noting that the "building restoration" field saw a 100-fold increase in efficiency due to generative AI applications compared to traditional methods.
However, Nitzberg maintained a balanced perspective. "While these efficiency gains are exciting, their sustainability still depends on whether we can overcome the issue of technological reliability," he cautioned.
He also emphasized the importance of ensuring that we fully understand the pre-AI operational processes before integrating generative AI into every field. "Only then can we ensure that when AI is introduced, it operates more efficiently and quickly without automating the 'problem areas' from the previous manual processes," Nitzberg added.
Furthermore, Nitzberg believes that the potential for generative AI to evolve into intelligent agents is vast, but this transition comes with complex technical challenges and safety risks. He identified reliability and control mechanisms as critical research directions for future development. Regarding future research priorities, Nitzberg pointed out two key areas of focus: integrating multimodal data and enhancing causal reasoning.
