Abstract

Just a few years after the first truly massive large language models were created, it’s already clear that the impact of generative AI and LLMs will be significant in almost every aspect of society.

Further development of generative AI is ongoing, and currently, there seems to be no end in sight. Some limitations and issues with generative AI applications have been fixed, some have been mitigated to be less problematic, and some challenges remain tough to completely overcome. Issues such as “hallucinating AIs” and security-related topics are still active areas of development. Generative AI today also goes beyond large models and frameworks. Solutions have been created that allow generative AI applications to fetch facts and data, program and execute code, browse the web, and use more advanced reasoning.

As generative AI solutions evolve, so does our use of them. Many people use generative AI solutions as brainstorming partners, quick tools to organize thoughts or as proofreaders. As generative AI solutions become more advanced, we also grant AI systems a higher degree of autonomy. In some cases, we don’t just want advice from LLMs that we can turn into action ourselves—we want the AI to figure out what needs to be done and perform the task itself.

Like many other types of organizations, statistical organizations have a wide spectrum of activities and decisions, ranging from highly important tasks where accuracy and correctness are key, to trivial activities where more pragmatic solutions can be applied. Some activities and decisions require extra verification procedures, and some can be completely automated by using AI or other automation technologies.

This presentation highlights how some of the advancements in generative AI could be applied in our statistical context and describes how we can leverage the capacity of LLMs to go from providing advice to action.