The Queensland Government Customer and Digital Group (CDG) supports whole-of-government IT and Data initiatives that offer better services to the people of Queensland. This is done through automation, technology innovation and making available relevant and up-to-date information to the community.

CDG works closely with other Queensland Government agencies to build their digital capabilities and experiment with new models and their adoption in the workplace. A key development in the past year is the advent of Generative Artificial Intelligence and, in particular, Large Language Models (LLMs) as mainstream technologies that can be deployed in the corporate environment. CDG is developing the appropriate policies, resources, technology infrastructure and support services for Queensland Government agencies to effectively and responsibly adopt these new technologies.

An important challenge with LLMs is that they require a high level of technical skills, as well as advanced research skills to keep up with continuous advances in the area of Artificial Intelligence and its management. APR.Intern offered CDG the opportunity to access a large pool of graduate research students with the requisite skills and the interest in applying their domain knowledge to practical applications.

it is important to ensure that content generated by the software is aligned with the ethical principals and the values of the Queensland Government. To do so manually can be a challenge due to anticipated large volume of user interactions with LLMs.

“We were interested in investigating the possibility of using the LLMs themselves to ensure the alignment of their outputs with Queensland Government values,” said Marco Fahmi, Associate Director Data Strategy at CDG.

“However, this is not a readily available feature nor a product that is currently available commercially. It was therefore necessary to put together a research project to build a prototype and evaluate its effectiveness aligning outputs with values. This required the technical and scholarly skills of academic researchers.”

Yue (Joy) Wang (QUT) was drawn to the project at CDG due to a strong skills alignment and interest in LLMs, and the opportunity to enhance her practical experience as she approached her graduation.

“During my internship, I worked on a project focused on improving GenAI safety. The goal was to develop an automated system that ensures AI-generated content is in line with the values of organisations in the public sector. My role encompassed several key tasks: conducting an extensive literature review to understand the current state of GenAI safety, developing a comprehensive project plan, designing an initial approach for the automated solution, and carrying out experiments to validate our hypotheses.”

– Joy Wang, Former APR Intern at CDG.

Image: Joy Wang, APR Intern (supplied).

Joy’s top learning outcomes from the internship include a comprehensive exploration of the AI safety research landscape and gaining proficiency with the OpenAI API.

“This experience has significantly broadened my understanding of the challenges and considerations in ensuring AI safety and has equipped me with practical skills in utilising advanced AI tools for research and development.”

“I believe I made significant contributions to the project by leveraging my existing knowledge and skills from prior research. My most substantial achievement was designing and testing a proof-of-concept solution that met stakeholder requirements and providing preliminary experimental results. These results not only validated our initial hypothesis but also outlined potential future research directions for the project.”

APR.Intern was useful to CDG because it gave the organisation access to technical and research skillsets from a broad range of disciplines and applications, and an enthusiastic cohort of students who would not have been available otherwise.

“The internship produced an architectural design called “AI Safety-as-a-Service” that has been validated and can be deployed in Queensland Government’s own LLM service. This is a valuable contribution that not only serves users in Queensland Government and has also proved useful as a general architecture that can be applied in other LLM domains. Indeed, one of the interns has reused this architecture to develop a novel LLM evaluation method for his dissertation on smart transportation systems.”

– Marco Fahmi, Associate Director Data Strategy at CDG.

Envisioning a future in a rapidly evolving technological landscape is challenging, yet Joy feels confident that this internship has equipped her with the resilience and adaptability needed to thrive in a rapidly evolving digital landscape.

“The experience has honed my skills in quickly acclimating to new projects and domains, which I believe will be pivotal in navigating the uncertainties of the coming years. APR.Intern will undoubtedly support my career progression and help me achieve my long-term goals.”

Joy Wang was a recipient of the WiSE subsidy.