It’s my great honour and pleasure to deliver the opening keynote at the Australasian Joint Conference on Artificial Intelligence 2024. https://ajcai2024.org/ 🎙️

In this keynote, I reflected on the journey and challenges of trustworthy AI: from models to systems to agents. 🚀 Although you might be familiar with much of the work CSIRO’s Data61 do in this area, I positioned it within an increasingly challenging context:

– What happens when human answers are no longer the ground truth, especially when considering the cost of collecting them at scale? 🤔
– Why do human-AI collaborations sometimes perform worse than either AI or humans alone? 🤷
– Why does AI sometimes improve junior professionals more, while at other times it provides more help to experts? 🧠
– What if inference-time scaling laws and tool access, rather than training-time scaling laws, are what really matter? What strategic shifts does that imply? 🔄

I introduced some of our latest work on runtime guardrails, AgentOps, and evaluation-driven out-of-model learning to help tackle these challenges, in addition to our systems-engineering approach to AI trustworthiness. 🛡️ https://lnkd.in/gPhid9tX

Thanks to the organisers for inviting me. It’s always particularly exciting to deliver a keynote to the most intelligent AI experts in the nation and region.

I will also be joining an industry panel tomorrow to discuss more about our work in Australia’s AI Safety Standard and to report back on my trip to the US last week, covering all things related to AI safety internationally.

Slides: dropbox link