🎉 Kicking off the new year with a big topic – “Europe’s Path in Artificial Intelligence”. I had the privilege of joining a stellar panel organised by Fraunhofer, with an online audience of ~500. I shared Australia’s AI regulation status, including the pivotal role of the Voluntary AI Safety Standards in accelerating AI adoption with trust and helping explore regulatory approaches.

🚀 I conveyed the following messages:
1️⃣ 𝐒𝐚𝐟𝐞𝐭𝐲 𝐯𝐬. 𝐅𝐮𝐧𝐜𝐭𝐢𝐨𝐧, 𝐧𝐞𝐚𝐫 𝐯𝐬. 𝐥𝐨𝐧𝐠-𝐭𝐞𝐫𝐦 𝐫𝐢𝐬𝐤 – 𝐀 𝐟𝐚𝐥𝐬𝐞 𝐝𝐢𝐜𝐡𝐨𝐭𝐨𝐦𝐲?
High-performing AI use cases and safety aren’t hard choices. The underlying science is the same – deeply understanding AI models/systems and steering them effectively. The same understanding ensures functional accuracy and reliability, and also control over risks—whether it’s bias or deepfake concerns today or out-of-control risks tomorrow.

2️⃣ 𝐖𝐡𝐲 𝐬𝐨 𝐦𝐚𝐧𝐲 𝐫𝐢𝐬𝐤 𝐚𝐬𝐬𝐞𝐬𝐬𝐦𝐞𝐧𝐭𝐬, 𝐦𝐞𝐚𝐬𝐮𝐫𝐞𝐬, 𝐚𝐧𝐝 𝐦𝐢𝐭𝐢𝐠𝐚𝐭𝐢𝐨𝐧𝐬?
Because we lack good science understanding, we have to spread many mitigations across the lifecycle—process and product. And we have to layer things up and allow different ways because we are not sure exactly how much risk eduction one measure/mitigation achieves . Only by advancing the science of concrete and quantified risk assessment/mitigation can we truly reduce the confusion, the cost, and the interpretation variations plagued by high-level regulations, standards, and frameworks that try to do the right thing by piling up many mitigations just in case.

3️⃣ 𝐁𝐞𝐲𝐨𝐧𝐝 𝐦𝐨𝐝𝐞𝐥𝐬 – 𝐒𝐲𝐬𝐭𝐞𝐦𝐬-𝐥𝐞𝐯𝐞𝐥 𝐢𝐧𝐧𝐨𝐯𝐚𝐭𝐢𝐨𝐧:
AI models don’t work by themselves. The real leap comes from pairing them with external tools, smarter and more inference-time compute, and external knowledge bases. This system-level innovation is within reach for everyone; there’s no need to chase endless scaling of GPUs or data. And if all models are learning the same underlying world model/understanding, they will converge on the same thing, so model-level competitive advantage may disappear. System-level innovation is where you truly differentiate, so start now.

🛠️ That’s why CSIRO’s Data61 is focusing on the science of measuring and controlling AI “systems”—from concrete risk assessments/mitigation to inference-time/system-level innovation for industry and government. A bottom-up, foundational science approach is key to cutting through the noise of vague mitigations and ensuring meaningful, efficient risk control alongside accelerated innovation. More here: https://lnkd.in/gPhid9tX

We are also developing version two of the AI Safety Standard right now in tandem with AI safety institutes around the world on model/system evaluation (assuring the system) and the ISO, NIST and EU AI Act’s Code of Practices for GPAI developers (assuring the processes). Stay tuned.