The International Science Council has released a White Paper titled ‘Preparing National Research Ecosystems for AI,’ which examines how different nations are approaching AI opportunities through their research ecosystems. I contributed to the Australian Case Study, including the pivotal role of CSIRO in AI for Science, responsible and safe AI, and general research-driven AI innovation.

But where are we heading from here? Here are my personal takes:

1. 𝐌𝐨𝐯𝐢𝐧𝐠 𝐁𝐞𝐲𝐨𝐧𝐝 𝐓𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐃𝐚𝐭𝐚/𝐂𝐨𝐦𝐩𝐮𝐭𝐞 𝐭𝐨 𝐒𝐲𝐬𝐭𝐞𝐦-𝐋𝐞𝐯𝐞𝐥 𝐀𝐈 𝐈𝐦𝐩𝐫𝐨𝐯𝐞𝐦𝐞𝐧𝐭𝐬
Many nations’ current discourse often fixates on model training data and compute. However, evidence suggests that pre-trained base models (no matter how trained independently) are increasingly converging in their capabilities. Given that these models learn from a large amount of different but overlapping data sources, they inevitably identify the same underlying real-world governing laws, both physical and social. This suggests that the future of AI advancement lies less in brute-force training and more in system-level innovations that enhance AI’s applicability and performance.

𝟐. 𝐈𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞-𝐓𝐢𝐦𝐞 𝐎𝐩𝐭𝐢𝐦𝐢𝐳𝐚𝐭𝐢𝐨𝐧 𝐚𝐧𝐝 𝐒𝐲𝐬𝐭𝐞𝐦 𝐒𝐚𝐟𝐞𝐠𝐮𝐚𝐫𝐝𝐬
Recent studies have shown that inference-time improvements can significantly enhance AI performance without requiring the same level of training compute. Implementing smart inference techniques and robust safeguards can drastically elevate overall system capability and reduce risks. This has major implications for AI deployment in research and industry, as it enables efficiency gains without the escalating costs of massive training/finetuning efforts.

𝟑. 𝐒𝐜𝐚𝐥𝐢𝐧𝐠 𝐀𝐈 𝐰𝐢𝐭𝐡 𝐒𝐞𝐥𝐟-𝐒𝐮𝐩𝐞𝐫𝐯𝐢𝐬𝐞𝐝 𝐚𝐧𝐝 𝐄𝐱𝐭𝐞𝐫𝐧𝐚𝐥 𝐊𝐧𝐨𝐰𝐥𝐞𝐝𝐠𝐞 𝐀𝐩𝐩𝐫𝐨𝐚𝐜𝐡𝐞𝐬
As demonstrated by models like R1-𝒁𝒆𝒓𝒐, with a longer lineage back to AlphaGo 𝒁𝒆𝒓𝒐 and Alpha𝒁𝒆𝒓𝒐, human-labeled data and reasoning annotations and feedback (which is a major bottleneck cost-wise and, 𝑐𝑜𝑛𝑡𝑟𝑜𝑣𝑒𝑟𝑠𝑖𝑎𝑙𝑙𝑦, also a performance bottleneck) could be removed from the training loop entirely. The space between training base models and the emergence of highly capable reasoning models—potentially smaller, more efficient reasoning models augmented with externalized knowledge—presents a crucial area for further exploration.

CSIRO’s Data61 is actively exploring these dimensions, focusing on system-level AI improvements and the stages after the pre-training/base model part. If you are interested in collaborating with us on these cutting-edge AI research areas, we welcome potential partnerships and discussions.

Full report https://council.science/publications/ai-science-systems/

Australian case study: https://www.dropbox.com/scl/fi/ltyxcmql58vrkcudxb7gf/Australia-Preparing-National-Research-Ecosystems-for-AI-second-edition.pdf?rlkey=dhvz8fk4atr3nfdm194ltlvi7&dl=0