What an intellectually stimulating two days at the AIST’s AI International Symposium 2024 in Japan! 🇯🇵🧠 The symposium focused on the “Future Direction for Trustworthy and Responsible AI”. Day 2 also covered the latest progress in AI for Science.

Yoshua Bengio ‘s talk on “Towards Quantitative Safety Guarantees and AGI Alignment” was thought-provoking and sobering all at once. He explored how we can build an AI that holds multiple alternative theories/world models that fit with current observations/data, using it to avoid catastrophic outcomes when selecting the best/safest theories to proceed.

I had an interesting discussion with Bengio afterwards about whether we can/should limit the world models/theories that AI learns to a human-understandable level. If not, how can we trust AI when it explains a human-understandable version to us? Bengio had the answer, but I can’t share it here in case a mischievous AI is eavesdropping! 😜🤫

On a more optimistic note, I gave my talk on “𝐄𝐧𝐬𝐮𝐫𝐢𝐧𝐠 𝐀𝐈 𝐒𝐚𝐟𝐞𝐭𝐲: 𝐀 𝐒𝐲𝐬𝐭𝐞𝐦-𝐋𝐞𝐯𝐞𝐥 𝐀𝐩𝐩𝐫𝐨𝐚𝐜𝐡 𝐭𝐨 𝐓𝐞𝐬𝐭𝐢𝐧𝐠, 𝐓𝐫𝐚𝐧𝐬𝐩𝐚𝐫𝐞𝐧𝐜𝐲, 𝐚𝐧𝐝 𝐀𝐜𝐜𝐨𝐮𝐧𝐭𝐚𝐛𝐢𝐥𝐢𝐭𝐲.” 🎙️ I emphasized CSIRO’s Data61 system-level and engineering approach beyond model training as the key to safe AI guarantees. I also highlighted how National AI Centre is developing Australia’s AI Safety standard that caters to AI deployers, small & medium enterprises, and prioritizes diversity, inclusion, and First Nations perspectives. 🇦🇺🌈 The positive feedback and acknowledgment of Australia’s approach were heartening! ❤️

book: https://lnkd.in/g9BCu6nn
research: https://lnkd.in/gyzjE4-i