Why do you care about AI Existential Safety?

I believe as we progress towards more capable AI, potentially approaching artificial general intelligence (AGI), the risks and rewards become increasingly consequential. Ensuring that these systems are aligned with human values, robust against failures, and transparent in their operations is crucial to prevent unintended consequences and harm. AI safety is not just a technical challenge but a moral imperative. The decisions and developments we make today could shape the trajectory of human civilization. By focusing on AI existential safety, we are working to ensure that the immense potential of AI is realized in a way that benefits all of humanity, rather than posing existential risks.

Please give at least one example of your research interests related to AI existential safety:

One of my key research interests related to AI existential safety is enhancing the robustness of large vision-language models. This work is crucial for ensuring that advanced AI systems, particularly those that integrate visual and linguistic information, behave reliably and safely in diverse, real-world scenarios. For instance, my research on developing techniques like PromptSmooth aims to certify the robustness of medical vision-language models through prompt learning. This work ensures that these models can withstand adversarial attacks and function reliably in critical applications, such as healthcare, where errors could have life-threatening consequences.

Building on this, I am exploring how these robustness techniques can be scaled to larger, more complex multimodal large language models that are increasingly being deployed in high-stakes environments. My work focuses on not only defending against known adversarial attacks but also developing proactive strategies to anticipate and mitigate potential future vulnerabilities. This forward-looking approach is essential in the context of AI existential safety, as it helps prepare our AI systems for unforeseen challenges, ensuring they remain safe and beneficial for society.