Stanislav Fort: Robustness and Security in the Age of Massive Multi modal AI Models (HAAISS 2025)

The talk centers on the robustness and security challenges facing large multimodal AI models, particularly regarding adversarial attacks. It begins by tracing AI's ambitious origins and highlighting the surprising ease with which digital minds have been created, requiring vast computational resources and data rather than deep insight. Despite AI's average robustness and generalization capabilities, the speaker demonstrates how easily classifiers can be deceived through crafted perturbations, even in the largest models like GPT-4, using examples such as a Rickroll-encoded Stephen Hawking image. The speaker explores differences between human and machine vision, suggesting micro and macro saccades contribute to human robustness. While standard solutions like adversarial training are deemed unscalable, the talk concludes by encouraging new approaches to address AI's vulnerabilities, emphasizing the importance of high-dimensional geometry.