TLDR: Tom Lue, General Counsel and Head of Governance at Google DeepMind, highlights the company’s dedication to developing robust systems for identifying and mitigating potential risks in advanced AI models, especially as multimodal AI becomes more prevalent. This effort involves a combination of human oversight and AI-driven “auto evaluations” to analyze and predict threats in nascent foundation models.
Google DeepMind is intensifying its efforts to understand and control the capabilities of its advanced artificial intelligence creations, particularly with the emergence of multimodal AI. Tom Lue, General Counsel and Head of Governance at Google DeepMind, emphasized the company’s commitment to building sophisticated systems designed to identify and mitigate potential dangers inherent in these evolving models. This initiative comes as the AI industry grapples with the complexities of models trained on diverse data types, including text, video, and audio.
The challenge is significant, as the most advanced AI models are now multimodal, making their behavior and potential risks more intricate to assess. Lue acknowledged that while Google DeepMind employs some of the world’s leading experts in this field, the science of AI evaluation is still in its early stages, requiring further investment, research, collaboration, and the establishment of best practices.
In a competitive landscape where AI safety is increasingly seen as a strategic advantage, Google DeepMind has revealed details of its approach to preemptively identify dangerous advancements in AI capabilities. This strategy involves a dual approach: leveraging human expertise alongside “auto evaluations,” where AI models are designed to analyze other models, or even themselves, for potential threats.
Also Read:
- Global Alliance Formed to Advance AI Safety and Control
- xAI Commits to EU’s Generative AI Safety Code
Google has been working on AI safety for years, having recently released a comprehensive framework for assessing AI model dangers. The urgency of these efforts has escalated with the global, multibillion-dollar race to enhance AI capabilities, driven by the success of foundation models like GPT and DeepMind’s Gemini. Despite their rapid development, these powerful foundation models are still in their infancy and do not yet pose imminent threats. However, researchers are proactively designing methods to analyze a technology that is still being created, aiming to ensure safety as AI capabilities continue to expand.
