DeepMind's Ayoub: Companies Can Fence In AI Like Banks Did

Google DeepMind's Kareem Ayoub says companies can build a fixed perimeter around their AI systems to ensure safety, much like banks did in the late 1980s and early 1990s. He spoke at HumanX in Amsterdam about the tension between exponential capability and linear governance.

Ayoub noted that while it's difficult to fully govern an AI system itself, a deterministic governance perimeter can be built around it. This approach is based on three key parts: the role of AI in research and development, transparency, and control.

Ayoub also emphasized the importance of partnerships beyond tech companies in regulating AI. He mentioned Google's Frontier Safety Framework as an example of how to set capability thresholds and mitigations. Ayoub highlighted a gap between what technology can do and how companies use it, with examples including speeding up drug discovery, translating ancient texts, and detecting gravitational waves.

He also discussed the shift from models to agents as products, where an agent wraps a model in memory, tools, and internet access for long, multi-step tasks. Ayoub cited software engineering as one of the strongest areas with returns so far.