IN PERSON EVENT: What Does It Mean for a Machine to Understand?

We will explore the deceptively simple question with enormous stakes: what does it mean for a machine to understand? Just as climate science needed a shared definition of the problem before action was possible, AI research now needs a clearer consensus on understanding itself.

For seventy years, intelligence research has been split between two paradigms. The logic-inspired view sees knowledge as symbolic rules and propositions manipulated by reasoning; learning is secondary. The biological, neural view treats learning as primary and locates knowledge in patterns of connection strengths. The turning point came in 2012, when deep neural networks outperformed hand-engineered systems in vision, resulting in the way AI is predominantly practiced today.

These models don’t store sentences; they construct them, using high-dimensional “feature vectors” that shift with context. And because digital agents can copy and share their internal weights, they can exchange trillions of bits of learned structure, compared to the few hundred bits carried by a spoken sentence. That makes their way of sharing knowledge radically more efficient than ours, with unsettling implications.

Our conversation will explore Geoffrey Hinton’s reflections https://www.youtube.com/watch?v=6fvXWG9Auyg on understanding, meaning, and neural networks. We’ll ask whether machines and humans model reality in surprisingly similar ways, how this challenges older symbolic theories of mind, and what happens when “understanding” itself becomes scalable and copyable.

Sources:
What Is Understanding? – Geoffrey Hinton | IASEAI 2025
https://www.youtube.com/watch?v=6fvXWG9Auyg

Questions to Think About

1. If neural networks and humans both build meaning through language, what sets our understanding apart?
2. How does our notion of meaning change if we stop thinking in terms of stored propositions and start thinking in terms of flexible, context-dependent representations?
3. What might it mean for society when digital agents can share knowledge at a scale and speed that no group of humans can match?