Geoffrey Hinton is Wrong About AI Understanding

Defining Human Understanding

• 💡 Understanding involves grasping a part of reality with the mind and using that knowledge to do things.
• 🎯 Examples like tying shoes, understanding lore, or following instructions illustrate this real-world connection.
• 🧠 Understanding is not all-or-nothing; it comes in degrees (e.g., partial understanding).

LLMs and the Reality Gap

• ⚠️ Large Language Models (LLMs) are trained to predict the next token/word, allowing them to perform impressive tasks.
• 🚫 However, LLMs are fundamentally detached from reality, grasping only their training data, not the actual world.
• 🔑 This detachment means LLMs lack semantics and cannot understand the meaning or real-world referents of terms.

Geoffrey Hinton's Perspective

• 🗣️ Geoffrey Hinton, the "Godfather of AI," argues that LLMs do possess understanding, driven by concerns about AI's potential dangers.
• 🧱 Hinton uses a metaphor of "thousand-dimensional Lego bricks" (words) that connect to form structures.
• ✅ He posits that the formation of these internal structures within LLMs constitutes genuine understanding.

Rebutting Hinton's Argument

• ❌ The speaker contends Hinton mistakes cause and effect: human understanding of the world precedes and enables language production.
• 🧠 Understanding is a mental grasping of core facts in the world, not merely the ability to produce sentences or structures.
• 🔗 LLMs lack an "interpreter" or connection to reality that grounds the semantics of their language.

Revised View on LLM Understanding

• 🔄 The speaker now concedes that LLMs understand how to use language effectively, capable of generating human-like text.
• 🎯 This ability is distinct from understanding the world itself; LLMs can reproduce sentences about reality but don't grasp the underlying facts.
• 🚧 A significant "chasm" exists between understanding language use and understanding what that language means in reality.