#74 Dr. ANDREW LAMPINEN - Symbolic behaviour in AI [UNPLUGGED]

Human vs Machine

Andrew Lampinen explores the nuanced differences between human and machine reasoning, highlighting the limitations of AI in replicating human cognitive processes. He illustrates this with the Wason selection task, where humans excel when the task is framed in familiar contexts, unlike AI, which struggles with such biases 1. Keith Duggar adds that human rationality is often misunderstood due to evolutionary influences, which are essential for exploration and adaptation 2. Andrew argues that while logical reasoning is a tool humans have developed, it is not inherently part of our intelligence, suggesting that AI should similarly learn to use these tools rather than have them built-in 3.

   

Language Models

The discussion shifts to the capabilities of large language models and their mathematical understanding. Andrew Lampinen notes that these models often memorize rather than reason, as seen in arithmetic tasks where tokenization issues arise 4. He shares an example of a language model learning long division with detailed guidance, emphasizing the importance of structured learning environments 5. Andrew also points out that both humans and AI exhibit biases in reasoning, which can be mitigated through systematic training 1.

   

Interacting with Humans

Andrew Lampinen discusses the importance of training AI in environments that require human interaction to enhance adaptability and learning. He suggests using datasets of human interactions as a foundation for AI training, allowing systems to be shaped by human culture 6. Keith Duggar emphasizes the need for AI to be embedded in human culture to ensure compatibility and usefulness, warning against isolated AI development 7. Andrew's paper on symbolic behavior argues that understanding symbols through their use, rather than content, is crucial for developing human-like intelligence in AI 8.