AI vs Human Thinking: How Large Language Models Really Work

Summary

The video compares human cognition with large language models (LLMs) across six areas: learning, information processing, memory, reasoning, errors, and embodiment. Its central point is that humans and LLMs can produce similar outputs through fundamentally different mechanisms.

Learning

Human brains adapt through neuroplasticity, changing neural connections in response to experience. People can sometimes learn from only a few examples. LLMs are trained using backpropagation, which adjusts model weights across large datasets and many training passes. Unlike people, a trained model’s weights generally do not update through ordinary use.

Information Processing

The brain processes information in parallel across specialized, interconnected regions and works with meaning and concepts. LLMs represent text as tokens and use attention across layers to help predict the next token. The video characterizes this as pattern completion rather than human-like comprehension.

Memory

People have sensory, working, and long-term memory, with memories connected by meaning, context, and emotion. An LLM’s learned knowledge is encoded in its weights, while its temporary working context is limited by its context window. Information may no longer be available once it falls outside that window.

Reasoning

The video relates human thought to fast, intuitive “System 1” thinking and slower, deliberative “System 2” thinking. LLMs can be prompted to produce intermediate reasoning steps, but the speaker argues that plausible-looking reasoning does not necessarily mean the model consciously understands or applies logic. This may help explain failures on seemingly simple tasks, such as counting letters.

Errors and Confabulation

LLMs can hallucinate, producing confident but false statements. The speaker compares this with human confabulation, in which someone unknowingly supplies a false memory or explanation and believes it to be true.

Embodiment

Human understanding is shaped by direct interaction with the physical world. LLMs are software without sensory or physical experience, so their knowledge of the world comes indirectly from text. The video links this lack of grounding to weaknesses in common-sense reasoning.

Conclusion

Humans contribute lived experience, meaning, and comprehension, while AI offers speed and broad knowledge. Combining them effectively can make use of both.

Source and Format

• Main speaker: Martin Keen, an IBM Technology presenter and narrator
• Format: Explanatory comparison; not a review, guide, or tutorial