The author, a cognitive scientist, suggests that understanding is not about the manipulation of strings (words, phrases, etc.), but about how these strings are related to things in the world or to concepts in our heads. This process, known as grounding, can be achieved in two ways: relating strings to sensori-motor experiences (direct experience) or relating strings to concepts (cognition). The author concludes that the latest generation of LLMs, which use knowledge graphs to systematically relate strings, sensori-motor inputs, and concepts, are much closer to achieving human-like understanding.
Key takeaways:
- The article discusses the ongoing debate about whether large language models (LLMs) like ChatGPT and Bard truly understand the questions and responses they generate, or if they are simply mimicking understanding.
- The author argues that understanding is not about being right or wrong, but about the difference between LLM performance and human understanding. This could lead to a deeper understanding of LLM limitations and potentially improve AI systems.
- Understanding, according to the author, is related to what we do with strings (words, images, sounds, etc.). To understand strings, we need to be able to relate them systematically to things in the world or to our existing knowledge. This process is often referred to as grounding.
- Knowledge graphs play a crucial role in enabling the latest generation of LLMs to systematically relate strings, sensori-motor inputs, and concepts. This has brought us closer to AIs that understand in a way that is very similar to how humans do.