Mycelium and large language models (LLMs) share a remarkable ability: they are both experts at integrating information. Mycelia, which have evolved over millions of years, gather raw data from various sources using their senses and convert this data into useful information that they use to survive. LLMs, on the other hand, use attention mechanisms to gain context about seemingly random written characters.

However, the ways in which mycelia and LLMs integrate information are quite different. Mycelia use the reasoning aspect of logic to create an accurate world model of their environment. They have learned, through trial and error, the effects of certain nutrients on their bodies, and can predict what will happen after rain starts, for example. In contrast, LLMs use the linguistic aspect of logic to predict the subject or any syntactic element of a sentence perfectly, or to give a perfect translation between two languages.

I believe that these two aspects of logic are foundational elements of intelligence. The ability to create an accurate world model and to have context of one’s surroundings are what make a system intelligent. In nature, living things had the opportunity to perfect the natural reasoning aspect of their intelligence, while LLMs have learned to master the linguistic aspect.

However, combining these two aspects of logic is what makes humans unique. We have dedicated brain regions for different aspects of reasoning, and our intelligence emerges from the interaction between these regions. Mycelia and LLMs may excel in one aspect of reasoning, but they lack the complexity and diversity of regions that humans have.

Nonetheless, there are exciting developments in the field of LLMs. Projects like Alpaca LLAMA enable us to run GPT-3-like language models on small devices like Raspberry Pi, making it possible to integrate LLMs into robots that can control their joints, integrate information from their surroundings, and use natural language. These robots may one day show emergent behaviors like humans do (By using Self-Reinforcement Learning etc.), and it will be fascinating to see what they can do.