Do Agents Dream of Mycelium Networks?
Looking to fungi for biological inspiration
“Intelligence is based on how efficient a species became at doing the things they need to survive.”
― Charles Darwin
A lot of people question if AI thinks. I'm curious why people believe AI would think in a similar manner to humans? Different structures for reasoning provide different forms of intelligence. We can see examples of this in the world around us. Arguably, the organisms most different from humans that display vast forms of intelligence are fungi.
As humans, we tend to put ourselves at the top of intelligence and then assume decreasing intelligence the less similar something is to us. Despite this, we are unable to completely manipulate the brains and motivations of other species. Yet there are organisms in the fungal kingdom that can completely take over insects and control their bodies and motivations. The closest "humanity" has come to doing something similar are algorithms for social media feeds that keep people scrolling for hours.
Let's explore other forms of intelligence to understand how different approaches might help us improve the technology we build.
Mycology
Over a decade ago, a Japanese lab conducted a famous experiment with fungi. Researchers created a physical map of Tokyo and placed oats at the locations of surrounding cities that were connected via railway. The researchers then let a yellow slime mold grow over the map. In about a day, the slime mold created a network design just as efficient as Tokyo's subway system. Paul Stamets, a famous mycologist who has written several definitive guides on mushrooms along with patenting many applications of fungi, explains how this experiment worked.
As noted by Mark Flicker of the University of Oxford, "the slime mold had no central brain or indeed any awareness of the overall problem it is trying to solve, but manages to produce a structure with similar properties to the real rail network." This type of experiment has been repeated over and over again. These organisms don't have what we would consider a typical brain and yet they make complex decisions. How does that work? Are they simply physical manifestations of algorithms? What functions are these organisms using to make these highly optimal decisions? Are they minimizing energy transport? Do they detect change in fuel rates as their input? How do they decide how long to keep searching to find food vs taking a different path? These organisms have been around for hundreds of millions of years and have become adept at highly complex decision making, yet we still do not know much about how these organisms work.
Fittingly, there is a lab at the University of West England Bristol called the Unconventional Computing Lab (founded in 2001) that studies other forms of computation and intelligence. For instance, they've created methods for using fungi as electrical conductors and even created fungal computers. The research has shown an array of sensors and electrical spiking activity in fungi that operate in a manner similar to the human brain. Even without a brain, it is obvious they have some form of intelligence.
The best engineering and designs have been inspired by nature. As we seek to build better models and networks, we should look to fungi for inspiration. For instance, fungi appear to have multi-channel signaling, which is a combination of electrical, chemical, and tensile signaling. Multiple, separate types of inputs zero in on the correct decision. In contrast, if you look at the current state of LLMs, they combine both their instruction and data information into a single input channel. This causes massive issues for security, jailbreaking, guardrails, and general performance. Alternatively, studying how fungi grow and make decisions can be applied to improve the underlying algorithms used in AI models.
Symbiosis
A symbiotic relationship is "the living together of unlike organisms". While there are actually three types of symbiosis, we'll talk here about the one most people think of - where each organism benefits from the relationship. As we begin to embed AI everywhere, we're on a path to have a symbiotic relationship with AI. As we benefit from the performance improvements from using AI, AI increases its knowledge about the world and improves itself. Understanding what good symbiosis looks like will help us improve the benefits of this relationship. Unsurprisingly, fungi form many, many symbiotic relationships.
Around 90% of plants create symbiotic relationships with mycorrhizal fungi. The fungi use their broad reaching mycelium to engage with the plant and to wind their way through the soil to create vast networks which connect different plants together. At the same time, the plants exchange energy from photosynthesis for hard-to-reach resources such as water and phosphorous. This allows the trading of resources between different connected plants to trade resources and improves the overall health of all connected species. The papers that describe these mycelium networks attempt to understand the complex relationship along with what triggers nutrient and chemical transport.
What's interesting is that your view of how a symbiotic relationship or network works is based on your perspective. If you take a plant based perspective, then it appears that the various plants and trees are trading resources among each other, using fungal networks as nothing more than infrastructure to make the sharing of resources happen. However, if you take a fungal based view, then the fungus is shifting resources between plants to maximize their health in order for the fungus to maximize its own health. Shifting the perspective changes your interpretation of what is happening. However, in a symbiotic relationship, as in an ecosystem, both are likely true.
As we build out our symbiotic relationship with AI, how should we view it? What happens when we shift from a human-centric view to an AI-centric one? Likely, the paths of how one thinks about building these systems changes. I've spoken to many who have taken an AI-centric view. They are building workflows that are easier for AI's to interact with and consume, even if somewhat incomprehensible to humans. Consequently, these systems work superbly.
As AI becomes embedded everywhere, what will be the equivalent of network trading layers? How do we manage the flow and assessment of data to be used by AI agents? Are they able to find useful data sources on their own? What type of screening procedure are needed to assess the veracity and validity of the data? How can AI agents adapt dynamically to changing environments and changing user behavior? How can AI agents become great symbiotes just like fungi?
Deep Questions
The parallels between fungi and AI agents are uncanny, particularly around networks and symbiosis. We're clearly on a path of AI agents being one of the main ways we interact with AI models. We can broadly define AI agents as autonomous entities that take in information, process it, and take actions. The future we are building is starting to look like a network of ecosystems with intelligent components, each having their own motivations. Our future won't be as simple as modifying a machine part in a deterministic way.
However, it is important to remember that AI agents are separate beings. While they will work in concert with us, they are not us. This means they will have different decision-making criteria, different contexts, different understandings, and different goals from us. That's a very different paradigm to work with instead of servitude. What mental model is used when you have a set of agents interfacing with agents of other individuals and companies in order to accomplish a task of solve a problem for you? Naturally, we will lean towards thinking in terms of influence and complex networks. Which is why we should look to fungi for inspiration.
Fungi are masters at resource allocation and building and manipulating networks. They are participants in an ecosystem, just like AI agents will be. Both live in a world of dynamic equilibrium. An understanding of ecology and the mental models of its basic ideas will be critical to being successful in what's to come. The mental models extend such as:
Energy flow: the movement of energy through an ecosystem -> how do data and inferences flow through a series of agents?
Nutrient cycling: the movement and exchange of organic and inorganic matter -> how do inferences get reused and trigger actions by agents?
Population dynamics: how populations change over time and space -> which agents, agent architectures, and underlying systems succeed under which conditions?
Carrying Capacity: the maximum population size that an environment can sustain -> how much compute is required to achieve a given set of tasks at X scale? How many agents are required to automate Y number of functions?
Each of these concepts apply to both AI agents and fungi. Similar to fungi, AI models are poorly understood. This is exciting as it leaves a lot of room for opportunity. If you are building AI models or products using AI, it seems prudent to look to fungi for inspiration.