Creating a dynamically integrated network (DIN)

A dynamically integrated network (DIN) is a computing architecture that balances local specialisation with global connectivity. The problem with Large Language Models (LLMs) is that they are a single massive system that activates for every token. The alternative is calling up multiple agents to deal with each specialist problem causing latency and creating inconsistences that need resolving.

There are features of DINs that may offer advantages over LLMs such as state dependent routing, structural grounding of the data and the ability to maintain a slow changing database (memory) whilst allowing fluid mathematical activations. Functionally the system would be multimodal by design and features such as common sense, creativity and emotional intelligence are likely to emerge. 

The problem with creating a DIN is that the only example we have studied is the human brain and this has proven too complex to understand. As scientists unlock the brain mechanisms of flies the lessons they learn will help us make sense of the dynamic systems in more complex brains. Adapting LLMs to become DINs is a plausible idea but it may be easier to take a reductionist approach informed by neurological research.

Why is the human brain a dynamically integrated network?

Intuitively the DIN makes no sense, how do synaptic changes store information, how does it know which of the situations does it apply to? The cortical columns must play a part but also the internal wiring; if an area is connected to the eyes then it does not need to store that this is optical information. The finding that most cortical columns have some location tracking information and there are topographical maps of function across the cortex suggest how information is grounded in DINs. 

The inhibitory pathways in a DIN dominate the system and this is likely to be a version of the attention system. The brain is close to a phase shift so can use frequency matching to link distant areas for instance when creating long term memories. Memories are created by synaptic growth so change slowly whereas nerve activation is near instant. The DIN uses the difference in speed between these processes to prevent interference between thoughts and memories.

Could a LLM become a dynamically integrated network?

LLMs have significant barriers to creating a DIN and would need frozen weights to be replaced with dynamic alternatives without causing model collapse. The activation pattern would need to be reduced from the whole model to localised areas. The LLM would not rely upon a single sub model but require multiple sub models to ensure reliability. LLMs will need to move from unlabelled to labelled data so that it can keep track of where as well as what. 

Comparing the human and LLM versions of the DIN offers a sobering conclusion. The brain’s inhibitory pathways, cortical columns, synaptic memory, reusing of cortex and linked modularity appears many times more complex than the proposed LLM version. Functionally, this gap of complexity will mean that emergent features will be lost and the system will lack stability. These experiments will likely have advantages but fall short of the DIN. Diffusion systems could be used as prediction engines but otherwise has limited advantages over LLMs.

How can we create a dynamically integrated network?

Humans create things by building the parts and then putting them together. In the case of a DIN this means starting with simple and small components, putting them together and then learning how the system works. This iterative approach will allow flexible comparison of different approaches such as analogue and digital versions. The focus would be on whether it can respond to a continuous input rather than prompts. Basically, plugging the system into sensors or a robot and watching it learn. 

There are at least 3 other elements, one is hardwired reflex nodes which includes emotions and other responses to stimuli. A sensory motor prediction system like a little brain that brings together different modalities to predict what is going to happen next. A control area which integrates the other information and allocates attention by inhibiting areas that are not important. The control area coordinates complex responses such as communication and actions towards a goal such as breaking down the problem, choosing the right methods, simplifying and generalising.

Conclusions

The dynamically integrated network (DIN) is considered by some to be the holy grail and others to be too complex to achieve with current technology. Designing a system that works on the edge of chaos is likely to require a different mindset to the controlled approach taken to LLMs. Taking a functional viewpoint provides the answer that order emerges from chaos in most biological systems. Similar to the finding that noise helps LLMs find stable solutions, linking unstable systems is likely to create resonance. 

The creation of dynamically integrated networks is likely to require input from experts in analogue systems, chaos theory, self-driving cars, clinical neurology, neurobiology and disability. Our current understanding of DIN is limited and history indicates that humans do not have the cognitive capacity alone to make progress. We should explore DIN because it has the potential to solve AGI but there is another deeper reason, DIN research may help us understand our brain. 
 

August 2026