This is an automated rejection. No LLM generated, assisted/co-written, or edited work.
Read full explanation
This post is an attempt to formulate an architectural hypothesis, not to argue that it is already correct.
I am presenting it as an open research question, not a proposal for achieving AGI or artificial consciousness.
Most current LLM progress comes from scaling a single model: more parameters, more data, longer context windows, and more inference-time computation. That approach works well. But I keep wondering whether we spend too little effort asking if a more complex surrounding architecture could produce qualitatively different reasoning.
This post explores one possible direction.
The general area is not new. It overlaps with cognitive architectures, multi-agent systems, Society of Mind, Global Workspace Theory, and agent-based systems. I am not claiming to have invented any of those things.
The narrower question is: what happens if an LLM is embedded in a system with continuously running autonomous background processes that are not explicitly invoked by the main model?
What is actually new here?
This is not simply “multiple agents.” Multiple agents can still behave like tools: the main system calls one, waits for a result, and moves on.
The distinguishing property is persistence. Background processes continue computing even when their work is unrelated to the current task. They are not tools selected by the planner. They are not called. They keep generating information, most of it probably irrelevant, and occasionally something they produce may redirect the main model’s reasoning.
In other words, the system is never cognitively idle.
What might run in the background?
Some processes could loosely resemble features of biological cognition:
continuous sensory monitoring;
autobiographical memory retrieval;
spontaneous associations;
shifts of attention.
Others would not need a biological analogue:
detection of statistical anomalies;
search for structural symmetries;
estimation of environmental novelty;
generation of distant associative links.
Most of these signals might be useless. But an occasional memory, anomaly, or analogy could interrupt the current line of reasoning and suggest a different path.
Why might this be useful?
Biological cognition appears to exhibit one feature that current LLMs largely lack: continuous internal activity.
We do not start thinking only when somebody asks us a question. Perception, memory, bodily state, and loose associations continue in the background. Much of this may be noise. Still, some useful ideas seem to come from unrequested collisions between ongoing processes.
The question is whether a structured version of persistent internal activity could help a system escape a narrow reasoning trajectory.
A simple thought experiment
Imagine an autonomous engineering agent working on a long-running design problem. While its main planner works, several independent processes continue running:
searching for structural analogies;
monitoring statistical anomalies;
estimating long-term environmental changes;
retrieving distant memories.
None were explicitly requested. Most outputs go nowhere. But one process might notice that the design resembles an old failure case, or that a slow environmental change invalidates an assumption. That signal enters the main context and changes the reasoning trajectory.
The interesting part is that the specialist was already working before anyone knew it would be useful.
This is not an attempt to imitate humans
The point is not to build a digital human. Human cognition is only one example of a system with persistent internal dynamics, and probably not the ideal template.
A machine version could contain processes humans do not have. The relevant property is ongoing computation that is partly independent of the immediate prompt.
A possible research hypothesis
A direct experiment would compare otherwise similar autonomous systems: one receives inputs only from explicitly invoked tools, while the other also receives a limited stream of unsolicited signals from persistent processes.
I would want to know whether the second system behaves differently on tasks involving:
long-term planning;
creative problem solving;
generation of remote associations;
recovery from local minima in reasoning.
The result could be an improvement, a change that is merely strange, or nothing at all. I do not think the answer is obvious.
What could be wrong?
The hypothesis may simply be incorrect:
persistent background computation may only increase computational cost;
sufficiently large transformer models may already approximate some of this behavior internally;
no qualitative improvement may emerge.
There are practical problems too: how signals compete for attention, how to prevent distraction loops, and how to evaluate a qualitative change without fooling ourselves.
Even if the hypothesis is wrong, the architectural question seems worth investigating. Scaling a model and changing the organization around it are different experiments.
An open question
Current LLMs are almost entirely reactive.
Biological cognition appears to remain continuously active.
Perhaps reproducing this property changes nothing.
This post is an attempt to formulate an architectural hypothesis, not to argue that it is already correct.
I am presenting it as an open research question, not a proposal for achieving AGI or artificial consciousness.
Most current LLM progress comes from scaling a single model: more parameters, more data, longer context windows, and more inference-time computation. That approach works well. But I keep wondering whether we spend too little effort asking if a more complex surrounding architecture could produce qualitatively different reasoning.
This post explores one possible direction.
The general area is not new. It overlaps with cognitive architectures, multi-agent systems, Society of Mind, Global Workspace Theory, and agent-based systems. I am not claiming to have invented any of those things.
The narrower question is: what happens if an LLM is embedded in a system with continuously running autonomous background processes that are not explicitly invoked by the main model?
What is actually new here?
This is not simply “multiple agents.” Multiple agents can still behave like tools: the main system calls one, waits for a result, and moves on.
The distinguishing property is persistence. Background processes continue computing even when their work is unrelated to the current task. They are not tools selected by the planner. They are not called. They keep generating information, most of it probably irrelevant, and occasionally something they produce may redirect the main model’s reasoning.
In other words, the system is never cognitively idle.
What might run in the background?
Some processes could loosely resemble features of biological cognition:
Others would not need a biological analogue:
Most of these signals might be useless. But an occasional memory, anomaly, or analogy could interrupt the current line of reasoning and suggest a different path.
Why might this be useful?
Biological cognition appears to exhibit one feature that current LLMs largely lack: continuous internal activity.
We do not start thinking only when somebody asks us a question. Perception, memory, bodily state, and loose associations continue in the background. Much of this may be noise. Still, some useful ideas seem to come from unrequested collisions between ongoing processes.
The question is whether a structured version of persistent internal activity could help a system escape a narrow reasoning trajectory.
A simple thought experiment
Imagine an autonomous engineering agent working on a long-running design problem. While its main planner works, several independent processes continue running:
None were explicitly requested. Most outputs go nowhere. But one process might notice that the design resembles an old failure case, or that a slow environmental change invalidates an assumption. That signal enters the main context and changes the reasoning trajectory.
The interesting part is that the specialist was already working before anyone knew it would be useful.
This is not an attempt to imitate humans
The point is not to build a digital human. Human cognition is only one example of a system with persistent internal dynamics, and probably not the ideal template.
A machine version could contain processes humans do not have. The relevant property is ongoing computation that is partly independent of the immediate prompt.
A possible research hypothesis
A direct experiment would compare otherwise similar autonomous systems: one receives inputs only from explicitly invoked tools, while the other also receives a limited stream of unsolicited signals from persistent processes.
I would want to know whether the second system behaves differently on tasks involving:
The result could be an improvement, a change that is merely strange, or nothing at all. I do not think the answer is obvious.
What could be wrong?
The hypothesis may simply be incorrect:
There are practical problems too: how signals compete for attention, how to prevent distraction loops, and how to evaluate a qualitative change without fooling ourselves.
Even if the hypothesis is wrong, the architectural question seems worth investigating. Scaling a model and changing the organization around it are different experiments.
An open question
Current LLMs are almost entirely reactive.
Biological cognition appears to remain continuously active.
Perhaps reproducing this property changes nothing.
Perhaps it changes everything.
I don’t know.
But it seems like an experiment worth running.