Monthly Researcher Note by @_agoodtraveler
For decades (centuries, really!) we have, reasonably, assumed that creating intelligent machines is a matter of designing the right algorithm. The wall we hit was the unexpected complexity of the problem itself, which led to brittle, unsatisfactory results. What was discovered instead is that we can specify a different algorithm: one that defines how to search for the method that implements intelligence, and let computers do the work. Current advances in AI are the pinnacle of this approach.
Neural Cellular Automata (NCAs) are one possible continuation on this trajectory. They begin with almost no imposed structure, consisting of local rules on a grid, each informationally coupled only to its immediate neighbors. In this way interesting message-passing patterns, and by that the organization of computation itself are discovered rather than designed. This lets us evaluate a wider range of solutions than pre-assumed topologies allow.
Here, the properties (e.g. world modeling, agency, intelligence) we are looking for are emergent from the interactions of identical units across the substrate (the NCA grid). This emergence from simple units, inside-out instead of outside-in, also seems to align with what we see in biological systems.
The engineering work at CIMC involves building systems that let us specify the properties of interest and efficiently run the search, evaluation, and analysis of the resulting automata. Our current hypothesis is that consciousness might be one of those interesting properties, since we observe it as apparently indispensable part in many biological organisms.
The aim is not to declare any particular instance conscious, but to maintain a controllable lab in which the conditions for that kind of organization can be found.
Check out our latest Substack: cimcai.substack.com/p/cimc-newslet…
For decades (centuries, really!) we have, reasonably, assumed that creating intelligent machines is a matter of designing the right algorithm. The wall we hit was the unexpected complexity of the problem itself, which led to brittle, unsatisfactory results. What was discovered instead is that we can specify a different algorithm: one that defines how to search for the method that implements intelligence, and let computers do the work. Current advances in AI are the pinnacle of this approach.
Neural Cellular Automata (NCAs) are one possible continuation on this trajectory. They begin with almost no imposed structure, consisting of local rules on a grid, each informationally coupled only to its immediate neighbors. In this way interesting message-passing patterns, and by that the organization of computation itself are discovered rather than designed. This lets us evaluate a wider range of solutions than pre-assumed topologies allow.
Here, the properties (e.g. world modeling, agency, intelligence) we are looking for are emergent from the interactions of identical units across the substrate (the NCA grid). This emergence from simple units, inside-out instead of outside-in, also seems to align with what we see in biological systems.
The engineering work at CIMC involves building systems that let us specify the properties of interest and efficiently run the search, evaluation, and analysis of the resulting automata. Our current hypothesis is that consciousness might be one of those interesting properties, since we observe it as apparently indispensable part in many biological organisms.
The aim is not to declare any particular instance conscious, but to maintain a controllable lab in which the conditions for that kind of organization can be found.
Check out our latest Substack: cimcai.substack.com/p/cimc-newslet…