Cognitive Architectures for AGI: Integrating Human Intelligence Models

The podcast, featuring Nate Derbinsky, introduces cognitive architecture as a crucial approach to achieving Artificial General Intelligence (AGI). Derbinsky argues that current AI/machine learning systems, while powerful in specific domains, lack the robustness, persistence, and generalization capabilities required for human-level intelligence, especially when facing unforeseen tasks. He highlights the challenge of integrating disparate, task-specific theories (like Fitts' Law or the Power Law of Practice) into a unified framework capable of tackling complex problems like chess. The core proposition of cognitive architecture, stemming from Alan Newell's "Unified Theories of Cognition," is to consolidate fundamental assumptions about intelligent agents' fixed mechanisms—such as representations, learning, and memory systems—into a single, implementable theory applicable across diverse tasks. This integration, by forcing interconnections, imposes constraints that refine the design space for intelligent systems. Derbinsky distinguishes between three motivations for AGI research: pure curiosity, cognitive modeling (understanding human intelligence to predict behavior), and systems development (building functional systems). He contrasts different benchmarks for intelligence, from the behavioral focus of the Turing Test to the internal cognitive processes sought by cognitive modelers, and the optimal rationality pursued by some AI researchers. The concept of "Lakatoshian science" is introduced to explain how scientific progress occurs in this field, with a growing core of established beliefs surrounded by ephemeral hypotheses. The discussion also delves into Newell's Time Scales of Human Action, illustrating how regularities occur at different temporal levels, from neuronal to sociological, with most cognitive architectures targeting the "deliberate act" level. Furthermore, Herbert Simon's concept of "bounded rationality" is emphasized, acknowledging that human intelligence operates under computational, cognitive, and temporal constraints, often leading to "satisficing" rather than optimal solutions. The episode showcases several prominent cognitive architectures, including Act-R, Soar, Sigma, and Spawn, detailing their core components and applications. Act-R, for instance, is used for predicting human performance and errors in HCI, allowing for rapid iteration in interface design. Sigma is presented as a newer architecture capable of integrating modern machine learning techniques, vision, and optimization on a uniform substrate. The discussion on Soar highlights its chunking mechanism for learning new rules through sub-goal reasoning. For those interested in the field, Derbinsky recommends exploring these architectures, attending workshops, and engaging with the provided resources. The underlying message is that by understanding and modeling the fixed mechanisms of human cognition, we can build more robust, adaptive, and general-purpose AI systems. The pursuit of cognitive architectures has profound implications for understanding the nature of intelligence itself, bridging disciplines from neuroscience and psychology to computer science and economics. It suggests that a truly general AI might not emerge from simply scaling up current narrow AI techniques but from a foundational understanding of how intelligence is structured and operates across tasks. The shift from purely symbolic representations to hybrid approaches incorporating non-symbolic (perceptual, statistical) processing reflects an evolving understanding of intelligence. Ultimately, this research aims to create AI that can learn, adapt, and interact with humans in a more natural, teachable, and robust manner, moving beyond the "Sorry, I don't know that one" limitations of current virtual assistants like Alexa, towards systems that can ask, "Do you have time to teach me?"

The bigger theme here is I want to talk about what's called cognitive architecture and if you've never heard about that before that's great and I wanted to contextualize that as how are we what how is that one approach to get us to AGI.

We want systems that are going to persist for a long period of time we want them robust to different conditions we want them learning over time and here's the crux of it working on different tasks and in a lot of cases tasks they didn't know we're coming ahead of time.

Sorry I don't know that one" which is okay right that's a lot of people have no idea what I'm asking let alone how to do that so what I want Alexa to respond with after that is do you have time to teach me and to provide some sort of interface by which back and forth we can kind of talk through this.

It's really understanding how humans think how humans operate human intelligence at multiple levels and if you can do that one there's just knowledge in and of itself of how we operate but there's a lot of really important applications that you can think of if we were able to not only understand but predict how humans would respond react in various tasks.

His argument to the psychology community at that point was the argument on the prior slide they had many individual studies many individual results and so the question was how do you bring them together to gain this overall theory how do you make forward progress.

To bring together your core assumptions your core beliefs of what are the fixed mechanisms and processes that intelligent agents would use across tasks so the representations the learning mechanisms the memory systems bring them together implement them in a theory and use that across tasks.

Decision-makers can satisfice either by finding optimum solutions for a simplified world just to say take your big problem simplify in some way and then solve that or by finding satisfactory solutions for a more realistic world.

A weak interpretation of the idea that symbol systems are necessary and sufficient for intelligent systems a very weak way of talking about it is the claim that there's nothing unique about the neuronal infrastructure that we have but if we got the software right we could implement it in the bits bytes Ram and processor that make up modern computers.

The idea is that these primitive cycles allow for the agent to be reactive to its environment so if new things come in that has react to if the Lions sitting over there I better run and maybe not do my calculus homework right so as long as this cycle is going I'm reactive but at the same time if multiple actions are taken over time I'm able to get complex behavior over the long term.

He reimplemented the the core functionality of soar all using factor graphs and message passing algorithms under the hood he got to that point and then said there's nothing stopping me from going further and so now it can do all sorts of modern machine learning vision optimization sort of things that would take some time in any other architecture to be able to integrate well.