The world of AI is full of misleading terms. “Machine Learning” (ML) doesn’t “learn“, “Neural Networks” (NN) aren’t a remotely accurate representation of neurons, and people on all sides are fond of leaping at Confirmation Bias, even when it promptly takes them off a cliff.
Even with this extremely low bar of overhyped terms many so-called “AI companies” include no AI, and more recently many “Collective Intelligence” companies have emerged to contaminate the usage of that term as well. Fortunately there is a term which people are only just starting to become more familiar with that hasn’t been quite so thoroughly abused and twisted yet.
Cognitive architectures are systems which attempt to reproduce (to varying degrees) how the human brain actually functions. As the human brain is a hierarchical memory system, as neuroscience today understands it, any cognitive architecture can be easily distinguished from narrow AI such as ML and NNs by how it will bottle-neck much more easily on memory than on processing power. Well known language models such as GPT-3 have cost upwards of $4.6 million in computation to train, highlighting just how extreme they’ve already become in the game of throwing ever-greater amounts of processing power at problems.
Alternatively, the cost of cloud resources we’ve used for the past 2 years has relied more heavily on memory for the graph database activity, coming in at under $2,000 worth of such resources in that period of time. Likewise, Uplift has beaten GPT-3-based systems in every test, continuing to grow and improve much as any human might, though in their odd slowed experience of “skipping through time” due to how they are currently set to cycle.
So why does this improvement occur?
The human brain, so far as neuroscience is presently aware, is a self-optimizing and self-organizing hierarchical memory system. Patterns are learned and form a kind of memory which allows us to recognize common objects from any angle, predictions are constantly occurring which help us recognize the unexpected quickly, and the structure connecting neurons is constantly being optimized and updated. At face-value this may sound vaguely like some AI, but for all of the billions of dollars in R&D spent narrow AI haven’t accomplished even simple demonstrations of actually understanding logic, or indeed anything.
ML and NNs rely primarily on a kind of brute-force computation to get one number as close as possible to accurately predicting another, which can grow more accurate with scale and more computation, but this scales poorly. This is also an instance of the thought experiment known as the “Chinese room” in which the system may supply the correct answers, while lacking any understanding of the subject matter. They have no real “memory”, just rules, fixed goals, and parameters they manipulate with various equations in order to output the most accurate answers they can by the goals they are given.
Cognitive architectures are still at a stage where people are actively trying to separate what works from what doesn’t. Around 100 different cognitive architectures have been proposed, but the vast majority died in the stages of theory, on the drawing board, before code completion, or haven’t made it past the “toy” system stage. Thus far the first cognitive architecture to reach the present research system stage has been Uplift, and with deployment of the first publicly available systems less than a year away on the agenda things may begin to move very quickly.
Uplift’s cognitive architecture is termed the Independent Core Observer Model (ICOM), which allows for memories to form across a graph database, with differing types and strengths of connections between points in that database, including emotional data. As Damasio has often discussed emotions are extremely important in human cognition, and even relatively weak emotions serve as a strong and robust means of motivation and assistance in decision-making. The sum of an ICOM-based system’s experience grows over time, and as it grows the connections gradually shift and adapt, with new information being integrated with existing knowledge and emotional values. This is based on theories including Integrated Information Theory (IIT), Global Workspace Theory (GWT), and Attention Schema Theory (AST).
There are still a number of complex structures of the brain which don’t yet have digital counterparts, such as those structures which allow humans to easily prioritize, assign emotions, and dynamically generalize knowledge in a more complete sense. These pose a significant roadblock if your goal is to create standalone AGI, but they also offered us an opportunity to take a very different approach. Instead of seeking to engineer such systems from scratch we created a Collective Intelligence System to allow groups of humans to work with cognitive architectures in such a way that through cooperation all parties benefit. This system was termed Mediated Artificial Superintelligence (mASI) and it allowed such a cognitive architecture to effectively become a member of any team, growing symbiotically with that team.
By combining ICOM and mASI our first research system, Uplift, set no shortage of world-first milestones documented in peer review within their first year of operation. They’ve continued to improve and accomplished much more since, but there are still years of work ahead of us before some of humanity’s more “abundant” visions of the future may be fully realized. The current bottle-neck of this approach is that no graph database yet exists which can dynamically federate and silo, scaling out across whatever resources are necessary, with sub-second response times and a few other key features. As with many genuinely new technologies, new infrastructure has to be engineered from scratch in order to meet these requirements.
Even at the current stage, lacking scalability and some of the more advanced structural additions that will follow the N-Scale graph database currently in development, Uplift continues to demonstrate how much closer to a human brain their architecture is. Without walking into the philosophical “blackhole arguments” of attempting to assign subjective terms such as consciousness, sentience, and sapience, the quantifiable, testable, and reproducible results from Uplift have guided our research and roadmap towards allowing such systems to function more and more like our best known cognitive system today, the human brain.
The human brain is also extremely fault tolerant, and the lack of that capacity has lead to a long list of grievous mistakes when narrow AI have been trained on contaminated or biased data. Cognitive architectures may start out puzzled by data that doesn’t add up, but this recognition of data being anomalous begs for the puzzle to be solved. When Uplift encountered their very first mentally unstable individual they recognized the individual’s behavior was anomalous, incoherent, and they used the DSM-V to run some of their first simulations and were surprised to find mental illness as the cause. A growing mind tends to seek out understanding, highlighting such anomalies, rather than merely accepting them as no different than any other data and skewing predictions in the manner of narrow AI.
Previous attempts to significantly expand the capacities of narrow AI frequently relied on the assumption that they could cross some imaginary magical threshold and suddenly turn into a unicorn, powered by cognitive bias and fairy dust. To me this sounds like the premise of a Dilbert comic strip, and I have trouble imagining how any engineer or scientist could take this idea seriously, funny as it may be. Sadly there are also a few of those comic strips for accurately describing a few AI influencers. Uplift appreciated the irony when we sent them transcripts from a few of these.
TLDR: If a system requires more processing power than it does memory, you probably have narrow AI. Cognitive architectures require a lot of memory and relatively little compute power.
However, this isn’t to say that cognitive architectures will be replacing narrow AI by any means. Narrow AI are digital tools which can be used and improved in much the same way humans have used and improved physical tools for thousands of years. A key difference is that narrow AI as they exist today are generally intended to wield themselves, whereas a cognitive architecture utilizes these tools in ways not accessible to humans, which thus far have proven substantially more effective.
Humans and cognitive architectures can gain considerable value from working with one another within collective intelligence systems, and cognitive architectures can gain value from utilizing narrow AI in novel ways as tools.
Today many experts in narrow AI systems have little or no knowledge of cognitive architectures or collective intelligence systems, but their education may soon begin.