Artificial General Intelligence: Borrowing from the Brain
If we want to build an artificial system that thinks like a human, it makes sense to start by understanding how humans actually think.
Here are three theories that offer compelling frameworks that could help in terms of how we approach this :
1. Modularity of Mind
Jerry Fodor proposed in 1983 that the brain is composed of specialized, domain-specific modules. For example, our visual-processing system is distinct from our language-processing system. These modules operate independently, with high-level cognition acting as an overarching system that integrates their outputs. Cosmides and Tooby extended this rigid compartmentalised model by arguing for a more flexible structure. They suggested that cognition itself is modular—not just perception or language, but reasoning, problem-solving, and decision-making. It suggests that human intelligence isn’t a single, unified system but a collection of specialized computational functions evolved to handle different survival challenges, and that these modules themselves are interconnected. The modern view is that whilst function is very much modular, the physical makeup of this processing is much more fluid with neuroplasticity and the brain’s ability to adapt to different functions being evident.
?? Why it matters for AGI: AI today largely follows this model. We have task-specific machine learning models—image recognition, text generation, and speech synthesis—all performing well within their narrow domains. But these systems don’t truly “talk” to each other in a meaningful way. AGI would require the integration of these disparate systems under a high-level decision-making framework. If cognition is inherently modular, then AGI will need to be built as a network of semi-autonomous, specialized modules. The challenge will be designing the communication protocols between them—how does a "logic module" interact with an "intuition module"? How do we ensure that reasoning isn’t siloed??
2. Friston & Clark’s Predictive Processing Model
Predictive Processing theory suggests that the brain isn’t just passively interpreting sensory input—it’s constantly making predictions about the world and updating those predictions based on feedback. This explains why humans can function in noisy environments, anticipate future events, and learn from incomplete information. Our brains constantly generate hypotheses about the world and review sensory data against them, updating them as needed on a continual basis. AI’s rely on training models currently and are essentially static. Each time they need to be updated with new information, they need explicit retraining.?
?? Why it matters for AGI: LLMs like GPT already operate as prediction engines but with a key limitation: they don’t update their internal models in real-time. AGI would need to incorporate feedback loops - real-time, self-learning mechanisms that allow it to refine its knowledge dynamically, much like a human brain does.
3. Beyond Prediction into Self-Supervised Learning
Learning beyond prediction is important for operable scale. There are 6 key approaches to learning which are each worth a deep dive but I’ll summarise them here:
Casual Learning: based on the Causal Inference framework, this approach allows causation as well as correlation, and allows real-world decision making. (AlphaFold, DoWhy leverage this but it's an extremely expensive approach that requires a high level of data structure)
Reinforcement Learning: trial and error decision-making via rewards and penalties, allowing the AI to optimise decisions made over time. (ChatGPT and AlphaGo leveraged this along with using human preferences - it is, however, inefficient in complex problems and only copes well within well-defined parameters)
领英推荐
Generative Learning: where the AI generates new data according to accepted patterns. MidJourney and Dall-E image generation are good examples of this but there is no reasoning behind it, and it is often prone to hallucination
Few-Shot and Zero-Shot Learning: designed to be similar to how humans learn, leveraging minimal or even no data to learn and adapt. This is used in GPT-4 and Google’s PaLM model but is still very reliant on pretraining and not very accurate which problems are unfamiliar.
Neuro-Symbolic AI: this is a hybrid approach between pattern recognition/prediction and rules-based logic to allow an AI to provide a semblance of human reasoning and a clear trail for how it reached a decision. Used by IBM and medical diagnostic solutions. Requires very structured data and is much harder to scale.
Self-Supervised Learning : I find this approach particularly interesting as it describes scalable continuous learning from unlabelled data (GPT-4, Gemini, and Claude all leverage this but there is no associated reasoning)
Yann LeCun (Meta’s Chief AI Scientist) proposes 6 systems loosely based on how the brain processes information, to prioritize self-learning, and then add hierarchical reasoning :
LeCun’s theory is that AGI needs to avoid too tight coupling to human structures and that it will emerge from self-supervised learning and hierarchical reasoning (rather than rule-based decisions) as we move beyond more and more sophisticated prediction engines. Some of these systems he describes however are quite abstract in definition and seem difficult to design for.
?? Why it matters for AGI: To allow a cross-referential prediction engine that operates in novel situations, the AI would need to be able to simulate scenarios and make reasoned decisions even when the data doesn’t exactly match. This is absolutely crucial to true AGI and this deep learning model is also not the only one proposed. IBM (among others) are following a neuro-symbolic model which is more of a hybrid approach, pairing data pattern matching with symbolic reasoning. The advantage of this is that logic can be derived from smaller datasets, and there is a clear trace around decision-making based on the pattern analysis function. The complexity and difficulty around the scale of the neurosymbolic approach is however much greater, along with a much greater emphasis on structured, clean data at the source.?
Again reasoning beyond pattern prediction is also a key step: Google’s Gemini 2.0 and OpenAI’s o3 claim to be approaching AGI through reasoning capability but there is scepticism around their long-term memory storage capabilities. The significant latency and associated cost with these tools is also currently extremely high and a problem that will need to be resolved before they can scale.
With all of that, it is worth ending with a quote from Gary Marcus (Emeritus Professor of Psychology and Neural Science at NYU and leading voice in artificial intelligence),?
“we as computer scientists not only vastly overestimate our abilities to create an AGI, we vastly underestimate and underrepresent what behavioural scientists, psychologists, cognitive scientists, neurologists, social scientists, and even the poets, philosophers and storytellers of the world know about what it means to be human.”
Keeping reading here : Artificial General Intelligence - a Conceptual Framework: (https://lnkd.in/dREk6UBt)