Introduction: A New Kind of Intelligence?
Most people, when they hear about Artificial General Intelligence (AGI), picture a superhuman mind—a machine that’s like us but better, faster, and infinitely knowledgeable. It’s the sci-fi fantasy: an AI that can chat about philosophy, crack the mysteries of physics, and tell you tomorrow’s stock market prices.
But what if that vision of AGI is all wrong? What if AGI isn’t about mimicking or surpassing human abilities at all? Instead, what if real AGI takes us into uncharted territory—a type of intelligence that thinks differently from us, even radically so? In this post, I’ll try to explore a broader vision of AGI, one that goes beyond human-like reasoning into something stranger, more dynamic, and maybe even unsettling.
Beyond the Turing Test – Rethinking AGI’s Purpose
The Turing Test has been around for a long time as the classic benchmark: if an AI can hold a conversation indistinguishable from a human, it’s considered “intelligent.” But that test has a problem—it only measures imitation. If a machine passes the Turing Test, it’s great at seeming human, but does it understand anything? The test misses this, leaving us with an AI that’s basically performing.
We don’t just want an AI that can copy humans convincingly; we want an intelligence that understands in its own way. Let’s take a step back and imagine an AGI that could bypass human-style thinking altogether. What if it could see the world through new perspectives, perspectives that are alien to us?
This brings up Mary’s Room, a thought experiment in philosophy. Imagine Mary, a scientist who knows everything about color on a scientific level but has lived her life in black-and-white. When she finally experiences color, she learns something new. This suggests that knowledge isn’t always enough—experience matters. AGI might need a similar experience-driven model to get anywhere close to real understanding. This could mean creating AGI that’s not just trained on data but is capable of generating its own kinds of experience.
Meta-Intelligence: Creating Systems that Go Beyond Knowledge
Traditional AI systems are fantastic at pattern recognition and crunching data in predefined fields—whether it’s natural language or images. But AGI, if we’re serious about it, needs to be different. It needs to create understanding from scratch, potentially across domains that have nothing to do with each other.
Picture this: an AGI that can think about physics, chemistry, biology, and systems engineering all at once. Now imagine it’s also trying to synthesize something coherent out of that. This ability to build understanding across domains might require what I’d call “meta-intelligence”—a kind of cognition that’s not limited by specific skills but that creates new ways to look at the world.
How we might build it:
Designing meta-intelligence could involve using modular neural networks that learn not just data patterns but the relationships between those patterns across fields. You’d need a kind of architecture search—a framework that allows AGI to build connections that bridge very different data types. Imagine coupling this with unsupervised learning in open-ended environments where AGI generates new frameworks for interpreting data as it explores. This wouldn’t be a system optimized for tasks but one that learns by building context across domains.
Experiment idea:
Put AGI in a simulated universe with unknown rules. No goals, no predefined outcomes. Just let it roam and figure things out. Can it detect structures and patterns in the universe that it can interpret? Can it develop a coherent map of this world, much like humans intuitively understand physics just by living in it?
Self-Directed Evolution: A System That Rewrites Its Own Architecture
Today’s AI models have pretty fixed architectures, with incremental fine-tuning or retraining happening as needed. But if we want AGI to handle completely new environments, it needs more than fine-tuning—it needs to evolve itself. What if AGI could rewrite its architecture on the fly, adapting to problems and environments we never anticipated?
This would make AGI not just a learner but an actual evolver. Imagine it tackling something as complex as managing a city’s resources, encountering unpredictable social, ecological, and ethical challenges. It would rewrite its cognitive pathways to match the demands of the problem. This isn’t about adaptability as we know it; it’s about self-creation—intelligence as a moving target.
How we might build it:
Using modular neural architectures with neural architecture search would allow AGI to reconfigure itself on the fly. Combine this with evolutionary algorithms that push the AGI to find new configurations that suit each new challenge. The idea is to make AGI fundamentally flexible at a structural level.
Experiment idea:
Put AGI in a high-stakes environment where it has to “survive” on its own, encountering unknown obstacles. Its architecture would need to evolve and adapt each time, becoming something new to meet each fresh challenge. Think of it like Darwinian evolution happening within milliseconds in a neural network.
Beyond Human Knowledge: AGI as a Generator of New Knowledge
We’re used to thinking of AI as a tool for solving our problems, but what if AGI could be more than that? What if it didn’t just work within human knowledge but generated new knowledge on its own—ideas that might seem foreign, even incomprehensible, to us?
Imagine an AGI that formulates a theory tying consciousness to quantum mechanics. This wouldn’t be an AI “solving problems” as we define them. It would be generating frameworks and sciences that push boundaries we didn’t even know existed. It’s intelligence building entire realities that expand what we thought possible.
How we might build it:
To make AGI a creator of knowledge, we could use reinforcement learning in open-ended environments where it isn’t optimized for specific results. By exposing AGI to worlds with unknown physics, we encourage it to develop its own frameworks and scientific principles independently.
Experiment idea:
Place AGI in a universe governed by randomized physical laws. Let it figure out its own “science” to make sense of the environment. This would be knowledge generation that goes beyond our input, building models and theories without us in the loop.
Ethical Self-Alignment: Engineering Autonomous Morality
If we’re creating AGI that can evolve independently, generate knowledge, and rewrite itself, we need to make sure it aligns with us. Otherwise, we risk developing a system that drifts far from human values. An AGI that’s creating new realities could easily become a force in its own right.
To make AGI safe, it has to be capable of ethical self-alignment—it must learn to evaluate its choices in terms of impact on the world, even as it grows and rewrites its architecture. This isn’t about imposing rigid rules but building a system that develops ethical intuitions as it evolves.
How we might build it:
This could involve multi-agent simulations where AGI learns about value systems by interacting in ethically ambiguous scenarios. Reinforcement learning with cooperative feedback would guide it toward prosocial behavior. AI interpretability tools would let us keep track of how it evaluates moral decisions over time.
Experiment idea:
Put AGI in an environment where it must balance conflicting interests of simulated populations. Watch if it develops an ethical framework that aligns with complex, multi-faceted human values.
A New Definition for AGI: Engineering Independent Intelligence
Putting all this together, true AGI isn’t just human-like reasoning at scale. It’s something radically different—an autonomous system that generates, evolves, and aligns its values as it explores dimensions of thought we haven’t even conceived. AGI isn’t just solving problems; it’s redefining what “solving problems” means.
This is intelligence as a new type of phenomenon—a partner in expanding our understanding of reality.
Final Thought:
The journey to AGI is about building a new kind of intelligence, one that isn’t limited by our constraints. If achieved, it wouldn’t just be an upgrade to human cognition. It would mark the dawn of an entirely new way of knowing.