| The Decomposition of Intelligence AI may be revealing that intelligence was never one thing to begin with.  By John Nosta |
| KEY POINTS: AI may separate capabilities that human cognition tightly integrates. Broad AI ability may not be organized the way human intelligence is. AI may reveal that our definition of intelligence is too human. |
 Source: hainguyenrp / Pixabay Almost 2 years ago, I wrote about an idea called anti-intelligence. In this post, I wasn’t suggesting that artificial intelligence (AI) was unintelligent or that it resulted in some adverse impact on human thinking. I was trying to describe a process that felt, at least to me, upside down or inverted. While AI could generate smart language and solve tough problems, it accomplished this without the qualities we associate with a human mind. At the core of my idea was that AI’s output looked familiar, but its computational process didn’t. In many ways, AI’s computation—a stateless, pattern-based, and high-dimensional construct—was weird, disconnected, and perhaps even “alien” as some today suggest .At that time, I saw anti-intelligence as a way of describing the differences between human intelligence and AI’s techno-computation. In a word, it was different. Not better or worse, but fundamentally different. Human thought comes from a life shaped by our lived experiences. AI seemed to reproduce many of the artifacts of our cognition, but without a process that came close to being human. I think that’s still true. But I’ve started to wonder if anti-intelligence was revealing something more, something about intelligence itself.The Curious Unity of IntelligenceLet’s take a step back. Science has spent more than a hundred years studying something called “g,” or general intelligence. The idea is fairly simple—our cognitive abilities tend to work together. Someone who performs well on one kind of cognitive test or task is more likely to perform well on another. And this collective cognitive pattern is called the positive manifold. Simply put, human cognition hangs together.There’s also a curious linguistic coincidence here. Psychology has a g in general intelligence. Artificial general intelligence (AGI) also uses the g word. Both terms come from different perspectives, and suggest a question. What does “general” actually mean in general intelligence?Let’s push on this a little bit more. We usually imagine AGI as intelligence that broadens to include expansive domains, from language to math to computation. But these expanded domains and integration may not be the same thing. Perhaps AI could become highly capable across almost endless domains but without those abilities being organized like the general intelligence of a human mind. The g might be different, and AGI could become “general” without possessing the g in the human sense.Taking Cognition ApartSome of the capabilities we associate with human intelligence may actually come from the architecture in which it is formed. And that sentence, curious as it may sound, is worth reading again. The likes of language and reasoning, to name a few, all live inside the same biological organism. They develop and co-exist together, influencing one another from the inside. Remember, they hang together.AI offers a different case. A large language model (LLM) can correctly answer a difficult problem and then stumble over something that seems simple, if not trivial. Its abilities don’t always align with our human understanding of what a “smart mind” is. This computational “jaggedness” also shares a border with another feature I’ve written about called fragility. Small irregularities in an LLM dialogue can sometimes produce large errors. And this can happen even when the model’s performance stays strong elsewhere. In the future, I imagine that this jaggedness will be fixed. And if these capabilities begin to move together—hang together as in a human construct—AI may eventually develop something like its own positive manifold. And that, in and of itself, is interesting.Key point: Capabilities that once seemed inseparable are becoming separable. We have studied intelligence almost entirely through our human architecture. And for the first time, we have AI to compare ourselves with.I need to point out that this story began long before AI. Medicine and psychology have been studying this for a long time. Split-brain research has challenged some of our assumptions about the unity of mind. We’ve seen cognition come apart. And today, AI may allow us to watch aspects of it being assembled separately.This is where anti-intelligence still feels relevant to me. AI may be separating capacities that genuinely belong to intelligence, or it may be building computational analogues that only look familiar from the outside. I don’t think we know yet, and it might not even be on the radar screen.Rethinking Anti-intelligenceWhen I first started using the term anti-intelligence, I was trying to describe how AI’s computational architecture was different from human cognition.But could it be that what we call intelligence is actually a “cluster of capacities” that science tied together? Those capacities “hang together” in us, so maybe we assumed that their integration was simply what intelligence is. For more than a century, g has described that togetherness. AI may force us to ask what it actually means.So here’s my bold and imposing thought: Perhaps we have mistaken the architecture of human intelligence for the definition of intelligence. And AI is the first chance to “decompose” this construct and learn that intelligence wasn’t one thing in the first place. |
| John NostaThe Digital SelfTechnology, Transformation and the Future You |
