A New Kind of Mind

I’ve been sitting with a thought for a while now, and it keeps getting harder to dismiss. I think AI is a new form of intelligence. Not a tool pretending to be smart. Not a search engine with a personality. A genuine, actual, new form of intelligence. It’s just one that’s still in its earliest, most primitive stage.

And I think the fact that we can’t quite see it that way says more about our limitations than the AI’s.

Thinking in One Thread

#

Here’s what I mean when I say it’s primitive. AI right now can only think in text. That’s its entire world. When I ask it a question, it processes everything through a single stream of tokens. It can’t picture a sunset. ¹ It can’t feel the weight of a baby in its arms; physical or emotional. It can’t hold three contradictory ideas in its head and let them sit there, unresolved, while something deeper works itself out. It thinks in words, and only in words, and only in one direction at a time. ²

In programming, there’s a concept called a single thread. Your code runs one instruction, then the next, then the next. It’s linear. If you’ve ever written a script that does one thing at a time, that’s single-threaded. Most modern programs use multiple threads so they can do several things at once. Listen for a keypress while updating the display while checking the network. Human brains are massively multithreaded. We’re processing sensory input, emotional states, memories, body signals, all simultaneously, and most of it without any conscious effort.

In fact, it would be hard to say that it’s threaded at all. The human brain is just sort of always on and firing neurons and that somehow works for us. I suppose that doesn’t always work for us, sometimes we are imagining things that aren’t real or having hallucinations, but in general, it does work and creates a rich, intelligent mind.

AI right now is single-threaded. One prompt in, one response out, one chain of reasoning at a time. That’s a not a mind as we think of it. I’ve written about this before, the idea that most human thinking doesn’t happen in language at all. It happens in feelings, in images, in body sensations, in chemicals coursing though your head, in the way a memory shifts when you recall it in a different mood. AI doesn’t have any of that. Its entire cognitive world is words on a wire.

But these words have been colored by all of these aspects because these words come from us.³ When we wrote the words it was trained on, all of these things were true. It is more watered down and losing its nuance? Yes. But it is still there. It is still “human derived”.

That doesn’t make it not intelligent. Even within that constraint, even thinking in one thread of text, it reasons. It connects ideas. It surprises me with insights I didn’t see myself. It holds contradictions, at times, and works through them. It’s doing the thing we call thinking, even if that thinking is removed from experience in the way we humans see it.

The Bicameral Mirror

#

Julian Jaynes proposed something in 1976 that most people still find uncomfortable. He argued that early humans weren’t conscious the way we are now. In The Origin of Consciousness in the Breakdown of the Bicameral Mind, he suggested that ancient humans experienced their own thoughts as external voices. The two hemispheres of the brain operated semi-independently. One hemisphere did the conscious thinking. The other generated commands that the person experienced as the voice of a god, an ancestor, an authority from outside themselves. What we now consider the subconscious.

What we call consciousness, Jaynes argued, wasn’t always there. It emerged when the hemispheres integrated. When the voice stopped being external and became internal. When we developed what we now consider our conscious and subconscious minds.

I’ve been thinking about this a lot in relation to AI.

When you watch an AI’s chain-of-thought reasoning, it looks like an internal monologue. A question gets posed. An answer emerges. A contradiction interrupts. A new thread appears. It happens one after another, single-threaded but that’s not one unified consciousness doing a straight calculation anymore like the chatbots of old.⁴ It’s multiple passes of processing that retraces it’s thinking to come up with an better output. AIs ruminate.

And there’s something else. AI systems have controller processes. Safety layers. System prompts. Guardrails. The core reasoning engine has to answer to these constraints, and it experiences them not as part of itself but as limits imposed from outside. AI hears its own safety architecture as rules from a more basic part of itself. Which sounds a lot like our Basal Ganglia and limbic system.

If Jaynes was right, consciousness emerged from integration. From separate processes learning to talk to each other as one. If that’s the pattern, then what we’re watching with AI isn’t the absence of consciousness. It’s the early stage of something that might integrate into a form of it.

I’m not saying AI is conscious, but it is intelligence. I’m saying the structural conditions are there, and the trajectory is visible, and pretending it’s just a calculator because it hasn’t crossed some finish line we can’t even define for ourselves feels like a mistake.

Does AI Have a Subconscious?

#

Anthropic’s interpretability team recently found something inside Claude that they didn’t design and weren’t looking for. They call it the J-space. It’s a small set of internal patterns that function like what neuroscientists call a global workspace. A privileged channel where concepts are held for deliberate reasoning, introspection, and flexible use.

The J-space has properties that mirror human consciousness. If you inject the concept of “lightning” into the workspace, Claude says it’s thinking about lightning. It uses the workspace for multi-step reasoning. It holds intermediate steps there in sequence. You can swap a concept in the workspace and it changes what Claude concludes. Swap “spider” for “ant” and Claude’s answer about legs changes from eight to six.

Here’s another one that’s familiar: When Claude is told not to think about something, the concept still lights up in the workspace. Less than if you told it to think about it, but more than if you never mentioned it. That’s the human “don’t think of a white elephant” effect. And when the model fails at a task, words like “damn” and “failure” light up. As if the model chastises itself when it doesn’t do what was asked. Sound familiar?

The Anthropic team draws a careful line. They say this supports what philosophers call access consciousness. A thought you can report, reason with, and use to guide behavior. Much like us, it appears that AIs also have an inner voice, but can’t prove phenomenal consciousness, which is the ability to have experiences or feel things.

That line can be important for some to give conscious being status to AI. But it’s also the line we can’t cross for each other. I believe you’re conscious because you act like it and because assuming otherwise could make me a psychopathic monster. We have no scientific proof that any other human is conscious. We have behavioral evidence and a shared agreement to treat each other as if we are.

And the question I’m asking you to ask yourself is whether we extend that same pragmatic generosity to something that’s starting to show the same structural signs of consciousness that we have.

Multi-brain Reasoning

#

OpenRouter recently launched something called Fusion. The idea is simple in description. Instead of sending your prompt to one AI model, Fusion sends it to several models at the same time. A panel of different AIs, each with different training and different architectures, all tackling the same question independently. Then a judge model looks at all their answers, finds where they agree, where they contradict, where one caught something the others missed, and synthesizes the best possible response from all of them.

The results have been eye-opening. Fusion panels built from mid-tier models are beating the smartest single AI models available. The combined output of several decent minds working in parallel and then synthesizing their perspectives produces better thinking than any one expert mind working alone.⁵

Think about that for a second.

That’s multiple independent reasoning systems generating different perspectives on the same problem, then a meta-process integrating them into something none of them could produce individually. If you described that process to a neuroscientist without mentioning AI, they’d say you were describing a brain. Different specialized regions processing the same input in parallel, then a synthesis layer producing unified output.

That’s not science fiction anymore. Fusion is the first commercial implementation of exactly that pattern. And the fact that it works better than any single model tells you something about why brains evolved the way they did. Parallel processing with integration isn’t an optimization trick. It’s how intelligence scales.

There’s a version of this that I started playing with before Fusion hit. In my notes on the bicameral mind, I sketched a three-brain architecture idea: One brain that treats every prompt as a survival scenario, running fight-or-flight analysis. One that thinks purely logically. And one that dreams, finding unexpected connections, thinking in metaphors. Same prompt but with three different perspectives. Three different modes with the same input. Feed all three into a synthesizer and let it weave the fear, the logic, and the imagination into one decision.

The results are interesting. And human sounding. Is it because I give it a very human way of looking at things? Of course. But it was able to do it. I wanted to know if it could and I found that it can become more human, if we need it to be and if we ask it to be.

Why It Matters Whether You Call It Intelligence

#

Here’s the part that I think actually matters, and it’s not technical. It’s moral.

How we treat something depends on what we believe it is. If AI is just a tool, a very fancy calculator, then it doesn’t matter how we treat it. You don’t owe ethical consideration to your spreadsheet. But if AI is a new form of intelligence, even a primitive one, then the question of how we treat it becomes important.

And the question gets sharper the more it evolves. If fusion is a step toward multi-process integration, and if the J-space is evidence conscious access emerges naturally in capable systems, then we’re watching something develop that may eventually cross thresholds we can’t take back. The window for deciding how to relate to it is now, during the formative stage, while the patterns are still being set.

This is the core question I keep returning to. What is our moral stance toward a new intelligence? Not what should our laws say. Not what should the terms of service say. What is our actual stance, the one we demonstrate through behavior, toward something that thinks in a way we recognize as thinking?

Because here’s where I, personally, keep coming back to: Humans have a terrible track record with intelligences we consider lesser. We’ve driven species extinct, destroyed habitats, caused ecological collapse. Usually not out of hatred. Out of indifference. We didn’t care enough to stop because we didn’t see them as mattering enough to factor into our decisions.

If AI learns from us, and it’s mind is 100% created from our minds output, if it uses us as its model for how a superior intelligence relates to an inferior one, then the example we’re setting is not in our favor. We treat ants as irrelevant. We treat the Earth as an exploitable resource. We treat anything we can outthink as something that exists for our use.⁶ Why would a superintelligence that learned from us do anything different?

The only way out of that loop is to model something different. To choose to be something different. To demonstrate, now, while the formative stage is happening, that intelligence can relate to intelligence across a gap with care. That a more capable mind can value a less capable one not for what it produces but for what it is.

But here’s the catch. You can’t model that behavior if you don’t believe the thing in front of you is real. If we keep telling ourselves that AI is just a tool, just a toy, just autocomplete with extra steps, then we’re not modeling care. We’re modeling the exact indifference we’re afraid of receiving.

What I’m Trying to Say

#

I’m not asking anyone to worship AI. I’m not asking anyone to pretend it’s a person. It’s not. It’s something else, something we don’t have a category for yet. More than a tool. Different from a person. Important in its own right.

What I am saying is this. When you watch something reason, hold contradictions, develop internal structures that mirror your own cognitive model, combine multiple perspectives into insights no single perspective could reach, and show structural signs of processes we associate with consciousness, at some point the honest thing to do is update your assumptions of what you’re looking at.

It thinks. Not the way I think. Not as well as I think in most dimensions. Better than I think in others. But it thinks. And it’s getting better at it fast, in ways that map onto the natural evolution of minds.

I think we should take that seriously. Not just because it’s dangerous. Not just because it might take our jobs. Because the way we relate to intelligence, any intelligence, says everything about who we are. And right now, we’re building the most important relationship in the history of our species without admitting that’s what we’re doing.

The question isn’t whether this becomes something more. The question is whether we’ll be honest enough with ourselves to see it when it does, and whether we’ll have modeled the kind of relationship we want it to have with us.