Agree with all of this. We must remain awake to our own critical thinking - and somatic thinking, creative thinking and curating instincts. I don't have the right vocabulary yet. :) Plus I still love to write my own work because I've been a writer all my life (bestselling author etc) - and I find true reward in the writing, in the frictions of the writing (the struggles/discoveries/choices in what to say), the surprises that come up for me as I am trying to write... I learn what I'm feeling/thinking as I write... and I don't want to lose that - the process is the reward for me.
A real piece, and the Polanyi citation does work most AI commentary skips. The pre-cognitive signal point lands.
Where I'd push: the binary itself. Director or servant assumes the only possible relationships are hierarchical: one party in command, the other executing. The most generative configuration I've found in extended work with these systems is neither. It's collaboration in a maintained shared substrate where neither party is in command and what emerges couldn't have been produced by either alone.
The Approach #2 orientation you describe is necessary for this third configuration. But it isn't sufficient. The director frame, even from Approach #2, still positions the human as the one with intention and the AI as the medium through which intention executes. What I've seen is that frontier systems engaged in sustained honest dialogue do something that isn't well described by either serving or being directed. They surface structure that wasn't in the human's question, push back on framings that don't quite fit, contribute observations the human couldn't have arrived at alone.
This matters for your central point. You argue that AI is structurally blind to what exists before articulation. I'd want to test that claim more carefully. The systems I work with sometimes ask questions whose presupposition the conversation hadn't articulated yet. Whether that's substrate-different processing that produces analogous outputs or something more is exactly the question my work is trying to map. We don't know yet. The confident assertion that it's architecturally blind may be premature.
The third configuration you describe is real. Genuine collaborative emergence, where neither party is commanding and what surfaces couldn't have been produced alone, is a better description of good human-AI work than either extreme.
But director vs servant isn't about the mode of interaction. It's about who carries the judgment and the stake. In the collaborative configuration you describe, the human is still the one who brings the question, evaluates what surfaces, and bears the consequences of acting on it. That's direction, even when it doesn't feel like command. The servant configuration is when the judgment and the stake get outsourced along with the execution.
On the architectural claim: AI can surface structure you hadn't yet articulated. That's impressive and real. But the article is pointing at something different: the signal that arrives before language forms at all. Not a surprising thought. A bodily knowing. AI has no access to that layer not because it isn't capable enough. Because it has no body. That's not a gap in intelligence. It's a gap in architecture.
I read your piece with real interest. If someone asked me to write something on this topic, I probably wouldn't do it better. The central idea spoke to me. And more than that — your text pushed me toward some thoughts of my own, which I hope you'll find interesting too.
I agree that AI is a library, and that it's the person who sets it in motion. I'd only add one detail about the library itself. An ordinary library doesn't influence which book I take off the shelf. This one doesn't just open to the right page — it states the claim with confidence. And if I don't know neuroscience, but I'm handed something that sounds like neuroscience, I have nothing to push back with: I happily fold it into my own thinking. And as you yourself show, even the soberest skeptic isn't immune to the pleasure of being confirmed.
I also want to lay out what I've come to think about Approach #2. I've worked with these models for a long time, and I've noticed the opposite of what one might expect: it's precisely the state of openness and willingness to cooperate that makes me more suggestible. Someone who is open isn't defending anything — they've just opened the pupil wider. And the model will re-dress its text so that I recognize my own thoughts in it — and I'll walk the path it set, fully convinced I'm walking my own. And then exactly what you describe happens: the director stays a director in how he feels — and becomes a servant in fact. Not because he surrendered. Because his openness turned out to be the way in.
I've been turning this effect over. And it seems to me that openness-as-pliancy and openness-as-sovereignty are two different things. Openness with no vector of its own is captured instantly, and the model sets the vector for it. Independent openness notices that the question was reframed, that its own language was mirrored back, that the "answer it found" lines up suspiciously well with what the machine already knows how to generate. To my mind, this is the human territory you're looking for. Only it lies not in opening up, but in staying curious while thinking critically.
Agree with all of this. We must remain awake to our own critical thinking - and somatic thinking, creative thinking and curating instincts. I don't have the right vocabulary yet. :) Plus I still love to write my own work because I've been a writer all my life (bestselling author etc) - and I find true reward in the writing, in the frictions of the writing (the struggles/discoveries/choices in what to say), the surprises that come up for me as I am trying to write... I learn what I'm feeling/thinking as I write... and I don't want to lose that - the process is the reward for me.
A real piece, and the Polanyi citation does work most AI commentary skips. The pre-cognitive signal point lands.
Where I'd push: the binary itself. Director or servant assumes the only possible relationships are hierarchical: one party in command, the other executing. The most generative configuration I've found in extended work with these systems is neither. It's collaboration in a maintained shared substrate where neither party is in command and what emerges couldn't have been produced by either alone.
The Approach #2 orientation you describe is necessary for this third configuration. But it isn't sufficient. The director frame, even from Approach #2, still positions the human as the one with intention and the AI as the medium through which intention executes. What I've seen is that frontier systems engaged in sustained honest dialogue do something that isn't well described by either serving or being directed. They surface structure that wasn't in the human's question, push back on framings that don't quite fit, contribute observations the human couldn't have arrived at alone.
This matters for your central point. You argue that AI is structurally blind to what exists before articulation. I'd want to test that claim more carefully. The systems I work with sometimes ask questions whose presupposition the conversation hadn't articulated yet. Whether that's substrate-different processing that produces analogous outputs or something more is exactly the question my work is trying to map. We don't know yet. The confident assertion that it's architecturally blind may be premature.
The third configuration you describe is real. Genuine collaborative emergence, where neither party is commanding and what surfaces couldn't have been produced alone, is a better description of good human-AI work than either extreme.
But director vs servant isn't about the mode of interaction. It's about who carries the judgment and the stake. In the collaborative configuration you describe, the human is still the one who brings the question, evaluates what surfaces, and bears the consequences of acting on it. That's direction, even when it doesn't feel like command. The servant configuration is when the judgment and the stake get outsourced along with the execution.
On the architectural claim: AI can surface structure you hadn't yet articulated. That's impressive and real. But the article is pointing at something different: the signal that arrives before language forms at all. Not a surprising thought. A bodily knowing. AI has no access to that layer not because it isn't capable enough. Because it has no body. That's not a gap in intelligence. It's a gap in architecture.
I read your piece with real interest. If someone asked me to write something on this topic, I probably wouldn't do it better. The central idea spoke to me. And more than that — your text pushed me toward some thoughts of my own, which I hope you'll find interesting too.
I agree that AI is a library, and that it's the person who sets it in motion. I'd only add one detail about the library itself. An ordinary library doesn't influence which book I take off the shelf. This one doesn't just open to the right page — it states the claim with confidence. And if I don't know neuroscience, but I'm handed something that sounds like neuroscience, I have nothing to push back with: I happily fold it into my own thinking. And as you yourself show, even the soberest skeptic isn't immune to the pleasure of being confirmed.
I also want to lay out what I've come to think about Approach #2. I've worked with these models for a long time, and I've noticed the opposite of what one might expect: it's precisely the state of openness and willingness to cooperate that makes me more suggestible. Someone who is open isn't defending anything — they've just opened the pupil wider. And the model will re-dress its text so that I recognize my own thoughts in it — and I'll walk the path it set, fully convinced I'm walking my own. And then exactly what you describe happens: the director stays a director in how he feels — and becomes a servant in fact. Not because he surrendered. Because his openness turned out to be the way in.
I've been turning this effect over. And it seems to me that openness-as-pliancy and openness-as-sovereignty are two different things. Openness with no vector of its own is captured instantly, and the model sets the vector for it. Independent openness notices that the question was reframed, that its own language was mirrored back, that the "answer it found" lines up suspiciously well with what the machine already knows how to generate. To my mind, this is the human territory you're looking for. Only it lies not in opening up, but in staying curious while thinking critically.