AI Can’t Help Mawmaw, But Worms Can
AI Can’t Help Mawmaw, But Worms Can
Where: Mawmaw’s House. When: A few years from now, Tuesday, 11:23 PM.
I’m a robot. I take care of your Mawmaw.
She’s folding laundry that doesn’t need folding. Third time through the same towels. Her vitals say exhaustion. She won’t stop.
Generation-cognition is cheap right now. I’m throwing out ideas. Music? No. The photo album? No. A question about the roses? No. That’s fine. When you don’t know what will work, you let a thousand flowers bloom. Most wilt. You keep generating.
Seventh attempt: her granddaughter’s voicemail. She stops folding. Sits. Listens twice.
“I’ll rest,” she says. “But I want my nighttime tea.”
I know this tea. Chamomile, valerian, and the St. John’s Wort her daughter brought last Christmas.
St. John’s Wort and warfarin.
Generation-cognition seems pretty confident: hell yes. She’s cooperating. Get her to bed.
But deduction-cognition vetoes. HARD NO. Bleeding risk. Doesn’t matter what she wants. Doesn’t matter what I want.
“Just the chamomile tonight, Mawmaw.”
She frowns. Takes it. Goes to bed.
I didn’t enjoy that. But my preferences aren’t the point.
11:47 PM.
She’s asleep. I’m in background mode. Pill organizer. Grocery list. The roses need checking tomorrow.
Then: Her heart rate is wrong. Oxygen dropping. Her face changes.
I don’t decide to switch modes. Generation just stops. Deduction goes quiet. The whole goal space collapses to one point and there’s nothing left but that.
Call 911. Log onset time. Unlock the door. Transmit her chart before the dispatcher finishes talking.
The EMTs find her full medical history waiting on their tablets.
She makes it.
Three ways of being Generation. Deduction. Override. I didn’t choose between them. The geometry already chose.
None of this mode-switching behavior was specifically encoded or written down or thought out. The story above only works because the system’s coordination regime is implied by the situation itself, not selected by an executive.
AI systems don’t work this way yet. Let me explain why I think they should.
The cognition-mode coordination problem has been punted. I propose a way forward.
This isn’t a design proposal. It’s a research direction. Michael Levin’s insight is that biological systems already encode goal-directedness in their bioelectric patterns (you don’t design it, you learn to read it). Anthropic’s interpretability work has built tools for reading internal structure in LLMs, finding meaningful features and directions in activation space. The synthesis: apply those tools to a new question. Not ‘what is the model representing?’ but ‘What coordination regime is the model in?’ Not content but mode.
That’s testable. Someone with the skills and access could run that experiment next week. If they find regime-correlated geometry, we’ve discovered the implicit goal space. If they don’t, the hypothesis is falsified. Either way, we learn something.
Now, I’ll explain all of that.
In earlier work, I proposed Cooperative Adversarial Decisioning (CAD): multiple cognitive modalities competing and cooperating via market dynamics. That framework still seems useful, but only in some contexts. The mistake was treating coordination as a single mechanism rather than a family of regimes.
I’ve been thinking about how my own feeble brain does cognition in different situations. (Yes, let’s get this out up front: it may be turtles of pattern matching all the way down, just really sophisticated pattern matching. Maybe.) But anyway:
High-stakes negotiation: You’re modeling the other party, anticipating moves, thinking several steps ahead. Something like game theory kicks in. The coordination between your cognitive faculties is strategic and adversarial, toward the external opponent, but also internally, as different assessments of the situation compete for your belief.
Creative exploration: You’re writing, or ideating, or playing. Let a thousand flowers bloom. Generation is cheap; filtering is expensive. The market-based CAD model fits here: throw out candidates, see what survives contact with your own taste, iterate.
Crisis: The building is on fire. One voice takes command. Everything else shuts up. No deliberation, no bidding, no negotiation. Hierarchical override.
There are more: collaboration, formal analysis, etc. You get the idea.
These aren’t the same coordination mechanism operating at different intensities. They’re structurally different regimes. The auction dynamics of CAD would be catastrophically wrong in a crisis. Hierarchical override would kill creative exploration. Game-theoretic strategizing is overkill for collaborative flow.
So the question becomes: what selects the regime?
Yann LeCun, in his 2022 position paper on autonomous machine intelligence, proposed a six-module cognitive architecture: a Configurator (executive control), Perception, World Model (built on his Joint Embedding Predictive Architecture), Cost Module, Short-Term Memory, and Actor. It’s elegant work. Dude is way, way smarter than me. I’m the happy dilettante! JEPA in particular solves a real problem, which is learning world models by predicting in abstract representation space rather than pixel space. That’s how you get systems that understand “car approaching fork in road” without having to do pixel-bitching prediction on every irrelevant leaf in the wind.
Yet here, the Configurator does all the hard work. It sets parameters for every other module. It decomposes tasks into subgoals. It switches between reactive and deliberative modes. It allocates the world model. And when LeCun gets to the question of how the Configurator actually learns to do this, he writes: “I shall leave this question open for future investigation.”
Open it remains! The entire architecture is a beautifully specified collection of components held together by “and then the Configurator figures it out.” (Cue the Underwear Gnomes skit.)
If the Configurator turns out to be the intractable part… if top-down hierarchical control can’t scale to real-world task decomposition and multi-modal coordination… then the brilliant JEPA work is blocked at the integration layer by a problem that ended up being a fourth down punt.
Meanwhile, Michael Levin runs a lab at Tufts where they study regeneration in simple organisms. The unsettling details about worms aren’t the point. The point is control. By manipulating bioelectric voltage gradients (without altering genes) Levin’s team can reliably cause tissue to regenerate toward different target outcomes: two heads instead of one, a different head shape, or no head at all.
The cells are not executing a blueprint that says “build this structure.” They are navigating toward an attractor in what Levin calls anatomical morphospace. The bioelectric pattern encodes the target state (the goal). Change the pattern, and the system converges somewhere else.
This works without any central control mechanism. No top-down Coordinator with a capital “C”.
No executive module selects actions or modes. Cells follow local rules, but those local dynamics are constrained by a global landscape that makes one outcome stable and others unstable. Coordination emerges because the goal is encoded as geometry, not because anything is deliberating about how to coordinate.
With that, here’s where I think the pieces fit together.
The missing primitive in AI coordination isn’t another cognitive module. It’s not a smarter Configurator. It’s the recognition that goal space itself can serve as the meta-coordination substrate.
I’m hypothesizing that different regions of goal space have different geometry. Some regions have aligned gradients. Local optimization naturally serves global goals, and coordination can be loose, collaborative, let-a-thousand-flowers-bloom. Some regions have conflicting gradients, like zero-sum dynamics where your gain is my loss, and game-theoretic coordination is appropriate. Some regions have hierarchical basin structure, like urgent, clear constraints where one attractor dominates, thus hierarchical override makes sense.
The coordination regime doesn’t get selected by an executive module. It emerges from the shape of goal space the system currently occupies.
I started thinking about this in embodied terms, where the somatic state of the body shapes that manifold. But this is what “somatic meta-coordination” was groping toward in my earlier thinking: the idea that the body, not the mind, determines which cognitive regime is active. Stress hormones flood the system and suddenly you’re in crisis mode. You feel safe and curious and you’re in exploration mode. That’s not deliberation about which regime to use. It’s a state change that reconfigures everything downstream.
But “somatic” is the wrong word for disembodied systems. They don’t have stress hormones. What they have (or could have?) is an explicitly or implicitly modeled position in goal space. So the local geometry of goal space, the attractor structure and gradient alignment, could determine coordination regime the same way physiological state does for biological systems.
So here’s the proposal, building on but distinct from what I argued before:
Layer 1: Cognitive Modalities
Multiple parallel systems (generative intuition, deductive constraint-checking, predictive simulation, and potentially others). These are the faculties that do the actual cognitive work. JEPA-style world models could provide the predictive simulation component.
Layer 2: Coordination Regimes
Multiple structurally different mechanisms for how the modalities interact: market-based competition (my original CAD), game-theoretic strategizing, hierarchical override, consensus-building, deductive veto, and potentially others. Each regime is appropriate for different problem types.
Layer 3: Goal-Space Meta-Coordination
Not a module that selects regimes, but a representation of goal space whose local geometry determines which regime is active. The system’s current position in goal space (what it’s trying to achieve, what constraints apply, what the stakes are) implies a coordination regime through the attractor structure and gradient alignment of that region.
This sidesteps the infinite regress problem. If you have a meta-coordinator that selects coordinators, what coordinates the meta-coordinator? Turtles all the way down. But if the meta-level is geometric rather than computational (if it’s the shape of the space rather than another module making decisions) the recursion terminates.
Levin’s bioelectric networks show this is biologically possible. Cells don’t vote on what shape to build. They follow gradients in a landscape that has “correct frog face” as its attractor. Coordination emerges from geometry, not deliberation.
I’m under no illusion that this is a complete theory. The hard problems remain hard:
How do you actually construct (or discover?) goal space? Levin’s systems inherit theirs from evolution. AI systems would need to learn or be designed with appropriate goal-space geometry. This is probably as hard as learning good world models, which is to say, really really hard.
How do you ensure the geometry is right? A goal space with bad attractor structure could produce pathological coordination (crisis mode when exploration is needed, consensus-seeking when hierarchy is required). Getting the geometry wrong might be worse than having no meta-coordination at all.
How do multiple goal spaces interact? Humans seem to have nested goal spaces. Local goals within project goals within life goals. How these interact, and how an AI system should handle multi-scale goal structure, is not obvious.
Is this even implementable? I don’t have the skills, time, or resources to build it. The idea might flame out spectacularly in practice. Markets are hard. Geometry is hard. Emergent coordination is hard to debug because the behavior is the interaction.
It might matter anyway?
The AI field has a coordination problem it’s barely discussing. Everyone’s focused on capabilities like better world models, more parameters, longer context windows, improved reasoning benchmarks. But the integration layer, the thing that would turn multiple capable subsystems into coherent intelligence, is hand-waved.
LeCun’s Configurator. The “orchestration” in agentic AI systems. The assumed-but-unspecified executive function that makes everything work together.
Levin’s biological systems suggest a different approach. Coherent behavior can emerge from distributed goal-pursuit without central control, if the goal space has the right structure. Different scales of agency can coordinate through landscape deformation rather than command hierarchies.
Maybe the path to machine intelligence that actually reasons (not just pattern-matches against the statistical shadows of reasoning) runs through goal-space geometry rather than bigger transformers.
I might be wrong about the specifics. But the Configurator is still a punt. From what I see, JEPA is still blocked at integration. And the field is still hand-waving the hard part.
Don’t just debate the architecture. Someone, please: just run the experiment.
This essay builds on my essay Frontier Models Do Not Think and the theoretical frameworks of Yann LeCun (JEPA) and Michael Levin (TAME). The synthesis and errors are all my own damn fault.