I have spent much of my adult life trying to understand what thought is in mechanistic terms. Not metaphorically, and not as a list of mental faculties, but as a process that could in principle be rebuilt. That project led me to two closely related claims. First, working memory is not simply a storage buffer. It is a continuously updated computational workspace whose contents overlap from moment to moment, and that overlap is what gives the stream of consciousness its felt continuity. Second, once you take that overlap seriously, “thinking” starts to look like a particular kind of search: a multi-cue, constraint-limited retrieval process in which several coactive items jointly recruit the next update.
This article is about how I backed into that second claim through a path that was not elegant at all. It started with small mental mistakes. The kind you notice and laugh at because they are obviously wrong, but only after they have already occurred. Over years, these little derailments became data. They revealed a pattern. When my mind was juggling several concepts at once, a cheap or superficial link could sometimes hijack the next association. The error usually snapped back immediately, but the moment itself was informative. It suggested that associations are often statistical attractor choices within an active set, rather than deliberate, logical steps selected in isolation. And it suggested that the same mechanism that enables rapid insight and creative pattern completion is also a mechanism that can drift when key constraints drop out of the workspace.
My goal here is to make this phenomenon clear and useful. I will describe the kinds of misfires that led me to formulate multiassociative search, and then I will connect them to a broader architecture of iterative updating. The broader ambition is practical: if we can specify the computational shape of these transitions, we can begin to design artificial systems that do not merely generate outputs, but that maintain context, test coherence, and refine their internal states over time in a brainlike way.
Readers who want the deeper foundation for iterative updating and its relevance to machine consciousness can find my related work at aithought.com and in the two papers cited below.
Reser, J. 2022. A Cognitive Architecture for Machine Consciousness and Artificial Superintelligence: Updating Working Memory Iteratively. arXiv: 2203.17255
Reser, J. 2016. Incremental change in the set of coactive cortical assemblies enables mental continuity. Physiology and Behavior. 167: 222-237
1. The phenomenon you can catch in yourself
Most of us are familiar with the big categories of cognitive error: forgetting a name, walking into a room and losing the reason, misplacing an object that was in our hand a moment ago. Those are obvious failures. What interested me more, and what eventually became the seed of a theory, were the smaller errors that are almost too fast to count as errors. They last a fraction of a second. They are not “beliefs” in any stable sense. They are brief candidate interpretations or candidate responses that win the competition for the next thought, and then immediately lose.
If you have ever had a moment where an absurd inference pops into mind and you laugh at it, you already know the kind of event I mean. You are not confused for minutes. You are not delusional. You do not even feel uncertain. You simply notice that the mind proposed something ridiculous, and then you watch yourself veto it. The experience is revealing because it makes the selection process visible. It lets you see, in miniature, how the mind chooses what comes next.
Over the years I began to notice that these slips have a characteristic structure. They are almost never random. They tend to occur when several concepts are active at once, when the mind is juggling. In that condition, a coincidental overlap between items can become the bridge to the next association, even when it has no legitimate causal standing. A cheap link wins because it is available, not because it is true. Then, if the system is healthy and the constraints are still online, the error snaps back quickly.
That snap-back is important. It suggests that there is a real evaluative process, a coherence check, that can reject a candidate despite the candidate being statistically plausible within the current activation landscape. It also suggests something more unsettling. If I can see the absurd candidates that get vetoed, then there are surely other candidates that do not get vetoed, candidates that drift by quietly because they are not absurd enough to trigger the alarm. The phenomenon is therefore not just a curiosity. It is a window into how thought can derail in subtle ways, and how the system normally protects itself from doing so.
Fig. 1. A Schematic for Multiassociative Search
Spreading activity from each of the neural assemblies (lowercase letters) of the four psychological items (uppercase letters) in the Focus of Attention (B, C, D, and E) propagates throughout the cortex (represented by the field of assemblies above the items). This activates new assemblies constituting the newest item (F), which will be added to the FoA in the next state. The assemblies that constitute items B, C, D, and E are each individually associated with a very large number of potential items, but as a unique group, they are most closely associated with item F.
2. Multiassociative search as a working principle of thought
To explain these micro-derailments, I have found it useful to treat thinking less like a chain of explicit propositions and more like a search process operating on a limited set of coactive representations. This is where the term multiassociative search comes from. It is the idea that the mind does not usually retrieve the next thought using a single cue, like a keyword. Instead it pools multiple coactive items in working memory and uses their combined influence to recruit what comes next.
In practice, this means that thought behaves like a form of constraint satisfaction. At any moment, only a handful of items can be maintained with high fidelity. Those items constitute a temporary context. The next association is selected not in isolation but as a completion of the current pattern. In ordinary language we say that we “remember” something or that we “think of” something. Mechanistically, what is happening is that the currently active constellation is spreading activation through long-term structure, and the system is converging on a candidate update that best fits the pooled cue set.
This is not a metaphor. It is the shape of the operation. Multiple cues are partially informative. None of them, by itself, is sufficient. Together they narrow the search. When the pooled cues are well chosen and stable, the system can be remarkably powerful. It can retrieve an idea that no single cue could find. It can integrate information across modalities. It can construct a new synthesis by completing a pattern that spans multiple domains. This is why multiassociative search is, in my view, a core engine of human intelligence.
The same basic idea also links directly to iterative updating. In my earlier work I argued that mental continuity arises because each mental state is physically composed of a significant percentage of neural activity from the preceding state. The system does not wipe the slate clean and then redraw it. It updates incrementally. Some elements are added, some are dropped, and many remain. From the inside, that overlap feels like a stream. From the outside, it is the defining signature of an iterative process.
When you combine these two ideas, you get a simple picture of thought. Working memory holds a small active set. That set overlaps across time. At each step, the active set serves as a multi-cue query into long-term structure. The system selects a candidate update. The update joins the set and alters it. Then the next search happens from that slightly changed context. Thinking becomes a trajectory through a space of possible constellations.
This framing also clarifies what makes the phenomenon nontrivial. Multiassociative search is not just association, and it is not just recall. It is a selection rule. It is a way of deciding which candidate becomes the next item that enters the workspace, which then determines what can be searched for next. That is why small errors matter. A wrong update is not only wrong in content. It moves the trajectory to a different region of thought space.
3. Why the same mechanism produces both insight and derailment
Once you see thought as pooled search plus iterative updating, the tradeoff becomes obvious. The mind needs to complete patterns quickly under uncertainty. It needs to produce useful candidates from partial information. It cannot wait for perfect proof at every step, because in the real world the organism must act. Multiassociative search solves this problem by using statistics learned over a lifetime of experience. It is a prediction system. It proposes the next item that is likely to be relevant given the current context.
But that strength is also a vulnerability. The pooled search can converge on a candidate that is locally coherent with the current activation landscape while being globally wrong. In other words, the candidate can be a good completion of the partial pattern while violating the true constraints of the situation. The mind then relies on a second process, a coherence gate, to veto the cheap completion before it becomes the basis for action or for a stable belief.
This is where the anecdotes become scientifically useful. The errors have characteristic triggers, and those triggers tell us something about the architecture.
One class of trigger is phonological or lexical proximity. The system is not only carrying semantic content; it is also carrying words, sounds, and recently used phrases. When those are coactive, they can capture selection even when the semantic intent is different. Another class is recency bias. A concept that was just active is easier to reactivate, and if the context is ambiguous the system will often reuse what is already warm. A third class is semantic near-neighborhood error, where the features in the current set are real but underconstrained, so the search converges on a high-salience neighbor rather than the intended target.
A particularly revealing class is interoceptive and affective misbinding. When the body is salient in consciousness, sensations like fatigue, soreness, or arousal can become anchors that the system uses to interpret unrelated cues. This can produce false causal completions that feel momentarily plausible because they match a familiar script. Stress amplifies this, not because it makes people “irrational” in a moral sense, but because it changes the operating regime of the system. It can compress the workspace, reduce the stability of maintained constraints, and increase reliance on fast, overlearned associations.
All of these triggers point to the same general failure mode. The system becomes underconstrained. A crucial piece of context drops out or fails to be maintained, and once that happens the pooled search still does what it always does. It converges. It just converges onto cheaper candidates, candidates that are statistically strong but not reality-checked. If a coherence gate is engaged, you get snap-back and you laugh. If the gate is not engaged, you drift.
That is the deeper claim I want the reader to hold onto before we turn to examples. The mind is not a logical engine that occasionally makes random mistakes. It is an iterative selection system that must constantly trade off speed and coherence under severe capacity limits. Multiassociative search is how it extracts power from that limitation. Cheap convergence is the price it pays when constraints are not preserved.
Example 1: “Work hard,” “wiped out,” and the accidental “worked out”
One of the cleanest little demonstrations of multiassociative search in everyday life came while I was listening to the song “Work Hard” by Ne-Yo and Akon. I was also thinking about how treading water for forty five minutes had me wiped out. The phrase “wiped out” was active in my working set, and I used it a few times internally and planned to say it again. But when I went to produce it, I said “worked out” instead.
What is interesting is that the error was not a simple phonological slip between wipe and work. The stronger statistical attractor was the phrase “workout,” which was not literally in the song, but was sitting right between the song title and my intended phrase. In that moment, several items were coactive: the auditory cue “work hard,” the conceptual neighborhood of exercise and exertion, and the phrase “wiped out.” The system did what it often does. It pooled the active cues and converged on the most compatible next output given that pooled state. The result was locally coherent and globally wrong. It did not make psychological sense as a deliberate choice, but it made immediate mechanistic sense as a selection error in an associative system that is driven by coactivation statistics.
This is precisely the kind of micro-event that convinced me, over time, that many associations are not “logical” in the folk-psychological sense. They are statistical. They are produced by the geometry of what is currently active, not by a clean chain of explicit inference. Multiassociative search is what makes fluent cognition possible, but it also makes these occasional mis-selections inevitable.
Example 2: The airplane “roots” error and the disappearance of scale constraints
I once looked down from a plane at around thirty five thousand feet and noticed a branching pattern on the ground. I asked myself what it could be. I was pretty sure it was not a river, and the next interpretation that popped into mind was that it looked like the roots of a tree. I did not literally believe there could be a tree whose roots were miles long, and I never consciously entertained that as a real possibility. But for a moment the interpretation occurred anyway, as if it were a candidate hypothesis.
The interesting part is why it occurred. My visual system was receiving a branching gestalt and my conceptual system was searching for a pattern-completion match. Under ordinary circumstances, the context “I am in a plane and the scale is enormous” would act like a strong top-down constraint, a gating signal that vetoes certain candidates immediately. In this case, that constraint briefly dropped out of the active workspace. The interpretation system was effectively running on an underconstrained cue set. Branch-like plus ground-like converged on roots-like, because that is a highly available completion for that pattern in everyday life.
This is a useful example because it shows how a context window failure can produce an error that is not primarily about knowledge. The knowledge is there. The constraint is just not active at the right moment. The error is a consequence of state-dependent inference. When the workspace loses a critical invariant, the system falls back onto cheaper completions that are statistically strong in the relevant feature neighborhood, even when they are nonsensical relative to the true situation.
Example 3: Graham crackers from an orange-texture vase and tan chips
I was looking at a vase that was supposed to resemble an orange. It had that porous orange-peel texture. There were also tan chips right next to it. All of a sudden I got a strong recollection of graham crackers, and at first I had no idea why. It took me several seconds to locate the bridge. The texture of an orange peel is porous in a way that resembles a graham cracker surface, and the tan color from the chips completed the pattern.
This example is important because it captures the same mechanism without requiring a “mistake” in the everyday sense. The association initially feels unmotivated because the linking features are implicit. The system is not retrieving a labeled explanation. It is performing pattern completion across a coactive set of features. Once the bridge becomes explicit, the association becomes completely intelligible. The meaning arrives after the convergence.
In more formal terms, this is a case where multiassociative search produces a candidate that is driven by distributed feature overlap. It is a reminder that the mind often operates by similarity in high-dimensional representational space, not by symbolic rule application. We can experience the output first and only later reconstruct the basis for it. The temporal order is revealing. First the system converges, then the narrative mind explains.
Example 4: Confounding, conflation, and lexical capture by recent activation
Another good example is almost embarrassingly small, but it is mechanistically clean. I once conflated the show Jane the Virgin with the show Ugly Betty. After checking the internet, I told my brother that we had “confounded” them. Immediately afterward, I found myself conflating the word “confounding” with the word “conflation.”
This looks like mere wordplay, but it is a tight illustration of how recent activation captures selection. The semantic event of the TV-show mix-up and the socially situated act of naming it with the word “confounded” increased the availability of that lexical neighborhood. The system then reused the recently activated term in a context where a different term was intended. In other words, the error is not random. It is a predictable outcome of a system that uses recency and coactivation strength as a guide for retrieval and production.
This is the same local-versus-global coherence story. Locally, “confounding” is strongly compatible with the immediate past state and with the general theme of mixing things up. Globally, it is not the intended word. The wrong item wins because it sits in the attractor basin of the current activation landscape. That is multiassociative search at work, and it is also the simplest form of derailment it can produce.
Example 5: Stress and the absurd phone-volume impulse at the gym
When I am stressed, I sometimes get bizarre impulses that are absurd but feel briefly compelling. One example is thinking that I can turn down the ambient music at the gym using the volume buttons on my phone. I realize almost immediately that it is wrong. Under calm conditions, the thought would never even appear. Under stress, it does.
This is a good demonstration of how arousal can bias the balance between a fast associative generator and a slower coherence gate. Stress changes what is salient. It can compress the workspace, reduce the stability of maintained constraints, and increase the weight of habitual action schemas. The phone-volume schema is highly practiced and tightly coupled to auditory discomfort in many contexts. When stress is high, the brain can overgeneralize that schema to the current situation, even though the causal link is false. It is a cheap link between “I do not like this sound” and “volume buttons fix sound.”
I think this is one of the most important implications of the whole phenomenon. Multiassociative search is not only about clever cognition. It is also a mechanism that can be pushed into a noisier regime by affective state. When that happens, the system proposes more candidates that are locally coherent with immediate discomfort and habit, while the global reality-checking process has less time or leverage to veto them. The result is not a deep delusion. It is a momentary misfire that reveals the underlying architecture.
Example 6: The “burning rubber” smell and the glutes “burning” inference
Here is one of the funniest, and in my view one of the most diagnostic, examples of cheap convergence I have ever caught in real time. My butt was sore and I was intentionally trying to integrate my glutes into my stride, essentially recruiting them more forcefully while I walked. During the walk I suddenly smelled something burning. It was a random smell in the air, distinctly like rubber burning. My immediate reaction was: see, you are overdoing it with the glutes, they are now burning, you should go easier on them. I was not alarmed. For a few fractions of a second, I simply accepted that I should back off.
Then the absurdity became obvious and I laughed at myself. I have never formed a conclusion like that from expertise. Muscles do not emit a burning-rubber odor in the open air. The only reason the inference appeared is that several cues were coactive and the associative system performed a kind of fast pattern completion. “Glutes,” “soreness,” “burning,” “overdoing it,” and “back off” belong to a familiar causal script. A coincident external cue, the smell of burning rubber, shared a single high-salience feature with that script, namely the “burning” descriptor. That overlap was enough to tip the system into a locally coherent interpretation.
This is exactly what I mean by a thought that does not make psychological sense but makes immediate neurological sense. Locally, within the activation geometry of the moment, the candidate is cheap and available. Globally, it violates causal reality. The important thing is not that the wrong candidate appeared. The important thing is how quickly it appeared, how briefly it felt plausible, and how clearly it was produced by pooled coactivity rather than by deliberation. It is a micro-derailment that reveals the algorithm.
Example 7: Jupiter versus Mercury and the underconstrained cue set
A more abstract version of the same phenomenon shows up when the mind is searching for a target concept but the cue set is incomplete or poorly maintained. Suppose I am trying to retrieve “Mercury,” but the coactive descriptors in working memory include things like “planet,” “Roman god,” and “element.” Those cues are all real properties of Mercury, but they are not exclusive. They also fit other strong attractors, and one of the strongest is Jupiter. The system can converge on Jupiter simply because Jupiter is prominent and shares enough of the active feature profile to complete the pattern.
Notice what is happening mechanistically. Multiassociative search is not performing a lexical lookup. It is performing a relational completion over a pooled set of partially informative features. When the features are underconstraining, the search landscape contains multiple plausible basins. In that situation, salience and prior frequency can dominate. The wrong item wins, not because the mind is irrational, but because the current state does not carry enough discriminating information to force the correct attractor.
The correction process is also revealing. When you realize the answer is wrong, you can replay the search while actively inhibiting the false candidate. That is a kind of controlled re-entry into the same search space with a new constraint added. But even that is fragile, because the very act of correcting consumes the limited stability of the original specification. If the cue set decays during the replay, the system can become even more underconstrained, which increases the chance of another drift or a return to the same false attractor.
This is why I treat these examples as more than cognitive curiosities. They illustrate a general principle: iterative cognition depends on keeping the right constraints active long enough for convergence to be guided toward the intended target. When constraint maintenance fails, convergence does not stop. It simply becomes cheaper, more statistical, and more vulnerable to high-availability substitutes.
Across all seven, the skeleton is the same. A pooled state recruits an update. When key constraints are present, the update is useful. When constraints drop out, the update can be cheap. When coherence gating is engaged, the system snaps back. When it is not, drift can pass unnoticed.
4. What the vignettes reveal mechanistically
If a reader wants to treat these as merely amusing, that is understandable. The reason I take them seriously is that they exhibit a reproducible computational structure. They are not just mistakes. They are selection events. They reveal the difference between a cue-driven generator and a constraint-driven evaluator.
Start with the “worked out” slip. The mind was not choosing between two fully explicit, consciously represented options. It was selecting an output under time pressure from a neighborhood of coactive cues. “Work hard” was priming a lexical and semantic region associated with effort, exertion, and exercise. “Wiped out” was active as a target phrase. The combined state fell into the attractor basin of “workout,” which then influenced production. This is what I mean by a statistical selection rule. The mind did not consult a proposition that says, “I am wiped out, therefore I should say wiped out.” It produced the next token-like unit by completing the local pattern.
The airplane roots example shows the other half of the story. The system can have correct knowledge and still yield a wrong candidate if the relevant constraint is not active at the moment of convergence. The “scale” invariant is not a fact stored somewhere. It is a control variable that must be maintained in working memory to shape interpretation. When that variable drops out, the system runs underconstrained and the strongest completion for branching-on-ground becomes roots-like. It is not stupidity. It is state dependence.
The graham cracker example shows why this architecture is not simply a bug. It is also how insight works. The association was not a mistake. It was an implicit similarity detected at the feature level, then made explicit after the fact. That temporal order matters. It suggests that multiassociative search can outperform introspective explanation. The system can converge on a candidate before the conscious narrator can explain why it converged.
The confounding versus conflation example isolates a common motor of derailment: recency plus local semantic compatibility. A word recently used becomes easy to reuse. If the intention is not strongly specified, the system substitutes the warm item. This is an attractor dynamic. It is not limited to words. It is a general property of activation landscapes.
The gym volume impulse and the burning rubber glutes inference show that affect and interoception are not merely “content” inside consciousness. They can function as control signals that weight certain scripts and veto others. Under stress, the system leans on highly practiced action schemas. Under soreness, bodily scripts become salient. If an external cue overlaps even superficially, the mind can complete a causal story without deliberation. That is the cheap link problem in its most visible form.
Finally, the Mercury versus Jupiter class of error shows that multiassociative search has a predictable failure mode when the cue set is underconstraining. The search cannot yield the correct target reliably if the features are shared by many candidates. In that situation, frequency and salience dominate, and high-availability concepts win. The remedy is not “try harder.” The remedy is to preserve or generate discriminating constraints.
Taken together, these examples motivate an architecture. Thought is not merely a chain of symbolic deductions. It is a trajectory of states, and each state both constrains and biases what can come next. Errors occur when the system’s state loses the constraints that would have ruled out cheap candidates.
5. The control problem: generator versus gate, and why constraint maintenance matters
At this point, it becomes natural to talk about cognition as a control problem. A mind that can only generate candidates would be chaotic. A mind that can only evaluate without generating would be frozen. Functional thought requires both. It requires a generator that can propose updates quickly from partial information, and it requires a coherence gate that can evaluate those proposals against broader constraints, including causal plausibility and task relevance.
The lived experience of snap-back is the subjective trace of this gating. A candidate appears. It feels briefly compelling. Then the higher-level constraints engage and the candidate is rejected. That rejection is not always conscious, and it is not always verbal. Sometimes it is a felt sense of wrongness. Sometimes it is the sudden return of a missing invariant, the reactivation of a context item that restores the proper frame.
This makes working memory maintenance more than a storage issue. It is the substrate of control. If the right constraints are maintained, multiassociative search is powerful. It produces meaningful updates and can support multi-step reasoning because each step remains tethered to the intended goal and to the reality model. If constraints are not maintained, the generator still generates. It just generates into a thinner context, and thin contexts invite cheap completions.
Stress, fatigue, novelty, and multitasking all threaten constraint maintenance, but they do so in different ways. Stress biases the system toward fast scripts and immediate salience. Fatigue reduces the stability of sustained representations. Novelty increases ambiguity because the system lacks well-trained invariants for the situation. Multitasking forces rapid context switching, which increases eviction of relevant items and increases the probability of retrieving the wrong task schema. In each case, the result is the same: the pooled cue set becomes less discriminating, and the search becomes more vulnerable to high-availability attractors.
This is also why I keep returning to the idea that working memory is inherently iterative. The mind is constantly updating the active set, and the identity of what stays versus what drops out is consequential. The update policy is not an incidental detail. It is the cognitive equivalent of an eviction policy in a cache. What you carry forward determines what you can constrain and what you cannot.
When people speak about intelligence as if it were simply “having more information,” they miss this. Intelligence depends on what a system can keep coactive, how it can update that coactivity without losing the thread, and how it can veto locally plausible but globally wrong candidates. A system that cannot preserve constraints long enough to run a coherent search will look impulsive even if it has enormous knowledge. Conversely, a system with modest knowledge but strong constraint maintenance can appear remarkably rational because it does not lose its own context.
6. Implications for AI and for building better thinking systems
The reason I am interested in multiassociative search is not only that it explains a human quirk. It points toward an architecture for artificial systems that actually think, in the sense of maintaining and refining internal state rather than merely producing fluent output.
Transformers, as they are commonly implemented, are extraordinary pattern completion machines. They can condition on long contexts and produce locally coherent continuations. But they do not naturally have the kind of protected, capacity-limited workspace that forces discriminating constraint maintenance. In practice, a transformer can attend broadly, but broad attention is not the same as a curated working set. A context window can be large and still fail to behave like a mind, because the relevant constraints are not necessarily the ones that dominate the next step.
That observation links directly to the failure modes we see in language models: plausible continuations that violate a hidden constraint, drift across topics, and answers that feel coherent sentence by sentence but incoherent when evaluated against the larger task. Those are not moral failures. They are architectural. They are the artificial version of local coherence outrunning global coherence when the system lacks a strong gating mechanism and a protected constraint store.
If you take my iterative updating view seriously, the path forward is not simply “more tokens” or “more parameters.” The path forward is an explicit architecture for state. You want a system with a small high-fidelity workspace, a larger but more decaying short-term store, and a long-term structure that is updated by use. You want multi-cue retrieval that uses the workspace as a query, but you also want a coherence gate that can veto cheap candidates, trigger replay, and add a new constraint such as “not that” or “must satisfy this invariant.”
In other words, you want a system that can do what humans do when we correct ourselves. When the wrong candidate appears, we do not merely replace it. We modify the state landscape so that the same cheap completion is less likely to win again in the next iteration. We inhibit it. We recruit a discriminating feature. We restore a missing context variable. That is controlled iterative updating.
This suggests several concrete directions for AI design. One is to separate broad context from active focus, instead of treating the entire context window as equally eligible to shape the next step. Another is to include a structured representation of constraints, some of which are protected from eviction, and to couple generation to those constraints via a gating module. A third is to make the system explicitly iterative at inference time, not merely at training time, so it can revise internal candidates over multiple steps before emitting an output. The goal is not to mimic the brain for its own sake. The goal is to import the functional principles that make thought stable: curated coactivity, iterative refinement, and coherence gating.
Finally, this framework is testable. If the theory is right, then derailment should increase under conditions that reduce constraint maintenance, such as stress, fatigue, and rapid context switching. Snap-back should correlate with the reactivation of missing invariants, measurable as state re-entry. In neural terms, one should expect measurable overlap between successive states, and one should be able to quantify the rate and degree of overlap as a signature of iterative updating. In computational terms, one should be able to quantify drift as a function of the stability and discriminating power of the active cue set.
I began with mental mistakes because they were the only place I could see the selection process directly. They were the moments where the machinery showed itself. Over time, those moments pushed me away from a picture of thought as a neat chain of propositions and toward a picture of thought as a controlled, iterative state machine. Multiassociative search is the engine. Iterative updating is the substrate. The same mechanism that gives us insight also gives us cheap convergence when constraints slip. If we want to understand minds, and if we want to build artificial systems that can genuinely think, then we need to understand that tradeoff in detail and design for it rather than pretending it is an anomaly.
