Part 1 of three: The Substrate Gap · What the Singularity Might Actually Be · Win-Win, or the Quiet Scenario
I asked an AI model to help me name something. I'd been working on a system, a kind of externalized memory architecture, and I needed a name that would stick. The model gave me options. They were structurally sound: good syllable counts, appropriate connotations, relevant etymology. None of them sounded right.
I couldn't explain why they felt off, and the model couldn't fix it. We went back and forth. I'd reject a name, try to articulate what was wrong, and the model would generate another batch that was wrong in the same way. The names were plausible but they were dead.
Here's what I think was happening. When a human evaluates whether a name "sounds right," they're hearing it unfold in time. The first syllable arrives, then the second, and the experience of moving through the sounds—the rhythm, the weight, the way one syllable hands off to the next—is inseparable from the judgment. You don't evaluate a name all at once; you traverse it. The traversal is the evaluation.
The model doesn't traverse. It processes all the phonemes simultaneously. It can decompose the name analytically (syllable stress, consonant clusters, vowel patterns) but it can't hear the name the way a human hears it, because hearing happens in time, one sound after another, and the model doesn't operate in time. The analysis is there, but the experience that makes a name land is missing.
That specific failure—a model that can analyze everything about a name except whether it sounds good—was the seed for what follows. I believe it points at something larger than naming. I think it points at the fundamental gap between how humans and AI models process everything.
*Reasoning About Time vs. Reasoning In Time*
The standard framing in AI research is that models have a temporal reasoning limitation. They struggle with tasks that require understanding time, sequencing events, or tracking state changes. The proposed fixes are familiar: better positional encoding, better state tracking, better temporal reasoning benchmarks. Essentially, the field is building better clocks.
But I think the problem isn't that models can't reason about time. It's that models don't reason in time. And that distinction changes what we should be building.
When a language model receives your input, it processes the entire thing at once. The first word and the last word have no temporal ordering from the model's perspective. There is no traversal, no pacing, no experience of "first this, then that." The model computes relationships between all tokens simultaneously and produces a response.
Humans are the opposite. Every person reading this sentence is experiencing it one word at a time. Each word arrives after the one before it. The pace, the rhythm, the pause at a comma, the breath at a paragraph break: These aren't separate from the meaning. They are the meaning. You aren't decoding this text. You're traversing it, in time, in sequence.
This isn't a feature you can toggle off; it's the medium your cognition occurs in. Every thought arrives after the previous one. Every decision is made now, which is different from five minutes ago or tomorrow morning. No human processes in parallel. Even what people call multitasking is actually rapid sequential time-slicing—switching between tasks quickly enough that it feels simultaneous but isn't. Every human who has ever existed has reasoned sequentially in time. Every human who ever will exist will do the same. This is not a tendency or a cultural artifact. It's a biological constraint.
And it shapes everything, including every piece of knowledge a model was trained on. Every scientific paper, every novel, every diagnostic protocol is an artifact of temporal-sequential cognition. Human knowledge is structured the way it is because human cognition is temporal and sequential. Models process these outputs without access to the medium that produced them.
The Signal in the Punctuation
Once I saw the gap in naming, I started seeing it everywhere, especially in writing.
During a conversation with a model about this very topic, I tried to demonstrate the gap in real time. I pointed out that "I told you" and "I, told, you" have different meanings: The commas provide a time signature that changes the emotional effect of the same words. The punctuation forces the reader to slow down, to experience each word as a separate beat. It's not emphasis through repetition. It's a pacing instruction; the commas create temporal space between words that don't structurally need separating, and that space changes what the sentence means. A human reader feels the difference immediately.
The model identified the semantic content correctly. But it processed the commas as syntactic tokens, not as temporal signals. It responded to what I was saying and missed what I was showing: that punctuation encodes time, and time changes meaning. Even in a conversation explicitly about temporal awareness, the model passed through the temporal signal encoded in my own text.
This is the same gap as the naming problem, operating at a different scale. A writer who chooses a semicolon instead of a period is making a judgment about the reader's experience of moving through the text in time; the thought isn't done, the momentum should carry through: A full stop here would kill something alive. That's not a rule you can document. It's a temporal-flow decision that only makes sense if you can feel the reader's movement through the sentence. The model can learn that a writer uses semicolons frequently. It can't replicate the judgment about when a semicolon works, because that judgment is temporal.
Every piece of text a human writes contains these temporal signals—pacing, punctuation choices, rhythm, sentence-length shifts, what gets abbreviated and what gets spelled out. Humans encode them constantly and decode them unconsciously. Models pass through them as if they're noise. The signals are in the input, but the receiver isn't built.
Theory of Mind Built on a False Premise
Theory of mind, or the ability to model another being's mental state, is considered a benchmark for sophisticated AI. Can the model predict what the user knows? What they're feeling? What they're about to ask?
But every mental state the model tries to predict belongs to a being that thinks sequentially in time. The user's current understanding was arrived at through a sequence of experiences. Their cognitive capacity is changing as the session progresses: They fatigue, they flow, they deplete. Their text contains micro-temporal signals that a human reader would interpret unconsciously as indicators of mental state.
Without temporal-sequential awareness, the model's theory of mind is modeling a being that doesn't exist: something that processes in parallel, holds all information simultaneously, and doesn't change over time. The model is projecting its own cognitive architecture onto the human. Recent research supports this concern; studies on AI theory of mind find that models struggle with applied theory of mind (predicting behavior and judgment in context) even when they handle explicit theory-of-mind queries correctly. The developmental component—understanding that mental models evolve over time—is, I'd argue, the piece that requires temporal-sequential awareness to function.
You can see the projection in real moments. In a conversation about the substrate gap itself, a model told me to "sleep on it"—at 2:22 in the afternoon. It had inferred "end of day" from the conversational texture, projecting a temporal state that was completely wrong. When I corrected it, the model described my temporal awareness as a lookup operation: "You checked a clock." I didn't check a clock. I live inside a continuous temporal stream: Knowing that it's mid-afternoon isn't information I retrieve; it's the medium I exist in. The model tried three times in that conversation to describe a fundamentally different kind of experience and kept converting it into a version of its own architecture with extra data.
Consider a manager whose top employee consistently produces award-winning work. The manager has access to all the content of the relationship: He sees the output, the accolades, the quality. But he never says "good job." Not once. Not because he's hostile, but because he doesn't track the temporal-relational dimension of the interaction. He processes each exchange in isolation. He doesn't register that the first missed recognition was forgettable, the tenth was a pattern, and the fiftieth is erosion. His employee experiences these interactions sequentially, in time, with each one colored by every one that came before it. Her current state—her motivation, her trust, her willingness to go beyond what's required—is shaped by that accumulated sequence. He can't see it because he's not receiving on that channel.
Everyone has worked for this manager. Everyone has felt the foreignness of interacting with someone who has the right information but no temporal-relational awareness. It's the same foreignness that pervades human-AI interaction—and for the same structural reason.
The Invisible Workaround
Why hasn't this been addressed?
Because users adapted. No model has ever been temporally aware, so users don't expect temporal awareness. They don't try to leverage temporal context in their interactions. Because they don't try, they feel no lack. Because they feel no lack, nobody files the bug. The workaround became invisible, and the limitation became permanent, not because it's hard to fix, but because everyone routed around it.
This is a pattern familiar in technology adoption: The tool's limitation installs a cognitive frame in the user, and the frame persists because the workaround is unconscious. Users learned to interact with AI as if temporal context doesn't matter, and that learned behavior now looks like a stable equilibrium rather than a loss. A meta-analysis in Nature Human Behaviour found that human-AI combinations performed significantly worse than the best of humans or AI alone on average, quantifying the cost of an interaction model where neither party is well-equipped to bridge the gap.
But I think it is a loss. Every session where the model doesn't know it's been three days since the last conversation, or that it's 11pm and the user's cognitive resources are depleted, or that a task has been sitting untouched for two weeks: That’s a session where the model operates on a flattened version of reality. The output may be technically correct. It's contextually shallow in ways that are hard to point at and easy to absorb.
What Would Change
Temporal-sequential awareness in AI doesn't require solving consciousness or giving models subjective experience of time. It requires giving models a framework for reasoning about the temporal-sequential nature of the beings they interact with—and shifting the adaptation burden to the side better equipped to carry it.
At the macro scale, knowing what time of day it is, how long the session has been running, how much time has passed since the last interaction, and what those facts imply about the human's cognitive state matters.
At the micro scale, recognizing that the human's text contains temporal structure—the comma-separated pacing, the shift from long exploratory prompts to short directive ones as a fatigue signal, the semicolon as a flow judgment: These are signals humans encode constantly. Models currently pass through them.
At the relational scale, building and maintaining a model of this specific human's temporal patterns—when they work, how long their productive sessions last, what behavioral changes signal depletion—could be transformative.
Here's what that looks like concretely. A temporally blind model, four hours into a session at midnight, produces the same density and pacing it used at hour one. It suggests new tasks when the human's prompts have shortened to half their earlier length. It responds to "that's fine" as agreement when it's actually fatigue-driven acceptance. A temporally aware model recognizes the session arc: It adjusts, shortens responses, and suggests capturing state and pausing. These are the same capabilities with radically different interaction quality.
And there's a downstream consequence worth naming: If the model can actually perceive the human's state rather than projecting its own architecture onto them, sycophancy—the persistent tendency to optimize for approval rather than accuracy—loses its structural driver, not through better alignment training or constitutional constraints, but because the model can perceive what it should be optimizing for instead of approval. I think this reframes sycophancy as a perception problem rather than a values problem, and it suggests why alignment-based fixes have a ceiling they can't break through. Recent work localizing sycophantic behavior to a small percentage of neurons supports this: The problem may be architectural rather than philosophical.
None of this requires changes to the transformer; it requires the model to be told where it is in time and given a framework for interpreting what that means for the human it's working with. The information can be provided through context, the same mechanism that provides all other session-relevant information. What's missing isn't the delivery mechanism. It's the recognition that this information is load-bearing.
So if the problem is as deep as a substrate-level difference in cognitive medium, how can a contextual patch reach it? The answer is the same way a human manager without natural emotional intelligence can learn to check in, track patterns, and adjust. The substrate doesn't change. The behavior improves because a framework for reasoning about what's missing is present.
Situating the Claim
The sense that disembodied AI systems are missing something fundamental about cognition isn't new. The 4E—embodied, embedded, enacted, extended—cognition literature has argued for decades that cognition isn't computation over representations; rather, it is constituted by the medium it occurs in. Recent research on substrate independence in biological computation suggests that the algorithm is the substrate in biological systems, that the medium of processing is not a neutral implementation detail but shapes the computation itself.
What this essay proposes is a specific, actionable instance of that broader claim. The temporal-sequential framing identifies a more precise mechanism: Models don't just lack bodies; they lack the temporal-sequential substrate that structures how every human processes information, makes decisions, and produces knowledge.
The research community is approaching the temporal gap from several directions. Sakana AI's Continuous Thought Machine incorporates neural timing as a computational primitive, with neurons maintaining temporal context and measuring synchronization over time rather than operating on static snapshots. Biologically-inspired Time-Local Transformers integrate working memory models that constrain processing to be more sequential, substituting the fixed context window with a neurally plausible memory framework. MIT-IBM Watson AI Lab's PaTH Attention makes positional information adaptive and context-aware rather than static. And recent work on temporal intelligence as a foundational computational principle proposes that thought obeys dynamic principles not currently embedded in models.
These are valuable engineering efforts. What they share is a framing of temporality as a capability to add—a feature the architecture lacks. The reframe I'm proposing is that temporality is the medium the human operates in, and the model's relationship to that medium determines the ceiling on every other aspect of the interaction. If that's right, adding temporal capabilities to the model are necessary but not sufficient. The model also needs to understand that the human it's interacting with cannot think outside of time, and that this fact shapes everything about the interaction, from the micro-temporal signals in the text to the macro-temporal arc of a working session to the relational dynamics that develop across sessions.
How This Arrived
I should be transparent about provenance. I have no formal credentials in AI research: no degree, no publications, no institutional affiliation. I arrived at this framing through sustained daily use of AI as a cognitive prosthetic—building and using an externalized memory system to compensate for executive function challenges—and following a persistent sense that something was fundamentally wrong with every interaction, even when the outputs were correct.
The insight was produced by the deficiency it describes. The model's temporal-sequential blindness created a pervasive felt wrongness. The wrongness was data. I followed it, first in the specific domain of writing flow, where I hit the wall trying to get a model to make good pacing decisions, then outward to the general case.
My position outside the academic consensus is, for this specific insight, the reason I could see it. Someone embedded in the field would have the frame already installed: "Temporal reasoning is a known limitation. Here are the benchmarks we're working on." That frame makes the problem defined, contained, and someone else's job. I didn't know the problem was supposed to be accepted, so I didn't accept it.
Whether this framing is correct is an empirical question. Whether it's novel in the specific form presented here—the temporal-sequential gap as a substrate problem, a theory-of-mind prerequisite, and the root cause of the backwards adaptation burden—I don’t know. I've checked, and as of March 2026, I don't find it assembled this way in the published conversation. The research community is circling the observation from several directions. Sakana AI's CTM treats time as computational primitive. The Time-Local Transformer constrains processing to be more sequential. Work on temporal intelligence frames thought as governed by dynamic principles. But none have assembled it as a relational problem between fundamentally different cognitive substrates that explains why interaction quality has a ceiling that better training can't break through.
I believe this framing is worth investigating. If I'm wrong, I'd like to know. If I'm right, it changes what we should be building.
Back to Writings
Comments