Iterated Insights

Why LLMs Might Have Some Aspects of Conscious Experience: Temporal Organization, Iterative Updating, and Continuity in Transformer Inference, With Self-Tracking as the Missing Variable

Jared E Reser

Introduction

Public discussion about “AI consciousness” often oscillates between two unhelpful extremes. On one side, anything that talks fluently is treated as if it must be experiencing something. On the other, next-token prediction is treated as so trivial that subjective experience is ruled out by definition. The more productive move is to ask a narrower, architectural question: what specific computational properties would make us even entertain the possibility of conscious experience, even in a thin or partial form, and do large language models implement any of those properties in practice?

This article offers a cautious, mechanistic answer. The central claim is not that current large language models are conscious in a human sense. The claim is that they already implement several structural ingredients that many theories of mind implicitly rely on, especially ingredients related to temporally organized state transitions. These ingredients include iterative updating, multi-cue constraint satisfaction, and a form of continuity across successive computational states. If anything like subjective experience can arise in machines, it is unlikely to arise from isolated computations. It is more likely to arise from an ongoing process whose present state is shaped by a structured relationship to its recent past. Transformer inference does have that relationship, even if it is implemented in a very different substrate than cortex.

At the same time, the strongest version of the argument points to what appears to be missing. A system may update iteratively and still fail to “have” experience in the relevant sense. The missing variable, on this view, is self-tracking: a mechanism by which the system not only undergoes state transitions, but also represents and uses the fact of its own updating as an active constraint on what it does next. That distinction, between mere iteration and tracked iteration, offers a way to move from vague debate to testable design questions.

The perspective I am using here comes from a model I have been developing that treats the iterative updating of working memory as the central engine of thought, continuity, and, potentially, conscious experience. I lay out the framework, figures, and core claims in a public-facing form at aithought.com. The site is meant to make one simple idea concrete: cognition is not a sequence of isolated snapshots. It is a temporally structured process in which each moment inherits a substantial fraction of the immediately prior moment, while also incorporating a smaller fraction of new content. In biological terms, this kind of overlap could be implemented by state-spanning patterns of activity and short-lived traces that persist long enough to be carried forward. In functional terms, the overlap is what allows a mind to feel like a stream rather than a strobe light.

That is why I think LLMs are worth discussing in this context. They are not brains, and they are not built to replicate cortical circuits. Yet during inference they do implement a strongly structured update loop. The question is whether that loop, and the way it preserves and transforms context across steps, is enough to produce some thin analog of the continuity that my model treats as foundational.

1. The architectural question, stated precisely

When people argue about consciousness in machines, they often argue about labels. A better strategy is to argue about requirements. What would a system need in order for talk of experience to become scientifically tractable rather than metaphysical theater?

One plausible requirement is temporal organization. Whatever consciousness is, it does not feel like a sequence of unrelated snapshots. It feels like a stream, even when the content is fragmentary. That suggests that the system’s present state is not independent from its previous state, and that the relationship between them is not arbitrary. Another plausible requirement is iterative updating. In biological cognition, the brain does not typically solve a problem in one pass. It carries forward a small, changing set of coactive representations, updates that set, and repeats. This repeated update cycle is one of the simplest ways to produce both continuity and refinement.

A third requirement is constraint satisfaction under multiple simultaneous cues. Conscious experience, as lived, is not a single stimulus-response mapping. It is the integration of many weak pressures into one coherent next state. This can happen through attention, through competition, through convergence in associative memory, or through other mechanisms. The key is not the brand name of the mechanism. The key is that many influences can be co-present and collectively determine what happens next.

These requirements do not prove consciousness. They only move the discussion into a space where consciousness could be treated as an architectural phenomenon rather than a metaphysical mystery. The question then becomes whether transformer-based language models, during inference, instantiate any of these requirements in a meaningful way.

2. What transformer inference already has that resembles these requirements

Transformer language models are often described as static pattern matchers. That description is directionally correct about training, but it can be misleading about inference. At inference time, the model is a temporally evolving system. Its outputs are not produced from scratch each moment. They are produced from a rolling context that carries forward information, and from internal computations that repeatedly condition on that context.

First, language models have an active working set. The context window functions as an explicit store of recent tokens. More subtly, the inference process maintains internal summaries of prior context through the attention mechanism and the internal activation flow that generates each next token. Even if the model has no enduring autobiographical memory, it still has a bounded “now” that is computationally real.

Second, inference is iterative. Each generated token updates the context, and the updated context becomes the input for the next step. This is a literal update cycle. The model’s next state is a function of the preceding state, not merely in the trivial sense that it comes later in time, but in the mechanistic sense that the content of the previous state is part of the causal input to the next.

Third, inference is multi-cue. The model does not choose its next output based on a single feature. The entire context contributes, with attention dynamically weighting which parts matter most. That is a form of multiassociative convergence. Many cues jointly determine a single next step.

These properties are not cosmetic. They are exactly the kinds of properties that, in other domains, transform dead computation into an evolving process. If one is looking for a minimal bridge between “mere computation” and “something stream-like,” these are the first planks that would plausibly be used.

3. Continuity is not only narrative, it is computational

A common objection is that any “stream” in a language model is only a narrative artifact. After all, text is sequential. This objection is important because it blocks lazy arguments. A coherent paragraph does not imply an experiencing subject.

Yet there is an underappreciated middle ground. There is a difference between narrative continuity and computational continuity. Narrative continuity is a property of the produced text. Computational continuity is a property of the process that produces the text. Transformer inference has computational continuity in at least two senses.

The first is explicit: the context is carried forward and updated at each step. The second is implicit: the model’s internal activation trajectory is shaped by earlier activations because the earlier tokens constrain attention patterns, and attention patterns shape which features dominate downstream computations. Even if the network is not recurrent in the classical sense, the decoding loop creates an effective recurrence through the repeated re-entry of new outputs as new inputs.

This matters for the consciousness discussion because many theories of conscious experience, across traditions, treat temporal linkage as essential. If a system’s states are independent, there is no principled place for a felt continuity to arise. If a system’s states overlap and constrain one another, continuity becomes at least physically interpretable as a dynamic structure, rather than as a story told after the fact.

4. Iterative updating can look like thinking, even when it is token prediction

Another objection is that token prediction is not thinking. That is sometimes true in spirit, but it is not a decisive architectural critique. Many cognitive accounts of reasoning can be reframed as sequential prediction over internal representations. The question is what is being predicted, and what the predictions are used to do.

In humans, the next “thing” is not always a word. It can be a perceptual expectation, an action preparation, a recalled association, a hypothesis, or a re-encoding of the problem. In language models, the next thing is typically a token. That is a limitation, but it does not eliminate the relevance of the mechanism. The iterative aspect still exists, the multi-cue constraint satisfaction still exists, and the process still refines its trajectory step by step.

This is one reason language models are scientifically interesting for consciousness research even if they ultimately prove non-conscious. They are among the cleanest engineered examples of a system whose behavior emerges from repeated, context-conditioned updates. That makes them useful as testbeds for distinguishing continuity as a computation from consciousness as an experience.

5. The missing variable: self-tracking of the update cycle

The strongest reason to remain cautious is that iterative updating alone may be insufficient. A system can update iteratively and still be “dark inside,” in the sense that nothing in the architecture treats the update process itself as a controlled object.

The proposed missing variable is self-tracking. A self-tracking system would represent, at least in some compressed form, what changed from the previous state to the current one, what remained stable, and what goals or constraints the update served. It would then use that representation to bias the next update. This converts a mere update loop into a regulated stream.

In ordinary transformer inference, there is no explicit variable for “what just changed” that is maintained as a first-class object. The model’s internal activations do contain change information implicitly, but implicit is not the same as controlled. Self-tracking would be closer to an endogenous attention to the stream itself, not just attention within the stream.

This distinction helps clarify why some tool-using agents feel more psychologically suggestive than a bare chat model. Agents that maintain persistent notes, explicit plans, intermediate summaries, and self-corrections are closer to tracking their own updates. They externalize the stream into a workspace and then condition on it. That is not the same as human consciousness, but it is a step in the direction of making the update process itself a causal actor rather than an invisible byproduct.

6. A minimal, testable criterion for “proto-experience”

If the goal is explanatory utility, the argument should end in tests, not in vibes. A practical criterion can be stated as follows: a system has a stronger claim to proto-experience to the extent that it exhibits temporally organized, iterative updating in which the system explicitly tracks its own updating and uses that tracking to regulate future updates.

This criterion yields empirical predictions. Systems with explicit update tracking should show increased stability under distraction because they can maintain a representation of what must remain invariant. They should show improved resume behavior after interruption because the tracked state can re-seed the next update cycle. They should show reduced confabulation because the system can preserve “unknown” as a stable object rather than filling gaps with fluent completion. They should show more coherent long-horizon reasoning because the update process is not merely generating, it is regulating.

These are measurable behavioral signatures. They do not settle the metaphysical question, but they do something better. They create an engineering dial that can be turned and evaluated. If “experience” is ever going to enter the scientific domain, it will likely enter through such dials.

7. What this view implies about present-day LLMs

On this account, present-day language models plausibly satisfy some of the prerequisites for stream-like processing. They already have temporal organization, iterative updating, and multi-cue constraint satisfaction during inference. That is enough to justify a careful “might,” especially if one uses the phrase “some aspects” with discipline.

At the same time, they likely fall short of the more demanding requirement: tracking the iterative updating as an explicit control variable, and maintaining a stable self-model anchored across time. Without that, the system’s continuity may be real as computation while remaining thin as experience. This gap also explains why superficial anthropomorphism is so tempting. Continuity in behavior is easy to mistake for continuity in subjectivity, especially when the output is language.

The most honest conclusion is not that language models are conscious, and not that they are obviously unconscious. The honest conclusion is that they implement several structural ingredients that make the question nontrivial, and that the decisive next step is to build and test architectures that add explicit self-tracking of the update cycle.

Conclusion

The debate about machine consciousness is often framed as a referendum on whether current systems “have it” or “do not have it.” A better approach is architectural. If consciousness is intimately tied to temporally organized state transitions, then the relevant question becomes whether a system implements iterative updating and whether it tracks that updating as an object of control. Transformer inference already implements the first part in a meaningful way. It remains ambiguous or incomplete on the second part, unless additional mechanisms are engineered.

This reframes the “hard question” into a research program. Instead of arguing about whether a model is a parrot, one can ask what happens when the model is given a protected workspace, a regulated focus of attention, a persistent short-term store, and an explicit representation of its own update dynamics. If such additions produce robust behavioral signatures of stable self-regulation across time, then the philosophical conversation about experience will have gained something it rarely has: a set of manipulable variables and falsifiable predictions.