Capability Is Not Continuity

Large language models have demonstrated capabilities that increasingly resemble substantial parts of human intellectual work. They can reason across domains, synthesise information, generate and analyse software, interpret multimodal inputs and interact with external tools.

These achievements have encouraged a tendency to treat increasing model capability as synonymous with progress towards increasingly general intelligence.

This paper argues that the two should be distinguished.

Large language models are extraordinarily capable computational models, but ordinary model inference remains predominantly episodic. A trained model receives context, performs inference and produces an output. Although its behaviour can adapt extensively to information available within that context, the interaction does not ordinarily produce a durable change in the deployed model itself.

This creates an important distinction between capability and adaptive continuity.

Increasing model scale, context length and reasoning performance can substantially improve what a model is capable of doing during an interaction. They do not, by themselves, establish that the model continuously accumulates and consolidates experience from the consequences of its previous activity.

External memory, retrieval, agent architectures and continual adaptation can provide forms of persistence at the wider system level. This does not invalidate the distinction. It establishes that continuity can be a property of the system surrounding a model rather than something implied by increasing model capability alone.

The argument is therefore not that large language models cannot form part of continuous intelligent systems. It is that capability during inference and adaptation across time are distinct properties, and evidence of progress in one should not automatically be treated as evidence of progress in the other.

1. Introduction

Large language models have transformed the practical capabilities of computational systems.

The Transformer architecture established a highly scalable attention-based foundation for sequence modelling^[1]. Subsequent scaling demonstrated that a single language model could exhibit broad zero-shot, one-shot and few-shot capabilities across tasks without task-specific gradient updates during inference^[2].

The resulting systems have progressed rapidly.

Modern foundation models can interpret ambiguous instructions, reason through unfamiliar problems, produce sophisticated written work, analyse software, synthesise large quantities of information and interact with external environments through tools.

Research has also demonstrated that LLMs can be incorporated into systems that interleave reasoning and action^[4], maintain external memories^{[5, 6]} and operate across interactive environments rather than merely producing isolated textual responses^[7].

The question is therefore no longer whether these models are capable.

They plainly are.

The more difficult question is what follows from that capability.

If foundation models continue to become more capable, does sufficiently advanced model capability eventually amount to continuously developing intelligence?

Perhaps.

But the conclusion does not follow from capability alone.

There is a distinction between two properties:

what a system can do, and what durably changes because it has done it.

The first concerns cognitive capability.

The second concerns what this paper calls adaptive continuity.

These properties can interact, but they should not be conflated.

2. Scope and Position

The argument advanced here is deliberately narrow.

It is not that LLMs cannot learn.

They clearly can during training, fine-tuning and other forms of parameter adaptation.

It is not that LLMs cannot adapt within context.

Few-shot and in-context learning demonstrate that model behaviour can change substantially in response to examples supplied during inference without corresponding parameter updates^[2].

It is not that LLM-based systems cannot possess persistent memory.

Retrieval-augmented generation explicitly combines parametric model knowledge with non-parametric external memory^[3].

Nor is it that LLM-based agents cannot use previous activity to improve subsequent behaviour.

Systems such as Reflexion preserve linguistic feedback across trials and use it to influence later decisions without changing the underlying model weights^[5]. Generative-agent architectures have similarly demonstrated persistent records, retrieval and reflection across simulated experience^[6].

These developments matter.

They also make precise terminology essential.

In this paper:

Model refers to the trained parameterised foundation model performing inference.

System refers to the wider deployed architecture within which model inference occurs, potentially including memory, tools, retrieval, state and other mechanisms.

The thesis is therefore not:

It is:

That distinction is the subject of this paper.

3. Intelligence Within an Episode

An LLM performs inference over information made available to it.

Within an interaction, this can produce remarkably adaptive behaviour.

A model can receive new evidence and immediately incorporate it into subsequent reasoning. It can discover that an earlier assumption was incorrect, reconsider a plan and produce a different conclusion.

Few-shot learning provides a particularly clear demonstration. Brown et al. showed substantial behavioural adaptation as task demonstrations were supplied within the context, while explicitly noting that this occurred without gradient updates or fine-tuning^[2].

This is genuine and useful adaptation.

But it is important to identify where the adaptation resides.

Consider a model that recommends an action because of assumption A.

Later in the interaction it receives evidence that A is false.

The model may correctly conclude that its previous recommendation should change.

Within the active context, subsequent behaviour can immediately reflect that correction.

If neither the model parameters nor any persistent system state is changed, however, an unrelated future interaction does not inherently inherit the lesson.

This produces a distinction between:

adaptation conditional on presently available information, and

durable adaptation resulting from previous interaction.

Both can affect behaviour.

They are not the same mechanism.

4. Training and Inference

Contemporary foundation models generally acquire their underlying capabilities through optimisation prior to ordinary deployment.

During training, model parameters are adjusted in response to an objective over training data.

Ordinary inference usually operates differently. The trained parameters are used to compute outputs conditional on the available input, but they are not ordinarily rewritten as a consequence of each individual inference episode.

The practical separation between learning and inference is therefore significant.

An interaction can alter the information upon which the model is conditioning without altering the model’s underlying parameters.

This gives rise to two useful categories.

Parametric adaptation

The model itself is changed through parameter updates, such as training or fine-tuning.

Contextual adaptation

The underlying parameterisation remains fixed while behaviour changes because different information is supplied during inference.

The distinction is not absolute across all modern machine-learning research. Continual learning, model editing and other adaptation techniques deliberately investigate mechanisms for changing model behaviour after initial training.

The relevant point is more limited.

Ordinary inference does not require durable modification of the model merely because the model has encountered new information.

5. Context Is Not Durable Learning

Increasing context windows can obscure this distinction.

A sufficiently large context can contain extensive histories, instructions, corrections and examples.

External retrieval extends this further.

Retrieval-augmented generation demonstrated that language generation can combine knowledge contained parametrically within a model with knowledge retrieved from an explicit non-parametric memory^[3].

This is a major capability.

But historical accessibility and adaptation from history should still be distinguished.

Consider a system containing records of one hundred earlier predictions.

Some were accurate.

Some were wrong because their underlying reasoning was poor.

Some were wrong because the environment changed.

Some were wrong because the available evidence was incomplete.

Others may merely reflect noise.

Possessing all one hundred records does not itself answer:

This suggests a distinction between two forms of persistence.

Informational persistence

Earlier information remains available to subsequent computation.

Adaptive continuity

Earlier activity or its consequences produce a durable difference in subsequent judgement or behaviour.

The distinction can be summarised as:

External memory can clearly participate in system-level learning.

Indeed, Reflexion demonstrates precisely that stored linguistic feedback can alter later decisions without modifying model weights^[5].

This strengthens rather than weakens the distinction.

The adaptation belongs to the behaviour of the wider architecture across episodes. It cannot be inferred merely from the capability of the underlying model.

6. The Temporal Problem

Many consequential forms of intelligence operate across time.

A prediction is made now.

Reality supplies evidence later.

A decision is taken today.

Its consequences may become apparent weeks afterwards.

An assumption may initially appear justified and only become questionable after repeated outcomes.

This temporal structure matters because the quality of a decision often cannot be determined at the moment the decision is made.

Let:

represent a prediction made at time t, and:

represent a subsequently observed outcome.

The significance of the original prediction may become apparent only once the prediction and outcome can later be compared.

A sufficiently capable LLM can analyse that comparison whenever the relevant information is supplied.

That is a question of inference capability.

A separate question is:

That is a question of adaptive continuity.

The distinction concerns not whether a model can understand experience, but whether past interaction becomes consequential to future cognition.

7. The Limits of Scaling

Foundation models have improved dramatically with advances in architecture, data, computation and training.

It is reasonable to expect further progress.

Future models may reason more reliably, use tools more effectively, process richer modalities and operate over substantially larger effective contexts.

None of these developments should be understated.

The issue is logical rather than predictive.

Suppose a future model possesses substantially greater inference capability than a present model.

Given the same problem, it may produce a much better answer.

Given the same history, it may interpret that history more effectively.

Given access to tools, it may choose and use those tools more competently.

None of these facts alone establishes that the system’s hundredth episode is different because it experienced the preceding ninety-nine.

Greater inference capability does not, by itself, imply greater adaptive continuity.

This does not mean the two cannot advance together.

They can.

Nor does it imply that increased capability is irrelevant to continuity. A more capable model may make a system-level adaptation mechanism substantially more effective.

The claim is simply one of non-equivalence.

Improvement in one dimension is not logically sufficient evidence of improvement in the other.

8. Beyond Static Benchmarks

Evaluation practice reinforces the importance of this distinction.

Many widely used model evaluations present a bounded task and measure whether the model produces a successful response.

Such benchmarks can measure important capabilities exceptionally well.

But the evaluation structure commonly asks:

rather than:

These are different questions.

The field is increasingly moving beyond purely isolated evaluation.

AgentBench, for example, evaluates LLM-based agents in interactive environments and identifies long-term reasoning and decision-making as significant challenges^[7]. SWE-bench similarly evaluates models on real software-engineering problems requiring interaction with substantial codebases and execution environments^[8].

Interactive and long-horizon evaluations represent an important development because they expose properties that isolated response evaluation cannot.

Even here, however, successful execution across a long trajectory should be distinguished from durable adaptation between trajectories.

A system may operate over many steps without necessarily changing because of the consequences of previous completed episodes.

Longer episodes and continuity across episodes are related, but not identical.

9. Capability and Adaptive Continuity

It is therefore useful to treat capability and adaptive continuity as separate dimensions.

Capability

Capability concerns what a model or system can accomplish when confronted with a task.

It includes reasoning, abstraction, linguistic competence, planning, coding, perception and other forms of cognitive performance.

Adaptive continuity

Adaptive continuity concerns whether earlier interaction produces durable changes that affect later cognition or behaviour.

A highly capable model can possess limited adaptive continuity when deployed without persistent state.

A less capable model could participate in a system that retains substantial cross-episode state.

The existence of one therefore does not establish the other.

This distinction becomes increasingly important as models improve.

Weak models make their limitations obvious.

Highly capable models create a different problem: exceptional performance within an interaction can make the boundary between sophisticated inference and accumulated experience increasingly difficult to perceive.

10. The Strongest Objection

The strongest objection to this argument is straightforward.

Perhaps adaptive continuity is simply an engineering problem around the model.

Persistent databases can store earlier events.

Retrieval can recover relevant information.

Agent architectures can maintain state.

Language feedback can influence future trials.

Model parameters can be updated where appropriate.

If so, why regard continuity as a limitation of LLMs at all?

The objection is substantially correct.

LLMs can participate in systems that possess forms of continuity.

ReAct demonstrated the value of interleaving model reasoning with environmental action^[4]. Reflexion demonstrated behavioural improvement across trials using persistent linguistic feedback rather than weight updates^[5]. Generative Agents demonstrated an architecture in which observations persisted and subsequently influenced reflection and planning^[6].

These systems show why the model/system distinction matters.

They do not demonstrate that increasing the underlying model’s inference capability automatically produces those properties.

They demonstrate that additional system-level mechanisms can produce behaviour not implied by standalone inference.

The argument of this paper therefore survives its strongest objection by becoming more precise.

It is not that continuity lies beyond LLMs.

It is that continuity lies beyond capability alone.

11. Continuity Is Not Automatically Improvement

There is a second important qualification.

Continuity is not inherently beneficial.

A system can retain false information.

It can learn the wrong lesson from an unusual event.

It can preserve obsolete assumptions.

It can become more confident in a mistaken generalisation.

It can adapt to noise.

A system capable of durable change is therefore not necessarily a system capable of durable improvement.

This distinction matters because descriptions of continuously learning systems can easily assume that more experience necessarily produces greater competence.

It does not.

Experience can improve future behaviour only if the resulting adaptation is useful.

The conceptual relationship is therefore not:

but:

and:

Adaptive continuity should therefore be regarded as a property of a system, not as evidence of its quality.

12. What LLMs Have Already Changed

None of this diminishes the significance of large language models.

The opposite conclusion may be more appropriate.

The Transformer and its descendants have demonstrated that a common learned computational substrate can support an extraordinary breadth of cognitive functions^{[1, 2]}.

Retrieval can extend the information available to that substrate^[3].

Tool use can connect it to external environments^[4].

Persistent memory can influence behaviour across episodes^{[5, 6]}.

Agent benchmarks increasingly test these systems in interactive settings^{[7, 8]}.

The relevant boundary is therefore not between “intelligent LLMs” and some categorically different future technology.

Foundation models may remain central components of increasingly sophisticated intelligent systems.

The important distinction is instead between the capability of the underlying model and the properties of the complete system in which that capability operates.

This distinction prevents two opposite errors.

The first is to underestimate LLMs because they do not independently possess every property associated with intelligence.

The second is to assume that improvement in model capability automatically supplies those missing properties.

Neither conclusion is warranted.

13. Discussion

The trajectory of foundation models has naturally encouraged capability-centred accounts of progress.

Can models solve harder problems?

Can they reason for longer?

Can they operate tools?

Can they outperform people on more benchmarks?

These are important questions.

But intelligence operating across time introduces another class of question.

Does history matter to the system?

Not merely because historical information can be inserted into the next prompt, but because previous interaction produces some persistent difference in subsequent behaviour.

That distinction is likely to become more important as individual model calls become increasingly sophisticated.

A model capable of expert-level reasoning across many domains may appear substantially more continuous than a weaker model simply because it can reconstruct context, infer missing information and behave coherently when supplied with historical records.

Such competence can narrow the behavioural difference between reconstructed continuity and persistent adaptation.

It does not eliminate the conceptual distinction.

This paper does not claim that adaptive continuity is necessary or sufficient for artificial general intelligence.

It does not propose an architecture for achieving it.

Nor does it claim that existing system-level approaches are incapable of providing meaningful forms of continuity.

Its claim is narrower and more defensible:

Consequently:

14. Conclusion

Large language models have dramatically expanded the boundaries of machine capability.

They can reason, synthesise, program, plan, retrieve information, operate tools and participate in increasingly sophisticated agent systems.

Further progress may make these capabilities substantially more powerful.

But capability should not be confused with adaptive continuity.

A model can reason about an earlier event without having changed because of that event.

It can adapt within a context without carrying that adaptation beyond the context.

A wider system can preserve information and use it later, but doing so introduces a property of the system that does not follow automatically from increasing the underlying model’s inference capability.

This is not an argument against LLMs as a foundation for future intelligent systems.

It is an argument for greater precision about what advances in LLM capability actually establish.

The question:

remains fundamental.

But for intelligence operating across time, another question matters:

The first concerns what intelligence can do.

The second concerns whether its history matters.

They are not the same question.

15. Commercial Conclusion

For a business, the practical consequence is straightforward. An organisation can buy increasingly capable AI, and it can equip its people with the same tools. Both are sensible, and both are often available to competitors on broadly similar terms.

The question that remains is what accumulates through the work: whether the second month benefits from the first, and whether the organisation is more capable at the end of an engagement than it was at the start.

That is the basis on which Eudira asks a business to allocate the next unit of work to Synthetic Labour, and it is what the resulting cognitive capital is meant to represent.

Discover Synthetic LabourExplore Cognitive Capital

References

1. [1]Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł. and Polosukhin, I. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems 30.
2. [2]Brown, T. B. et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems 33.
3. [3]Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-t., Rocktäschel, T., Riedel, S. and Kiela, D. (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems 33.
4. [4]Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K. and Cao, Y. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. International Conference on Learning Representations.
5. [5]Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K. and Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. Advances in Neural Information Processing Systems 36.
6. [6]Park, J. S., O'Brien, J. C., Cai, C. J., Morris, M. R., Liang, P. and Bernstein, M. S. (2023). Generative Agents: Interactive Simulacra of Human Behavior. Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology.
7. [7]Liu, X. et al. (2024). AgentBench: Evaluating LLMs as Agents. International Conference on Learning Representations.
8. [8]Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O. and Narasimhan, K. (2024). SWE-bench: Can Language Models Resolve Real-World GitHub Issues? International Conference on Learning Representations.