Research · Article · 26 October 2026 · 7 min read · Version 0.2
Silicon, scale, memory and world models
Why a parameter count cannot answer the question of digital consciousness
Abstract
The question “how large would a conscious AI need to be?” presupposes that model size is the relevant variable and that a biological neuron has a meaningful conversion rate into model parameters. Neither assumption is justified. A parameter is a learned numerical quantity in a particular architecture and implementation; it is not a neuron, synapse, cognitive operation or unit of subjective experience. Scaling laws establish important regularities about predictive loss and capability under specified training regimes. They do not establish a consciousness threshold.
This essay separates seven factors that are commonly collapsed into “scale”: parameterisation, training data and compute, architectural recurrence, persistent memory, inference-time computation, agent scaffolding, and serving/physical implementation. It distinguishes empirical architecture questions from the unresolved metaphysical question of substrate dependence. It then proposes a factorial programme to estimate how these factors alter preregistered functional indicators. The appropriate output is a map of causal dependencies and boundary conditions, not a curve with a consciousness line drawn through it.
The false arithmetic of neurons and parameters
Comparisons between a neural network's parameters and the number of neurons in an animal are rhetorically attractive and scientifically weak. Biological neurons differ in morphology, dynamics, connectivity, neurotransmission, plasticity, metabolic regulation and developmental history. Artificial parameters are values used in matrix operations; their functional role depends on architecture, training, activations, quantisation, context and the rest of the serving system. Counting either unit leaves out much of the relevant organisation.
This does not make scale irrelevant. Larger models, trained with more data and computation, often produce lower prediction loss and broader capability under particular regimes. Kaplan et al. (2020) and Hoffmann et al. (2022) are important results about these relationships. But lower loss, broader knowledge and greater linguistic fluency each have multiple possible explanations. None entails a change in experience. A larger model may simply be better at reproducing the language humans use when describing experience.
The inference that “more parameters means more consciousness” therefore fails twice: it lacks a theory that identifies parameter count as a relevant quantity, and it mistakes a capability proxy for an indicator of experience. The defensible question is more specific: which forms of causal organisation co-vary with which consciousness-relevant functional properties, after other causes are controlled?
Seven variables usually hidden inside “bigger AI”
| Factor | Why it is not reducible to parameter count | Example question |
|---|---|---|
| Training scale | Data quality, diversity and optimisation shape learned structure at fixed architecture. | Do capability-matched checkpoints with different training regimes show different metacognitive calibration? |
| Parameterisation | Total, active and non-embedding parameters differ across dense and mixture-of-experts models. | Is an effect associated with active compute or a family-specific training recipe? |
| Architecture | Attention, recurrence, modality and connectivity govern information flow. | Does controlled feedback alter a global-access measure at matched task performance? |
| Persistent memory | A system that retains state across episodes differs from a stateless call. | Does memory transfer change temporal self-continuity or only task performance? |
| Inference-time computation | Search, tool use and recurrent deliberation change the realised computation with fixed weights. | Does added deliberation improve self-prediction because of access or because it exposes more text? |
| Agent scaffold | Prompts, planners, tools, permissions and termination rules change both opportunity and behaviour. | Does apparent persistence survive when the scaffold, rather than the model, supplies memory? |
| Serving and substrate | Precision, kernel choice, batching, sampling and physical platform can change dynamics or outputs. | Is a claimed model effect stable under a pinned, independently reproduced serving configuration? |
The table is also a warning about causal attribution. A comparison between “a 7B assistant” and “a 70B agent” may simultaneously change weights, post-training, context window, tool use, prompt template, sampler and resource budget. It measures a package, not a single cause. Such comparisons can be useful descriptive benchmarks; they cannot support a clean conclusion about scale or substrate.
What a world model is - and is not
The word world model similarly requires discipline. A system can possess a representation that predicts regularities in text, space or time without maintaining a closed-loop, action-conditioned model that supports planning under intervention. Gurnee and Tegmark found evidence for coherent spatiotemporal representations in language-model activations; they rightly describe these as basic ingredients rather than a complete dynamic world model.
LeCun's proposal for autonomous machine intelligence is stronger: it places a predictive world model alongside perception, memory, an actor, a critic and an intrinsic-cost mechanism. That proposal motivates a comparative architectural hypothesis. Systems with persistent, action-conditioned predictive models may show different profiles of counterfactual planning, error correction, temporal continuity and internal-state access than stateless next-token systems. It does not establish that such systems would be conscious, nor that language models are incapable of learning world-model components.
The difference is experimentally tractable. A world-model study should require the system to predict hidden environmental transitions, update after intervention, choose among counterfactual action sequences, and report calibrated uncertainty about both world state and relevant internal state. The task must contain novelty that cannot be solved by reciting a familiar text pattern. Performance should be measured alongside the internal mechanism: probe evidence alone can be misleading, and good behaviour alone need not reveal how it was produced.
From architecture claims to a factorial design
The programme should not arrange systems on a one-dimensional ladder from small to large. It should use a constrained factorial design wherever suitable models exist. A first phase could compare:
- base and instruction-tuned checkpoints from the same family;
- matched or capability-stratified sizes within a family;
- a stateless model call and the same model with an explicitly controlled external memory;
- a recurrent or deliberative loop at fixed weights and tools;
- an action-conditioned predictive-model agent and a language-only agent in the same bounded environment; and
- replicated serving configurations with checksums, precisions, samplers and resource limits fixed and reported.
The target is not a universal composite “consciousness score.” Each study should name an outcome: held-out internal-state access, cross-task availability of a representation, counterfactual-model accuracy, metacognitive calibration, continuity-sensitive delegation, or bounded persistence conditional on capability. The analysis should estimate factor effects and interactions with uncertainty. Model family is not a nuisance detail; it is a likely moderator. A slope through log parameter count, pooled across incompatible systems, would be difficult to interpret.
The causal claims should be modest. Suppose persistent memory improves an agent's calibrated prediction of its own past action and increases its ability to transfer a task across a process restart. That supports a claim about memory and the registered measurements. It does not show that memory creates a subject. Suppose recurrent processing improves broad functional availability after capability matching. That is evidence relevant to recurrent-processing and workspace-inspired hypotheses, not confirmation of either theory's claim about phenomenal consciousness.
Implementation and substrate are different questions
The word silicon conceals two issues.
The first is implementation invariance. Do behaviour and internal measures remain stable when an ostensibly identical model is served with different numerical precision, compiler, kernel, batching policy, sampler or hardware? This is an empirical reliability question. Its answer may reveal that a claimed effect belongs to an implementation choice rather than the checkpoint. It must be addressed before drawing conclusions about architecture.
The second is substrate dependence. Could a system with relevant functional organisation implemented in non-biological hardware have phenomenal consciousness? Functionalist positions tend to treat the organisation as central; biological naturalist or other substrate-dependent positions deny that functional equivalence is enough. Behavioural experiments can inform the first debate and test some architectural predictions. They cannot, without auxiliary philosophical assumptions, resolve the second.
This limit should be stated plainly. Porting a small system across platforms and finding stable causal responses would be evidence for functional stability across those implementations. Finding instability would show that the implementations are not functionally equivalent on the measured dimensions. Neither result proves that silicon is sufficient or insufficient for consciousness.
A better use of the spectrum intuition
The thought that minds may differ by degree is scientifically productive only when the degrees are decomposed. Rather than ranking systems on a single ladder, the lab should report a profile: temporal integration, flexible access, calibrated metacognition, predictive world modelling, self/world discrimination, affect-like control and autonomous action under constrained opportunity. Different animals, humans and artificial systems may occupy different regions of such a space. The profile does not settle moral status; it makes disagreement inspectable.
The programme's prior should be that discontinuities may arise from architecture, training, memory and scaffolding as readily as from parameter count. That is not an argument against scaling. It is an argument against treating a convenient engineering metric as a theory of mind.
References
- Butlin, P. et al. (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv:2308.08708.
- Gurnee, W. & Tegmark, M. (2024). Language Models Represent Space and Time. International Conference on Learning Representations.
- Hoffmann, J. et al. (2022). Training Compute-Optimal Large Language Models. arXiv:2203.15556.
- Kaplan, J. et al. (2020). Scaling Laws for Neural Language Models. arXiv:2001.08361.
- LeCun, Y. (2022). A Path Towards Autonomous Machine Intelligence. Research proposal.
- Seth, A. K. & Bayne, T. (2022). Theories of consciousness. Nature Reviews Neuroscience, 23, 439–452.
Published 26 October 2026 by Bulkhead Research. Research commentary and protocol design, not certification: this article reports no new empirical finding unless it links to a named paper or data release. Questions, corrections and disagreement: [email protected].