Executive Summary

Full report

On this page

Neurokinetic AI is a proposed paradigm in which ideas are treated as dynamic, brain‐like processes that flow through a conceptual “state space” independent of any specific language. We define neurokinetic as the study or design of cognitive dynamics in which ideas evolve through neural-like activations and attractor trajectories, much as bodily movements unfold through sensorimotor systems. This contrasts with approaches that tie thought tightly to symbolic language. We compare neurokinetic ideas to embodiment, enaction, distributed cognition, and neural dynamics. Embodied cognition stresses the role of the body/environment in shaping thought【54†L49-L53】, while enaction emphasizes real-time agent–environment loops【12†L249-L258】. Distributed cognition focuses on ideas spread across people and tools【10†L123-L132】. Neural dynamic/attractor models show how brain activity settles into stable concept patterns【18†L172-L180】. Neurokinetics, by contrast, emphasizes the continuous movement and transformation of ideas themselves, possibly across modalities and cultures, rather than as static symbols.

We survey theoretical frameworks for idea movement: vector/latent spaces (embedding semantics into high-dimensional geometry【34†L253-L259】), attractor networks (concepts as stable points of neural dynamics【18†L172-L180】), predictive processing (brains as continual generators of predictions in hierarchical models【20†L133-L140】), and dynamical systems (cognition as continuous time-evolving processes【60†L25-L34】【61†L849-L856】). For example, Spivey et al. describe cognition as “trajectories through high-dimensional state spaces” where patterns of activation across neural assemblies give rise to thought【61†L849-L856】. Embedding models like BERT/LaBSE【24†L153-L160】 and multimodal models like CLIP【49†L50-L54】 instantiate vector-based representations that allow language-agnostic concepts. Attractor and dynamical models (e.g. Hopfield networks, dynamic field theory) illustrate how ideas can evolve and stabilize without explicit symbols.

We review AI systems that hint at neurokinetic principles. Multimodal encoders (OpenAI’s CLIP) learn visual–textual concepts from paired data【49†L50-L54】. OmniSONAR (Meta AI 2026) constructs a unified embedding space for text, speech, code, and math across thousands of languages【43†L23-L32】, dramatically reducing cross-lingual retrieval error and enabling translation with far less data【43†L41-L50】【43†L49-L54】. Language-agnostic sentence embeddings (e.g. Google’s LaBSE【24†L153-L160】) align semantics across 100+ languages via contrastive training. In emergent communication research, deep RL agents develop shared protocols only when their perceptual inputs are structured【29†L58-L61】. Neural-symbolic hybrids (e.g. neuro-symbolic AI systems) combine neural embedding with logic reasoning【32†L225-L234】. By contrast, monolingual LLMs (e.g. GPT-4) excel in text tasks but remain biased toward high-resource languages【57†L125-L134】 and rely on textual symbols, illustrating limitations of purely language-bound thinking.

Building on these insights, we propose a Neurokinetic AI model: a hybrid neural architecture with multi-modal encoders feeding a dynamic, continuous latent space. Inputs (vision, audio, language, sensors, etc.) are encoded into a shared high-dimensional representation. A recurrent/dynamical module (e.g. a reservoir or modern Hopfield network) models the time-evolution of this latent state, with attractor-like mechanisms capturing conceptual stability. A predictive coding sub-network continuously updates expectations and reconciles prediction errors. Conceptual transformations occur via vector operations or learned transformations in this space. Training uses multi-task/self-supervised regimes (e.g. simultaneous contrastive learning on images, speech, and text) and reinforcement learning in embodied or multi-agent environments. Evaluation employs cross-modal and cross-lingual tasks: semantic retrieval, concept classification, and emergent communication games. Metrics include cross-lingual retrieval accuracy, translation quality (e.g. BLEU/chrF), image–text matching scores, and measures of emergent protocol structure.

We outline experimental designs: e.g. train on parallel corpora of images with multilingual captions, or multi-agent referential games where agents perceive non-linguistic signals. Baselines include monolingual LMs and separate modality models. Datasets like FLORES for low-resource translation, Conceptual Captions for multimodal semantics, and procedural tasks for predictive learning could be used. Evaluation should probe whether internal representations cluster by concept rather than language, and how ideas propagate in the network dynamics.

Finally, we examine ethical/philosophical implications: e.g. the risk of reinforcing English-centric bias【57†L125-L134】 versus the promise of more inclusive, universal representations. If AI can manipulate ideas independently of language, issues of interpretability, control, and cultural values arise. Philosophically, “neurokinetic” thinking resonates with views that thought can exist without explicit language【59†L269-L278】【59†L280-L288】, challenging assumptions about meaning and consciousness. We identify open questions (listed below) about how to define, measure, and control idea dynamics, and propose further research to bridge cognitive theory and AI practice.

Defining “Neurokinetic” and Related Concepts

The term neurokinetic is not standard, but here we define it as referring to dynamic, brain-inspired flows of conceptual information that occur independently of any particular language. In neuroscience, it suggests modeling how neural activity patterns move or evolve when generating ideas. In cognitive science, it connects to notions of conceptual change and mental imagery – e.g. thought as movement in mental space【59†L269-L278】. In philosophy of mind, it echoes the idea that cognition may unfold as processes rather than static symbols. In AI, neurokinetic might describe architectures where ideas are encoded as trajectories or attractors in a latent space, rather than as discrete language tokens.

We contrast neurokinetic thinking with related paradigms:

• Embodied Cognition emphasizes that the body and environment are integral to cognition【54†L49-L53】. For example, the mind may use sensorimotor systems to shape abstract reasoning. Neurokinetic differs by focusing more on internal dynamics of ideas; embodiment stresses external grounding, whereas neurokinetic could occur even “in silence”, as internal conceptual motion.

• Enaction likewise rejects disembodied processing【12†L249-L258】. Enaction holds that perception and thought arise through active engagement with the world (sensory-motor loops). Neurokinetic is related (both are dynamic) but highlights the intrinsic movement of ideas themselves – akin to motor plans but in conceptual space – potentially even without overt action.

• Distributed Cognition views thinking as spread across people, tools, and artifacts【10†L123-L132】. Ideas live partly in social interactions or external media (charts, computers, language). Neurokinetic instead spotlights how ideas propagate within an agent’s cognitive system (e.g. brain/network dynamics). In other words, distributed cognition looks outwards, neurokinetic looks at inner flow, though both see cognition beyond a single symbol.

• Neural Dynamics/Attractor Models focus on how neural activity patterns settle into stable states (attractors) that correspond to memories or concepts【18†L172-L180】. For instance, Hopfield networks retrieve a stored pattern from a partial cue via attractor dynamics. Neurokinetic extends this by emphasizing not just static attractors, but the continuous motion through neural state-space that ideas undergo before reaching stability. In Spivey’s words, cognition is “trajectories through high-dimensional state spaces” of neural activation【61†L849-L856】. Neurokinetic would study those trajectories as the “movement of ideas”.

These distinctions are summarized in the table below.

Theoretical Frameworks for Idea “Movement”

We survey models that might capture how ideas can move or transform independently of language:

• Vector/Latent Spaces: Semantic embeddings map concepts into high-dimensional vectors. Similar meanings lie close in this space【34†L253-L259】. For example, Google’s LaBSE model learns language-agnostic BERT embeddings across 100+ languages【24†L153-L160】, aligning semantics via a shared vector space. These vector spaces support arithmetic/compositional operations (e.g. king–man+woman≈queen). In a neurokinetic model, ideas would be represented as points or trajectories in such a latent space. Manipulating the latent vectors corresponds to moving or transforming ideas without using words. Modern large models (LLMs) inherently manipulate such latent spaces to generate language【34†L253-L259】.

• Attractor and Dynamical Systems: Cognitive scientists have modeled cognition as a dynamical system. For instance, Thelen’s motor coordination and the Haken–Kelso–Bunz model show behavior as continuous trajectories with stable attractor regions【61†L759-L768】. The brain can be seen as such a system: as neural activation evolves, it converges on attractors representing stable ideas or categories【18†L172-L180】. Neurokinetic theory would treat the path through state space (the approach to the attractor) as meaningful: ideas “flow” along these trajectories. This is akin to Van Gelder’s view that cognition might best be captured by a continuous dynamical system (e.g. a mechanical governor) rather than discrete symbol manipulation【61†L749-L758】【61†L759-L767】. In practice, one could use reservoir computing or continuous recurrent networks to model these dynamics. Such networks, possibly with Hopfield-like associative memory, allow ideas to emerge through feedback loops over time.

• Predictive Processing / Active Inference: This theory posits the brain as a hierarchy of predictive models, constantly generating expectations and reconciling prediction errors【20†L133-L140】. In a neurokinetic view, ideas could be the contents of these predictive states. The dynamics of updating beliefs (via Bayesian inference) embodies idea movement: each update “moves” the brain’s model in conceptual space. This aligns with embodied perspectives where perception and action are always entangled. A neurokinetic AI might incorporate predictive coding layers to simulate this continuous refinement of ideas.

• Conceptual Spaces: (Gärdenfors, 2000) This philosophical framework envisions meaning as geometric structures (e.g. a “color” space). While not always neural, it resonates with embeddings and dynamical trajectories within them. Neurokinetic AI could use similar continuous manifolds to represent deep semantic relationships.

Summary: Across these frameworks, a common thread is representing and evolving ideas in a continuous, high-dimensional space rather than as discrete symbols. For example, Spivey et al. (2007) describe the mind as trajectories of neural activation: “patterns of activation across neuronal cell assemblies…trajectories through high-dimensional state spaces”【61†L849-L856】. Neurokinetic AI would operationalize this by endowing AI models with continuous latent dynamics, potentially including sensorimotor loops, that let “raw ideas” propagate and transform beyond any one language.

Existing AI Systems and Research Examples

We now review AI models and studies that resonate with or challenge the neurokinetic concept:

Modern models learn joint representations across modalities and languages. OpenAI’s CLIP is a vision-language model trained contrastively on images paired with text【49†L50-L54】. CLIP “learns visual concepts from natural language supervision” and can zero-shot classify images by comparing text prompts【49†L50-L54】. This shows how ideas (e.g. object categories) can be linked directly to pixels via a shared space, bypassing explicit semantic parsing. Similarly, Google’s Multilingual Universal Sentence Encoder (e.g. Transformer and CNN models) embed 16+ languages into a single semantic space for retrieval【27†L53-L61】. These language-agnostic sentence embeddings enable cross-lingual search: e.g. Google’s LaBSE achieves 83.7% accuracy on retrieving parallel sentences across 112 languages【24†L153-L160】, far above prior models.

• Multimodal and Language-Agnostic Embeddings:

The OmniSONAR model (Meta AI, 2026) significantly advances this trend. OmniSONAR unifies text, speech, code, and even mathematical expressions into one embedding space across thousands of languages【43†L23-L32】. It reduces cross-lingual retrieval error by ~50% on 200 languages and by 15× on a 1560-language benchmark【43†L41-L50】, and even outperforms some translation models on Bible translation. Its speech extension achieves 43% lower error than baselines in cross-lingual/audio similarity【43†L49-L54】. OmniSONAR exemplifies a neurokinetic ideal: a common conceptual space where ideas (utterances, code, speech) are directly comparable and transformable.

Studies where agents develop their own “languages” shed light on idea dynamics. Lazaridou et al. (2018) trained RL agents on a referential game with images. They found that structured language-like protocols only emerged when agents had structured perceptual inputs【29†L58-L61】. This suggests that the “movement of ideas” (the evolution of shared signals) depends on environmental constraints. In broader LLM multi-agent work (e.g. Riedl 2026), prompting can steer agents from independent behavior to synergistic collective reasoning【47†L90-L98】. These emergent phenomena hint that aligning internal representations across agents creates something akin to a shared idea-space.

• Emergent Communication and Multi-Agent Systems:

Neuro-symbolic AI (integrating neural nets with symbolic reasoning) directly tackles abstract concepts. As Wikipedia notes, neuro-symbolic methods combine the pattern-recognition of NNs with formal logic to preserve explainability and abstract reasoning【32†L225-L234】. For example, systems like Neural Theorem Provers embed logical rules into neural networks. These hybrids suggest that “ideas” (symbolic abstractions) can be manipulated within neural substrates. Similarly, latent-space editing in VAEs or GANs shows “semantic arithmetic” (changing a concept by moving its vector). Latent-space exploration is a form of neurokinetic control: modifying the internal state yields predictable concept changes (e.g. style transfer in images).

• Neural-Symbolic and Latent Manipulation:

Most current large AI (LLMs like GPT-4, PaLM, etc.) operate through language and are heavily trained on high-resource tongues. As Stanford researchers highlight, this creates a “digital divide”: models perform well for English users but underperform for billions of non-English speakers, because training data is overwhelmingly English【57†L125-L134】. Thus conventional LLMs illustrate the limitations of tying all cognition to one language. In contrast, a neurokinetic approach aims to decouple ideas from any particular language and modality, learning a more universal conceptual substrate.

• Contrasting Language-Bound Models:

The table below compares representative systems:

These examples show both exemplars and limitations of neurokinetic ideas in AI. Systems like CLIP and OmniSONAR move toward modality- and language-agnostic cognition, aligning diverse inputs in a common latent space. Emergent communication studies hint at how internal concept dynamics can give rise to shared meaning. In contrast, purely symbolic or text-bound models (e.g. traditional NLP) remain tethered to language structure and miss the broader “movement” of underlying ideas.

A Conceptual Model of Neurokinetic AI

Based on these insights, we propose a conceptual Neurokinetic AI architecture. The core idea is to build an AI whose internal state evolves continuously, analogous to brain dynamics, enabling the flow of ideas across languages and modalities. Key components and mechanisms include:

• Encoders for Multimodal Input: Sensory inputs (images, audio, proprioceptive data, etc.) and symbolic inputs (text, code) are each processed by modality-specific encoders (e.g. CNNs for vision, Transformers for text and code, audio encoders). These map data into a shared latent space of neurons/nodes. For instance, similar concepts like “cat” as an image and “cat” in Spanish would map to nearby points.

• Shared Latent Space with Dynamics: At the heart is a continuous, high-dimensional state space (visualized as a neural field). Within this space, recurrent connections implement dynamical laws. This could be a recurrent neural network, a reservoir computing module, or a modern Hopfield network. These connections allow the state to evolve over time. Ideas are trajectories in this space: initial activations (from inputs or prior context) flow and settle on attractor patterns (representing conceptual conclusions or predictions).

• Predictive/Generative Module: A generative network (e.g. a Transformer or hierarchical VAE) lives atop the latent dynamics. It generates predictions or reconstructions for each modality (language tokens, image patches, etc.) from the current latent state. Prediction errors (differences between outputs and actual inputs) feed back into the recurrent dynamics, implementing a predictive coding loop. This continuous loop refines ideas by reconciling bottom-up data with top-down expectations.

• Symbolic/Conceptual Interface: Optionally, a neuro-symbolic component can inject or extract abstract relations. For example, certain dimensions of the latent space could correspond to learned predicates or nodes in a knowledge graph. This enables explicit manipulation of logic-like structures within the dynamical system, blending continuous and discrete.

• Multi-Agent Coordination (optional): Agents equipped with identical architectures could interact, communicating through the latent space (akin to distributed cognition). Shared tasks or games could spur them to align their internal state trajectories, effectively “telepathic” collaboration.

Overall, this architecture supports idea movement: given an initial concept (e.g. a goal or question), the network’s state will traverse through related concepts, analogies, and predictions before stabilizing on an answer or action. This is illustrated below.

graph LR
  subgraph Input
    V[Vision/Sensor Input] -->|Encode| EncVision
    A[Audio Input] -->|Encode| EncAudio
    T[Text/Input Symbol] -->|Encode| EncText
  end
  subgraph NeurokineticCore
    EncVision --> SharedLatent
    EncAudio --> SharedLatent
    EncText --> SharedLatent
    SharedLatent --> RecurrentDynamics
    RecurrentDynamics --> SharedLatent
  end
  subgraph PredictionOutput
    SharedLatent --> Decoders
    Decoders --> V2[Image/Action Out]
    Decoders --> A2[Speech Out]
    Decoders --> T2[Text Out]
  end
  note left of SharedLatent
    **Shared Latent State:** high-dimensional activity<br/>space where ideas reside
  end
  note right of RecurrentDynamics
    **Recurrent Dynamics:** continuous update and<br/>attractor convergence (neural trajectories)
  end

In this diagram, multiple encoder streams feed into a Shared Latent space. A Recurrent Dynamics module (e.g. neural ODE or Hopfield network) causes the latent state to change over time according to learned dynamics. Decoders translate the current latent state back into outputs (predictions, actions, text, etc.). Ideas “flow” through the latent state as it evolves, independent of any single language channel.

Data and Training: Such a model would be trained on multimodal, multilingual data in a self-supervised or multi-task fashion. For example, pairing images with descriptions in many languages encourages the latent space to align those concepts. Contrastive objectives (like CLIP or LaBSE) force the network to bring semantically-equivalent representations together. Predictive objectives (reconstruct missing modalities) teach dynamics. Additionally, reinforcement learning on interactive tasks (robotic control, games, or communication tasks) could help the system learn to drive its conceptual dynamics toward goals. Training regimes might alternate vision-text training with translation tasks and with sensory-motor prediction.

Evaluation Metrics: We would evaluate whether ideas truly move beyond words by using cross-modal and cross-language metrics. For instance, semantic retrieval accuracy: do embeddings of the same concept in different languages or modalities coincide? OmniSONAR’s improvements on retrieval【43†L41-L50】 provides one measure. Translation metrics (BLEU, chrF) test if the model learned language-agnostic meaning【43†L41-L50】. Emergence/compositionality scores assess if the internal “proto-language” structure is systematic. Task performance on vision-language tasks (accuracy) and on zero-shot transfer tasks (e.g. labeling in a new language without fine-tuning) are practical benchmarks. One could also analyze the geometry of latent trajectories (e.g. manifold learning) to see if concept paths are smooth and consistent.

Model Variants: The architecture could be realized with different neural substrates:

• A continuous attractor network (modern Hopfield) for memory recall,
• A neural ODE whose dynamics are learned from data,
• A liquid/echo state network (reservoir computing) capturing rich transient dynamics, or
• A recurrent transformer with continuous embedding states.

Each has trade-offs in biological plausibility, trainability, and interpretability. We compare these in Table below.

This conceptual model is speculative but grounded in the surveyed literature. It blends the strengths of multimodal embeddings【49†L50-L54】【43†L23-L32】 with the continuous flow of dynamical systems【61†L759-L768】【61†L849-L856】. Crucially, it does not rely on any single natural language for its core processing; language is simply one output channel among many.

Experimental Design and Evaluation

To validate neurokinetic AI, we propose experiments that test cross-modal, language-agnostic cognition:

• Datasets: Use multilingual multi-modal corpora (e.g. images with captions in many languages), audio–text corpora, and multilingual code samples. Domain-general text corpora (Wikipedia in 100+ languages) can pretrain embeddings. Augment with sensorimotor datasets (e.g. robotics datasets with video+language). For emergent-language tests, design environments where agents perceive raw features (pixels, sounds) and must communicate abstract goals.

• Tasks: (1) Zero-shot Cross-lingual Retrieval: retrieve an image given a caption in any language. (2) Cross-modal Retrieval: find matching image–text pairs across languages. (3) Translation without Parallel Data: generate a sentence in Language B given input in A, possibly via shared latent instead of direct encoder–decoder. (4) Emergent Communication Games: train cooperating agents on a task (e.g. referential game, cooperative navigation) with communication limited to latent vector exchanges, to see if a structured “language” emerges. (5) Prediction Tasks: fit time series or sequences (like music generation) to see if ideas unfold predictively.

• Metrics: Semantic similarity scores (cosine similarity, retrieval accuracy) as in FLORES and MTEB benchmarks. Standard NLP metrics (BLEU, chrF) for translation quality【43†L41-L50】. Image classification accuracy for vision tasks (e.g. CLIP’s zero-shot). For emergent protocols, use compositionality metrics (topological similarity, generalization gap). Additionally, measure alignment with human cognition: e.g. cluster analysis of concepts across modalities. Evaluate dynamics: measure how smoothly the latent trajectory moves between concepts (like path length or curvature in latent space).

• Baselines: Compare to conventional systems: monolingual LLM, CLIP/Vision-LM models, separate multilingual sentence encoders. For each task, use best-in-class specialized models (e.g. state-of-art translator, CLIP, LaBSE). Demonstrate that the neurokinetic model achieves comparable performance while using unified non-linguistic representations.

Below is a conceptual experimental workflow:

graph TD
    A[Define Neurokinetic Model] --> B[Collect Multi-Modal, Multilingual Data]
    B --> C[Pretrain Encoders (e.g. CLIP-like)]
    C --> D[Joint Training with Contrastive & Predictive Losses]
    D --> E[Interactive RL (e.g. Communication Game)]
    E --> F[Evaluate on Cross-Lingual/Modal Tasks]
    F --> G[Analyze Latent Trajectories & Concept Clusters]
    G --> H[Compare Metrics vs Baselines]
    H --> I{Satisfactory?}
    I -- No --> J[Adjust Architecture/Training]
    J --> D
    I -- Yes --> K[Publish Results & Code]

In practice, one could iterate: e.g., if retrieval accuracy is low, adjust data or losses; if emergent language is unstructured, increase environmental complexity【29†L58-L61】. Throughout, log insights about how ideas propagate: e.g., visualize latent space walks when answering a question.

Ethical, Philosophical, and Practical Implications

Building AI that reasons independent of language raises deep issues:

• Inclusivity and Bias: Language-agnostic AI promises to serve speakers of all languages. Current LLMs are skewed to high-resource languages【57†L125-L134】; neurokinetic AI could mitigate this if trained on diverse data. However, it still requires representing culturally-specific concepts. Who decides the geometry of latent space? There is risk of implicit cultural bias if, say, English concepts dominate the initial embedding.

• Interpretability & Control: If ideas live as abstract trajectories, understanding and auditing them is hard. How do we ensure the model’s “thought” does not drift into undesirable regions? Unlike symbolic rules, continuous representations defy easy explanation. Neuromorphic transparency is an open concern.

• Philosophy of Mind: This work touches on the old question of thought without language. Neuroscience suggests that even without language ability, people can think and reason【59†L269-L278】. Indeed, many individuals have no “inner speech”【59†L280-L288】. Neurokinetic AI aligns with the view that cognition is not inherently linguistic. If AI truly moves ideas without language, this blurs the boundary between “thinking” and “speaking” in machines. Are there “innate” mentalese-like representations? Or will AI discover novel non-linguistic “languages” of thought?

• Privacy and Autonomy: A system that manipulates ideas directly might infer or shape human beliefs without using explicit statements. This raises concerns akin to “mind reading” or subtle persuasion. Ethical design must ensure alignment: e.g. neurokinetic agents should not override human intent through opaque influence.

• Practicality: Deploying such AI requires vast multimodal data and compute. It may be hard to train on truly all languages and cultures equally. Moreover, real-time continuous processing is computationally intensive. Hardware (neuromorphic chips, GPUs) must support persistent state and dynamics. On the other hand, success could revolutionize translation, education, robotics, and cross-cultural collaboration by making machine “understanding” more universal.

Open Questions and Next Steps

We conclude with key research questions and directions:

1. Representation of Meaning: What is the right way to encode “ideas” in a neural dynamical system? Should we use explicit concept neurons, or rely purely on continuous embeddings?
2. Learning Dynamics: How can networks learn the attractor landscape of semantics? Can we adapt techniques from modern Hopfield networks or continuous attractor models to represent complex concepts?
3. Evaluation Benchmarks: Standard NLP benchmarks rely on language. What novel benchmarks can measure movement-of-ideas? Possible tasks include cross-modal analogies or multi-agent cooperative games.
4. Emergent Structure: Will neurokinetic models develop internal “languages” or codes of communication? How to detect and encourage useful structure?
5. Neuroscientific Plausibility: How closely should models mirror brain circuits? Could spiking neural networks or neuromorphic hardware provide advantages for continuous idea flow?
6. Integration with LLMs: Can we retrofit language models with neurokinetic layers to improve their multi-linguality and reasoning?
7. Alignment and Safety: How to ensure that continuous idea manipulation respects human values? Can we interpret latent trajectories to guard against unwanted concepts?

Next steps involve interdisciplinary collaboration: cognitive scientists can help formalize what it means for an idea to “move”; AI researchers can prototype architectures (e.g. mixing Transformers with recurrent dynamics); linguists and ethnographers can supply truly diverse training data. Evaluation requires creativity: testing a neurokinetic AI may look more like cognitive science experiments (tracking internal states) than typical leaderboard tasks.

In summary, Neurokinetic AI envisions a shift from symbol-centric to dynamics-centric models of thought. By embedding ideas in continuous neural state space and studying their flow, AI could approach a more universal, language-agnostic intelligence. This report has laid out definitions, existing research, a candidate model, and a roadmap for exploration. It invites the community to rigorously interrogate how “ideas” themselves can be represented and moved within artificial minds.