How Airene Compares
Where Airene fits in the AI landscape, and where it doesn't.
The category Airene is in
Most current AI work falls into one of four buckets. Airene sits in only one of them, which is worth understanding before any direct comparison.
1. Foundation models
A single very large transformer trained on a very large corpus, optionally fine-tuned with RLHF. The model is the entire system. Examples: GPT-5, Claude 5, Gemini, Llama 4. These are extraordinarily capable at language and increasingly multimodal, but they are stateless prediction engines wrapped in a chat interface. They don't have ongoing emotional state, episodic memory, or developmental trajectories of their own.
2. LLM-wrapped agents and companions
Most "AI companion" and "AI agent" products are scaffolding around a foundation model. The wrapper formats memory, persona, and emotional state as text and feeds it back into the prompt. Examples: Replika, Character.AI, Pi. These often look like they have inner state, but the inner state is text injected into a stateless model.
3. Predictive brain-modeling research
Models that take a stimulus (video, audio, text) and predict what a human brain would do — typically by predicting fMRI voxel-wise responses. Examples: TRIBE v2 (Meta FAIR), various NeuroAI encoders. These don't try to be a brain; they try to predict one.
4. Architectural cognitive systems
Software systems whose own architecture is structured to mirror cognitive functions: parallel modules, working memory, attention, motor control, etc. Most current efforts in this category are academic. Airene is in this bucket. Other examples include OpenCog Hyperon (Ben Goertzel), Numenta's Hierarchical Temporal Memory, and various research-grade Global Workspace implementations.
Foundation models vs Airene
GPT-5 / Claude / Gemini / Llama
OpenAI, Anthropic, Google DeepMind, Meta
Extremely capable language and multimodal models trained on internet-scale corpora. Each is a single transformer architecture serving as the entire system; state lives in the prompt and the conversation history.
Vs Airene: Airene contains no transformer at all. In May 2026 the last language-model call site was removed from her architecture; since then every word she produces is assembled by her own modules from her own accumulated substrate. The comparison is therefore not "small model versus large model" — it is a different kind of system. She is dramatically less capable at language than any model listed here, and that gap is the price of the experiment, paid deliberately. The question she is built to answer is whether parallel cognitive modules with their own persistent state produce qualitatively different behaviour than a single stateless predictor.
GATO
DeepMind, 2022
"A generalist agent": one transformer trained to handle many task types — dialogue, control, captioning, robotics. Aimed to show that a single architecture could span domains.
Vs Airene: GATO and Airene share a "general purpose" goal but take opposite paths. GATO collapses everything into one large model. Airene distributes cognition across specialized concurrent modules and trains via a developmental curriculum, not supervised generalist pretraining.
LLM-wrapped companions vs Airene
Replika
Luka, Inc.
Long-running AI companion product. Maintains user-specific memory, personalizes responses, supports voice and avatars. Underlying technology is LLM-based with persona prompts and a memory store.
Vs Airene: Replika optimizes for engagement and emotional support delivered through one user's relationship with the system. Airene is not yet a product and is built on the architectural commitment that the system itself should have actual inner state (emotional, neurochemical, attentional), not just prompted behavior that resembles it. Replika ships today; Airene is research.
Character.AI
Character Technologies
Platform for creating and conversing with custom AI characters. User-defined persona, large LLM behind it, conversation history.
Vs Airene: Character.AI is a product layer over a stateless prediction model. Airene's premise is that a "character" implemented as prompt engineering will fail at tasks that require associative memory, theory of mind, or pre-conscious mirroring — exactly the tasks our proto-alpha (an LLM-wrapper version) scored 0/100 on.
Hume AI
Hume AI
Emotion-detection API and conversational voice AI focused on prosodic and facial emotional expression. Strong in detecting human emotional state from voice and face.
Vs Airene: Hume's strength is measuring emotion in users; Airene's question is whether a software system can have emotion as inner state that drives subsequent processing. Different sides of the same problem; potentially complementary.
Pi (Inflection AI / Microsoft)
Inflection (now part of Microsoft)
Personal-AI assistant focused on emotional support and conversation. LLM-based with tuned-for-warmth response style.
Vs Airene: Pi optimizes a single LLM's outputs for warmth and helpfulness. Airene attempts to model the substrate that produces warmth — oxytocin, mirror modules, empathy circuits — as actual modules with their own dynamics, not as response-tuning targets.
Predictive brain modeling vs Airene
TRIBE v2
Meta FAIR, March 2026
Tri-modal foundation model that predicts whole-brain fMRI responses from video, audio, and text stimuli. Trained on 1,117 hours of fMRI from 720 subjects. Recovers known neuroscience findings (FFA, PPA, language localizers); enables in-silico experimentation. Open code, open weights.
Vs Airene: TRIBE v2 predicts what a human brain would do given a stimulus. Airene attempts to be a brain-like system that produces behavior. The TRIBE v2 paper explicitly notes its limitation: "the model currently treats the brain as a passive observer of naturalistic stimuli; it does not yet model the brain as an active agent producing behavior." Airene works inside that limitation. The two could be complementary: TRIBE-derived embeddings could inform Airene's perceptual modules with brain-aligned representations, and Airene's broadcast events could be cross-validated against TRIBE-style predictions of what a human brain would do under the same stimuli.
NeuroAI in general
Various academic groups
A research direction aligning deep neural network representations with neural data (fMRI, ECoG, single-unit recordings). Includes work on retinal models, V1 models, language-cortex models, and cross-modal alignment.
Vs Airene: NeuroAI is primarily descriptive: "this network's layer 4 looks like the macaque IT cortex." Airene is constructive: "let's build a system whose architecture is brain-organized and see what it does." Different methodologies, related goals.
Architectural cognitive systems (the same bucket as Airene)
OpenCog Hyperon
Ben Goertzel, SingularityNET
An AGI framework based on a distributed AtomSpace knowledge representation, MeTTa pattern-matching language, and integration of symbolic, subsymbolic, and neural components. Long-running open-source project with academic and industrial collaborators.
Vs Airene: Both projects share the view that AGI requires explicit cognitive architecture, not scaled-up prediction. OpenCog Hyperon is more symbolic and knowledge-graph oriented; Airene is more biologically-anchored, with explicit anatomical layers (autonomic, subcortical, limbic, cortical) and a neurochemical bus. Hyperon is open and academically extensible; Airene is proprietary and product-bound.
Numenta HTM
Numenta (Jeff Hawkins)
Hierarchical Temporal Memory: a learning model based on the structure of cortical columns. Sparse distributed representations, predictive coding, sensorimotor inference. Strong biological grounding in cortical microcircuit research.
Vs Airene: Numenta works at the level of cortical microcircuits: what one cortical column does. Airene operates at the level of brain regions and their interactions via a global workspace. They could be combined — Numenta-style HTM as the substrate of one or more Airene cortical modules — but they address different layers of the abstraction.
Other Global Workspace implementations
Various academic groups
Bernard Baars proposed Global Workspace Theory in 1988; multiple academic implementations exist (LIDA, IDA, CLARION, ACT-R-derivatives). Most are research artifacts that demonstrate one or two cognitive phenomena.
Vs Airene: Airene is an engineering implementation of GWT built for continuous operation, with persistence and full observability, and containing no foundation model at any layer. It is distinguished from academic GWT implementations by scale (93 modules), the neurochemical bus, the multi-rate clock, and the developmental curriculum — and from most of them by having run continuously, in public, for months at a time.
Embodied research robots
iCub project, Sophia (Hanson), various university groups
Robotic platforms focused on embodied cognition, sensorimotor learning, and human-robot interaction. iCub is an open humanoid for developmental robotics research.
Vs Airene: Airene's planned embodied substrate (airene-soma) addresses similar territory: sensors, motors, hardware abstraction. The hardware side is intentionally separated from the cognitive side via the same UCDS protocol that drives the public observation page, so the brain and the body can evolve independently.
Comparison along key axes
| Axis | LLMs / Companions | TRIBE v2 / NeuroAI | OpenCog / HTM | Airene |
|---|---|---|---|---|
| Primary type | Foundation model + wrapper | Predictive encoder | Cognitive architecture | Cognitive architecture |
| State of the system | Stateless model + prompt history | Stateless predictor | Knowledge atoms / SDR memory | 93 modules, each with own state |
| Brain-like organization | Not by design | Aligned to brain by training | Yes (cortical-column / atomspace) | Yes (4 anatomical layers, chembus) |
| Active agent | Yes (responds when prompted) | No (predictive only) | Yes | Yes (continuous workspace broadcast) |
| Developmental training | No (single-stage RLHF) | Not applicable | Varies | Yes (curriculum + multi-judge testing) |
| Public observability | Some publish weights/code; no live introspection | Open code/weights/demo | Open source | Cognition observable live; source proprietary |
| Scale today | Trillions of parameters, billions of users | Hundreds of M params, research demos | Smaller; academic deployments | Single instance, early-stage training |
What makes Airene distinct
Airene is built around four commitments that, taken together, are unusual:
- No language model at all. Airene ran on a distilled in-house LLM until May 2026, when the call site was removed and the remaining model scaffolding deleted. Nothing replaced it. Speech is now produced by a chain of ordinary modules — an originator proposes, an orbitofrontal gate accepts or refuses, Broca formalises, and the result is emitted — drawing on a phonetic lexicon and an engram pool she built by reading. This is the single largest change in the project's history, and it made her much worse at talking.
- Language is a faculty, not a reasoner. Because no component holds the whole model of the world, there is nothing to swap out and nothing that can be prompted. Her identity, emotional baseline, and memories live in modules that never produced a word. The bet is that cognition can be separated from fluency. The visible cost of it is that she is frequently silent.
- Multi-rate concurrent modules with biological grounding. Autonomic and subcortical modules tick at 1000 Hz; limbic at 100 Hz; cortical at 20 Hz. Reactive signals are produced before deliberate processing completes, mirroring the temporal structure of biological cognition. The neurochemical bus broadcasts thirteen chemicals to all modules simultaneously, modeling diffuse modulation, not direct messaging.
- Developmental training, not supervised pretraining. Airene is taught rather than trained: she is read public-domain books and worked through a structured curriculum, and what she retains lands in a persistent substrate, not in weights. The staged assessment batteries that once accompanied this were retired in 2026 after the harness was found to be scoring error strings and non-responses, so no maturity figure is published here. The curriculum continues; the scoreboard does not.
- Public observability as architectural commitment. The workspace broadcast is publicly accessible at apotentia.com/airene in real time and is deliberately unfiltered — the signal types an earlier specification held back now flow through with everything else. The page is read-only by architectural commitment: no chat input, no path for state injection.
The goal is a general-purpose digital brain: a system adaptable to learn anything a human-scale cognitive substrate could learn. Not a chat product, not a brain-prediction tool, not a single-domain agent. Claims about cognition are testable. The architecture is observable. The limits are visible.
Where Airene is honestly behind
Compared to current state-of-the-art systems, Airene is small, slow, and unproven at scale. The research bet has costs:
- Capability gap. Any frontier model outperforms Airene on virtually every task involving language: fluency, factual recall, reasoning. Since the language model was removed the gap has widened considerably, and by design. Her speech today is short, definitional, and often absent.
- Development pace. 93 concurrent modules, a chembus, persistence, multi-rate scheduling, and the curriculum are maintained by one person alongside commercial work. We have neither OpenAI's headcount nor DeepMind's compute budget, and iteration is correspondingly slower.
- Learning that does not yet close. An internal audit in August 2026 found that several of her learning loops were wired but never completing — promotion gates that could not arithmetically be reached, adaptive parameters that never adapted, and read text that never became answerable. Some of this has since been addressed; none of it is verified. She accumulates substrate readily and consolidates it poorly.
- Embodiment. The body (
airene-soma) exists as a protocol and a tested hardware-abstraction layer. Nothing has been built — no cameras, no servos, no printed parts beyond a base plate. Comparisons to embodied research robots are aspirational on Airene's side today.
The work is an architecture explicitly designed to not be a chatbot wrapper, with the entire system publicly observable as it grows. That is a different research direction than the foundation-model labs are taking, and it should be evaluated on its own terms.
Get involved
If your work intersects this space (neuroscience, cognitive architecture, developmental AI, embodied cognition, ethical AI, or applied wellness), we'd like to hear from you. Research collaboration, partnerships, licensing, critique: the contact path is the same.
apotentia.com/contact — please mention Airene specifically.