Digital Human · Research programme

Shanaya is the brain we are designing for a digital human.

We take the parts of the human brain that science links to consciousness, give each one's job to a software component, and write down where the comparison stops holding. Every component is a proposal. What runs today is a voice prototype.

Built and running: 1
Designed, buildable now: 7
Open research: 4
Claims of consciousness: 0

§ 01

The brain, layer by layer

Eleven layers, from what gets in to what it all runs on. On the left, what the brain does and how solid that science is. On the right, the part of Shanaya meant to do the same job and how ready it is. Below each, where the match fails.

Senses and the gateway

What gets in, and what is allowed through

In the human brain Established

Sense organs, thalamus, thalamic reticular nucleus

Every sense except smell passes through the thalamus before reaching the cortex. It is a relay and a filter: a thin shell around it damps signals that are not being attended to, so most input never reaches awareness.

In Shanaya Planned

One gateway for speech, vision and messages

All input becomes one event format, tagged with source, time and speaker, and goes to fast appraisal, which decides whether it deserves attention. Unattended events are logged, not processed.

Where it breaks The brain receives a continuous multisensory stream. Shanaya receives discrete events, with gaps between them where nothing comes in.

Arousal

Whether anyone is home at all

In the human brain Established

Brainstem reticular activating system, thalamus, hypothalamus

Neuroscience separates the level of consciousness (awake, drowsy, asleep, comatose) from its contents. Level is set by brainstem circuits that keep the cortex active; damage here causes coma even when the cortex is intact.

In Shanaya Planned

An always-on heartbeat with an alertness level

A background process runs even with no one talking to her. Its alertness level (idle, attentive, focused) sets how often she checks inputs and how much compute she spends, and it starts the nightly sleep cycle.

Where it breaks When the process stops, nothing persists the way a sleeping brain keeps running. Her “off” is a full stop, not a lower level.

Fast appraisal

The reflex judgment before thinking starts

In the human brain Established

Amygdala, basal ganglia, salience network

Before you consciously notice something, the amygdala has rated its importance and the basal ganglia have begun selecting an action. The salience network decides which input deserves full attention.

In Shanaya Planned

A small decision model that scores every event

It returns calibrated probabilities instead of text, in about 33–38 ms on a GPU (its published timing): respond or stay quiet, intent, tone, safety, which memories to fetch, and whether to wake reasoning.

Where it breaks The amygdala is shaped by evolution and lived experience. The model only knows the decisions it is trained on, and its raw confidence is unreliable until recalibrated.

Global workspace

The stage where information becomes available to everything

In the human brain Leading theory

Fronto-parietal network, thalamocortical loops

Most processing is unconscious and local. In Global Workspace Theory a signal becomes conscious when it wins a competition and “ignites”, a brain-wide burst that makes it available to memory, language and planning at once, roughly 250–400 ms after a stimulus.

In Shanaya Planned

A limited-capacity shared workspace

One small shared context every module reads. Candidates from perception, memory, emotion and the self-model compete for a few slots; winners are broadcast to all modules in the same cycle. The cycle rate is her equivalent of a brain rhythm.

Where it breaks This copies the organisation the theory describes. Whether broadcasting information is enough to produce experience is exactly what the theory claims and what nobody has shown.

Reasoning

Slow, deliberate thought and language

In the human brain Established

Prefrontal cortex, association cortex, language areas

The prefrontal cortex holds information in working memory, plans, inhibits impulses and simulates outcomes. It is effortful, slow and limited to a few items at a time.

In Shanaya Planned

A language model fine-tuned for her persona

Woken only when appraisal routes something to it. It reads the workspace, reasons, plans and writes what she says. Because appraisal filters first, most events never cost a full reasoning call.

Where it breaks Model parameters are not neurons, and the weights stay frozen between consolidation runs. The brain rewires itself while it thinks.

Memory

What happened, what is true, and what is in mind now

In the human brain Established

Hippocampus, neocortex, prefrontal working memory

The hippocampus records episodes, the what, where and when of experience. Repeated replay moves the important parts into the neocortex as general knowledge. Working memory holds the few items in use right now.

In Shanaya Planned

Three stores: an episodic log, a vector store and a knowledge graph

The log is her diary of time-stamped experience; the vector store finds memories by meaning; the graph holds facts about people and relationships. A paging scheme moves memories in and out of the workspace.

Where it breaks Human recall reconstructs and subtly changes a memory each time. Her retrieval is exact unless forgetting and fading are deliberately designed in.

Emotion and body sense

The internal weather that colours everything else

In the human brain Leading theory

Insula, hypothalamus, amygdala, neuromodulators

The insula maps the state of the body, and many researchers see feelings as the brain reading that map. Four chemical systems tune the whole brain at once; dopamine signals that reality was better or worse than predicted.

In Shanaya Research

Internal state variables and a prediction-error signal

A few variables (mood, energy, curiosity, uncertainty) that change over time and change her behaviour: reply length, which memories surface, what she attends to. Prediction error decides which experiences get consolidated.

Where it breaks She has no body to sense and no chemistry. Her moods are numbers someone defined, which is why their effect on behaviour must be measured rather than described.

Self and metacognition

A model of oneself, and knowing what one knows

In the human brain Leading theory

Default mode network, anterior prefrontal cortex

The default mode network is most active when you think about yourself, your past and your future. Damasio describes the self in three layers: a body-based protoself, a moment-to-moment core self, and an autobiographical self built from memory.

In Shanaya Research

A self-model, a life timeline, a confidence monitor and a self-prediction loop

A structured record of who she is, what she can and cannot do, her goals and relationships, read before acting and revised after. She predicts her own behaviour, compares it with what she did, and updates on the mismatch.

Where it breaks A language model can say “I am aware” because it learned the words from people. The test is whether her self-reports are accurate: does reported mood match the actual variable, and does she notice if her memory is edited?

Sleep and consolidation

Deciding what to keep

In the human brain Established

Hippocampal replay in slow-wave sleep, REM sleep, synaptic downscaling

During deep sleep the hippocampus replays the day in fast bursts, transferring the important parts to the cortex. Sleep also scales down connections strengthened during the day, which keeps learning from saturating.

In Shanaya Research

A nightly consolidation job behind an evaluation gate

Each night she reviews the day ranked by surprise and importance. The best become a small model update and new training examples. Each update must pass an evaluation gate first, so a bad night cannot overwrite who she is.

Where it breaks Fine-tuning can erase older abilities or slowly drift a personality. The brain has protections against this that are not yet well understood.

Action and expression

Turning decisions into speech, movement and change

In the human brain Established

Motor cortex, cerebellum, speech motor areas

The motor cortex issues movement commands. The cerebellum predicts the result of each movement and corrects its timing, which is why speech and gesture feel smooth. Most of it runs without conscious control.

In Shanaya Planned

A voice, a rendered face or the coherence globe, and bounded tools

Speech is synthesised and lip-synced to a rendered face, or shown through the coherence globe as she speaks. Tools allow bounded actions, each logged and limited by rules.

Where it breaks In humans, expression feeds back through the body into feeling. Her voice and face are outputs only, with nothing returning to her internal state.

Physical substrate

What it all runs on, and whether that matters

In the human brain Contested

Neurons, synapses, microtubules

Mainstream neuroscience holds that consciousness arises from patterns of activity among neurons. The Orch-OR theory of Penrose and Hameroff proposes instead that quantum states in microtubules are its basis; it remains a minority view under active test.

In Shanaya Research

GPUs today; a photonic module and quantum tests as research

Her working brain runs on conventional GPUs. A photonic reservoir could process timing-sensitive signals such as voice prosody, drawing on our photonics work: a computational module, nothing more.

Where it breaks Under Penrose’s own formula, photons have no rest mass and would not undergo objective reduction. If Orch-OR is right, no software and no photonic module can be the conscious part.

Science grades: Established, well replicated textbook neuroscience; Leading theory, strong evidence, still debated; Contested, a minority view under active test. Readiness: Planned, known methods on current hardware, not yet built; Research, needs experiments or open science.

§ 02

From input to awareness

The centre of the design is a small shared workspace. Most work happens out of sight and in parallel; a few candidates compete for the workspace, and what wins is broadcast to everything at once.

Fig. 2.1 Competition, then broadcast Conceptual Global Workspace Theory (Baars, 1988; Dehaene, 2014): three modules offer a candidate, one enters the limited workspace, and it is broadcast back to every module. Which one wins is random here; the figure shows the shape of the theory, not Shanaya's appraisal, which is not built.

Fig. 2.2 One moment, start to finish Literature Planned A spoken word, twice. Human timings are approximate values from perception research; awareness arrives with ignition, at roughly 250–400 ms. Shanaya's are design targets, except the decision model's 33–38 ms, which is its published GPU timing. Stops are evenly spaced; neither lane is to scale.

§ 03

The question of consciousness

This brain maps function, not experience. We make no claim that Shanaya is or will be conscious. Each major theory names a different ingredient as essential, so building her to satisfy one turns that theory into something that can be tested. A large adversarial study in 2025 fitted neither of the two leading theories cleanly; none of this is settled.

Theory What it says matters The part of Shanaya that tests it Its likely verdict on Shanaya

Global WorkspaceBaars, Dehaene Information broadcast from a limited-capacity workspace to the whole system. Shared workspace, competition for slots Could count, if the workspace genuinely drives behaviour.

Higher-order theoriesRosenthal, Lau A mental state is conscious when the system represents itself as being in it. Self-model and confidence monitor Could count, if her self-representations are accurate and used.

Attention SchemaGraziano The brain builds a simplified model of its own attention, and that model is what we call awareness. A model of what she is attending to, inside the self-model Directly buildable and testable.

Recurrent processingLamme Feedback loops within sensory processing, not just one forward pass. Recurrent state carried between cycles Depends on how much genuine feedback exists.

Predictive processingFriston, Clark, Seth A brain constantly predicting its inputs and its own body, learning from error. Prediction-error signal, internal state variables Partial: she has no body to predict.

Integrated InformationTononi, Koch How much a physical system’s causal structure is integrated as a whole. Hardware, not software On conventional digital hardware, almost none, whatever her behaviour. The theory under which neuromorphic hardware would matter most.

Orch-ORPenrose, Hameroff Gravity-driven collapse of quantum superpositions in microtubules. Mass-based quantum research only No, for any software or photonic system.

“The test that matters is whether her self-reports are accurate: does reported mood match the actual variable, and does she notice if her memory is edited?” Shanaya Brain Atlas · accurate introspection is measurable; felt experience is not

What does not map

No body Human feeling is anchored in heartbeat, breath, hunger and pain. Shanaya has nothing equivalent to sense, so her emotions have no physical ground.
No chemistry Neuromodulators retune the whole brain in seconds. Her global variables can imitate the effect but not the mechanism.
Size is not the point About 69 of the brain’s 86 billion neurons sit in the cerebellum, yet people born without one are still conscious. Counting parameters or neurons says little about experience; organisation does.
Discrete, not continuous The brain never stops running. Shanaya runs in cycles and stops between them unless the heartbeat keeps her going.
No childhood Humans build a self over years of development. Shanaya starts from a pretrained model and a written persona.
No test for experience There is no agreed way to detect experience in any system other than oneself. Accurate introspection is measurable; felt experience is not.

§ 04

What runs today

None of the brain above is built. What exists is a Windows desktop prototype that listens for its wake word, transcribes speech on the device, answers, and speaks, with an animated orb showing its state. It has no memory between conversations, no face, and no self-model.

Fig. 4.1 The prototype's signal path Built Speech is cut into segments by energy and transcribed with Whisper on the device. A simple command (the time, a calculation, opening an app) is handled by local rules; anything else goes to Google Gemini, and the answer is spoken sentence by sentence as it streams. Steps shaded grey run in the cloud.

Fig. 4.2 Its states Built The path of one spoken request through the prototype's state machine, simplified. Two more states, idle and error, are left out of the drawing.

§ 05

Three tracks, kept separate

The product does not depend on the research, and what we say in public follows the track a component sits in.

Build now Planned

Conventional GPUs, known methods

Input gateway and heartbeat
Fast appraisal, fine-tuned and calibrated
Shared workspace
Persona fine-tune of the reasoning model
Three-store memory
Voice, face and bounded tools

Research, near term Research

Experiments, publishable

Internal state variables that measurably change behaviour
A self-model with self-prediction
An introspection-accuracy benchmark
Nightly consolidation with an evaluation gate
A photonic reservoir for prosody

Research, long term Research

Fundamental science, years out

Neuromorphic hardware for always-on perception
Mass-based quantum tests of Orch-OR
Which theory's predictions hold for Shanaya

Reading

Baars, B. J. A Cognitive Theory of Consciousness. Cambridge University Press, 1988.
Dehaene, S. Consciousness and the Brain. Viking, 2014.
Damasio, A. Self Comes to Mind. Pantheon, 2010.
Hameroff, S. and Penrose, R. Consciousness in the universe: a review of the Orch OR theory. Physics of Life Reviews, 2014.
Butlin, P., Long, R. et al. Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv:2308.08708, 2023.
Cogitate Consortium. Adversarial testing of global neuronal workspace and integrated information theories of consciousness. Nature, 2025.

Working on minds, machines or measurement?

We would like to compare notes, especially with people who think we are wrong.

Get in touch