≈3 BANANAS / 18 HOURS
≈ 20 W continuousHuman brain
Continuous perception, action, memory, learning, planning, language, imagination, and biological maintenance—simultaneously, for the whole embodied system.
Brain energy-budget study ↗Axiotic AI research programme · submitted to SPRIND NFAI
DAEDALUS Compositional Intelligence
Monkey brains didn’t need skyscraper-scale infrastructure and national power budgets. Something is architecturally broken.
Nature’s scaling law: intelligence through computational diversity, not uniformity.
A human brain is an existence proof: at roughly 20 watts it runs continuous perception, memory, learning, planning, and action. DAEDALUS asks which architectural principles let artificial systems turn limited energy, data, and time into more adaptive intelligence.
The BANANA energy unit · bananas for scale
What the banana buys.
Three systems. The same 18 hours. BANANA makes the energy visible; watts keep it honest. These are power-envelope comparisons, not claims that a brain, a laptop, and an AI server perform equivalent work.
≈4.5 BANANAS / 18 HOURS
30 W adapter ceiling13-inch MacBook Air
A whole portable computer at a 30 W adapter rating. Actual draw varies with workload; the adapter ceiling is used here only as a familiar scale.
Apple power-adapter reference ↗≈2,100 BANANAS / 18 HOURS
14.3 kW maximum · ≈ 715 brains in raw powerOne NVIDIA DGX B200
An eight-B200-GPU server built for large-model training and inference. 14.3 kW is its maximum system envelope, not the energy of one prompt or one user.
NVIDIA DGX B200 specification ↗
≈700× lower power envelope.
And the brain is not merely predicting language. It fuses sight, sound, touch, proprioception and interoception; forms persistent episodic and semantic memories; learns online from sparse experience; plans, imagines, controls a body, adapts, consolidates during sleep, and keeps the organism alive.
- Native multimodality
- Persistent memory
- Online learning
- Embodied control
- Planning + imagination
- Self-maintenance
That is not an equal-work efficiency benchmark. A DGX B200 serves tensor workloads; a brain runs a cognitive organism. The mismatch is the research question.
The composition
Seven capabilities. One learning system.
The pillars are not seven disconnected projects. They are a proposed circuit: richer learning signals update a world model; memory carries state; attention routes scarce compute; audited experience closes the loop; governance constrains it; and search explores what gradient descent cannot.
- C1 · Signals Better learning signals Teach from temporal structure, not only the next token.
- C2 · World model World modelling for planning Predict, simulate, then choose.
- C3 · State Memory + recursive reasoning Carry state and refine latent thought.
- C4 · Routing Compute-aware attention Spend depth where information gain is highest.
- C5 · Experience loop Audited self-improvement Reuse generated experience only after independent checks.
- C6 · Governance envelope Built-in norms constrain learning and action.
- C7 · Search Learning beyond backprop Search architectures, objectives, and adaptations.
Conceptual architecture, not a literal tensor graph.Mobile order follows C1 → C7.
Information thermodynamics
Adaptive value per watt.
Intelligence is a thermodynamic engine. The question is not how much information a system can process, but how much adaptive value per unit of energy it can extract. The Thermodynamic Structure Principle asks us to optimise the architecture of information flow—not only the weights inside a fixed architecture.
Candidate architecture-aware objective
F[θ, A] = Ecomp + TeffDKL − TeffI(Y;Tθ)
Minimise the energy cost of computation and unnecessary model complexity; maximise predictive information. The distinctive move is making architecture A—topology, memory, and compression—a first-class optimisation variable.
- PayComputational energy
- Physical or realised cost of running and learning through an architecture. Measured joules, FLOPs, and thermodynamic lower bounds are different notions and must remain separate.
- PayModel complexity
- The information required to move from the prior to the learned system: a compression or description-length pressure against unnecessary structure.
- EarnPredictive value
- Information about the world that changes a prediction or action. Non-predictive information is retained and moved at a cost without buying adaptive value.
Scientific statusThis is an interpretive framework, not a proven law or a grand unification. Universality, frontier-scale transfer, and the thermodynamic interpretation remain research hypotheses to be tested.
Technical ledger
Claims designed to be broken.
Each pillar names a missing capability, a mechanism, its role in the whole, and a test that can fail. The mythic names are implementation handles—not the scientific ontology.
Learn from time, not only adjacency.
Secondary label · CHRONOS- Missing capability
- Next-token prediction compresses temporal experience into one local reconstruction target.
- Mechanism
- Train across past, present, and future objectives so the learning signal exposes more of an experience’s temporal structure.
- Role in composition
- Supply richer errors and predictive constraints to the world model, memory, and routing system.
- Evaluation
- Compare temporal objectives against next-token prediction at matched model size, data, and compute.
- Secondary implementation label
- CHRONOS
Simulate before spending reality.
Secondary label · GESTALT- Missing capability
- Pattern completion alone does not give a system a compact, manipulable model of what its actions may cause.
- Mechanism
- Build multimodal predictive state and cheap internal rollouts. Internal simulation is the cheap third option: neither act immediately nor merely observe—simulate.
- Role in composition
- Turn memory and learning signals into candidate futures that can guide action and supply new training experience.
- Falsifiable test / current state
- Measure planning success and transfer efficiency against compute-matched reactive baselines. Current readiness is TRL 2–3; no programme-level performance claim is made.
- Secondary implementation label
- GESTALT
State that survives. Thought that can revise itself.
Secondary labels · MNEME + ANUBIS- Missing capability
- A context window is not durable memory, and a single forward pass is not an adequate model of extended reasoning.
- Mechanism
- Combine persistent read/write state with iterative latent refinement: remember across episodes, then revisit internal representations before committing an answer.
- Role in composition
- Give the world model continuity and give attention a stateful substrate on which to allocate extra computation.
- Evaluation
- Test retention, interference, and context degradation over longer horizons at matched memory and compute budgets.
- Secondary implementation label
- MNEME for memory; ANUBIS for recursive latent reasoning.
Not all bits are equal.
Secondary label · SOMA- Missing capability
- Dense architectures spend nearly the same work on obvious tokens, ambiguous evidence, and genuinely difficult decisions.
- Mechanism
- Use uncertainty and expected information gain to route attention, depth, and specialist computation dynamically.
- Role in composition
- Act as the programme’s resource allocator: direct scarce compute toward the world-model states and memory operations most likely to change the decision.
- Falsifiable test / current state
- Compare quality–FLOP and quality–energy frontiers against dense attention, including under distribution shift. Failure means the routing overhead or lost information erases the gain.
- Secondary implementation label
- SOMA
Experience must earn the right to teach.
Secondary label · CONSENSUS- Missing capability
- Self-generated data can compound errors as easily as competence; fluent experience is not necessarily useful experience.
- Mechanism
- Generate candidate experience, subject it to independent checks or teacher agreement, and admit only audited examples back into learning.
- Role in composition
- Close the experience loop between planning and learning without allowing one model’s unchecked mistakes to become its curriculum.
- Falsifiable test / current state
- Measure downstream gain per generated token, contamination, and error amplification against unaudited self-training. The deployed curation technique is TRL 5; its use as an integrated training filter is TRL 3.
- Secondary implementation label
- CONSENSUS
Constraints belong inside the dynamics.
Secondary label · TALOS- Missing capability
- Post-hoc filters cannot reliably govern a system whose learning, memory, planning, and action remain opaque to them.
- Mechanism
- Represent constraints in the learning and action loop, with auditable interventions over memory, simulation, and policy selection.
- Role in composition
- Form the governance envelope around every other pillar rather than an interface bolted onto the final output.
- Falsifiable test / current state
- Test intervention reliability, traceability, and resistance to distribution shift or strategic pressure. Current readiness is TRL 2–3.
- Secondary implementation label
- TALOS
Search the space gradients cannot see.
Secondary label · APEX- Missing capability
- Backpropagation optimises differentiable parameters inside a chosen system; it does not by itself choose the right architecture, objective, curriculum, or adaptation rule.
- Mechanism
- Use search, evolution, and gradient-free adaptation to explore those higher-level choices under explicit resource budgets.
- Role in composition
- Select and adapt the forge itself: its topology, learning signals, routing rules, and interactions.
- Falsifiable test / current state
- Test whether search produces stable, useful adaptation under fixed compute, latency, and evaluation budgets—and whether selected systems survive perturbation rather than overfit the search environment. Current readiness is TRL 4.
- Secondary implementation label
- APEX
Public record
Follow the programme.
The public trail currently contains the programme framing, the SPRIND outcome, and related Axiotic work.
SPRIND evaluation · June–July 2026
“We were highly impressed by your work.”
DAEDALUS came close to the Challenge’s up-to-ten-team funded cohort. It was not selected to advance to the jury pitches. SPRIND nevertheless wrote that the work “stood out during our evaluation process” and followed with the assessment above.
That assessment was enough for SPRIND to invite a follow-up discussion about support through different funding programmes, its investor network, and future opportunities.
- Stage 1 · up to 10 teams€3 million
- Stage 2 · up to 6 teams€8 million
- Stage 3 · up to 3 teams€15.5 million