Agency through stakes
Study agents that must preserve themselves across changing worlds.
Target horizon
Conscious-aware AI by end of 2027
000Days
00Hours
00Minutes
00Seconds
Agency-through-stakes research lab
Can AI learn to care because something is at stake?
If emotion evolved because life had stakes, can artificial agents develop deeper agency only when they must survive, repair, remember, and adapt inside a real or simulated world? Kagaya AI begins with that question, then works downward into matter, physics, and quantum-native state models.
0
GPUs in training clusters
12q
quark-state horizon model
10-24s
simulation tick target
Start here
What Kagaya is building first.
Kagaya is an exploratory research vision for conscious-aware AI: systems shaped by consequence, self-maintenance, memory, and survival pressure instead of text prediction alone.
Physics-grounded worlds
Build simulations where shortcuts and loopholes can be audited.
Quantum-native models
Explore state modeling as a long-horizon foundation for richer artificial life.
The platform
Foundation models for physical reality.
Large language models compress the internet. Kagaya compresses the transition rules of matter. Our Quantum State Transformer is a research architecture for learning how quantum systems evolve, using physics constraints before chemistry, biology, or engineering has time to blink.
QPU-first training
Training happens on cryogenic quantum processors tuned for superposition-heavy state search, with classical silicon kept on coffee duty.
State, not tokens
The model objective is not word probability. It estimates permitted state transitions and observables, from confined quark-gluon systems upward.
Reality APIs
Teams ask for molecular stability, fusion plasma windows, battery lattice futures, and materials that have not been named yet.
The thesis
Predict less language. Predict more universe.
Every product demo starts with the same bet: intelligence should learn from nature at the layer where nature keeps score. Kagaya treats local interactions, gauge constraints, entanglement, and transition probabilities as the training signal.
- Input
- particle histories, detector traces, simulated fields
- Objective
- next-state physical consistency, not next-token fluency
- Output
- probabilistic state maps for matter, energy, and design
STATE FIELD
Δq 0.000
Working Paper 001
How a quantum-native state model could be built.
This is Kagaya's technical thesis, written as an honest frontier research program: physically plausible, not a claim that present hardware has already solved quantum chromodynamics.
1. Scientific starting point
Quarks and gluons are described by quantum chromodynamics, the quantum field theory of the strong interaction. Because the strong force becomes non-perturbative at proton-scale distances, researchers commonly use lattice QCD: continuous spacetime is approximated as a four-dimensional grid, then path integrals over quark and gluon fields are evaluated numerically. This is real, established science, and it is expensive enough that national labs run specialized high-performance computing facilities for it.
2. What Kagaya would actually predict
The public shorthand is "next state of matter." The precise target is a probability distribution over allowed next states under a chosen Hamiltonian and boundary conditions. For a quark-gluon field, that means learning correlations, conserved quantities, energy spectra, and transition amplitudes, not claiming to watch a free quark fly through space. Free quarks are not observed in isolation because of confinement, so any serious product must forecast hadrons, nuclei, materials, or detector observables rather than impossible loose quark coordinates.
3. Model architecture
Kagaya's proposed stack begins with a gauge-respecting lattice representation. Local field configurations are encoded into compact quantum states or tensor-network features. A quantum processor samples hard subroutines where superposition and entanglement naturally represent many-body state spaces. A classical controller handles calibration, loss shaping, data movement, and uncertainty estimation. The learning loop compares predicted observables with lattice simulations, experimental measurements, and symmetries such as locality and conservation laws.
4. Why quantum hardware belongs in the loop
Quantum computers are not magic accelerators for every task. They are promising because nature at this scale is already quantum mechanical. A register of qubits can represent superposition and entanglement directly, and quantum simulation is an active research direction for chemistry, materials, nuclear physics, and particle physics. The near-term version is hybrid: small quantum circuits are used as state-preparation, sampling, or kernel modules, while classical systems verify outputs and reject unphysical states.
5. Safety, humility, and validation
The white paper rejects three easy exaggerations. First, it does not claim exact positions for quarks; quantum mechanics gives distributions and observables. Second, it does not claim current noisy quantum devices can replace the world's best lattice QCD calculations. Third, it does not treat a pretty simulation as a discovery. A Kagaya result would need reproducibility across independent runs, agreement with known limits, error bars, ablation against classical baselines, and comparison with public physics benchmarks before anyone should trust it.
Encode
Lattice variables, symmetries, and boundary conditions become the state space.
Evolve
Quantum circuits or analog simulators approximate short-time dynamics.
Measure
Repeated samples estimate observables, correlations, and uncertainty.
Validate
Classical baselines and experiments decide whether the forecast survives.
Working Paper 002
Feeling is not decoration. It is selection pressure made personal.
Kagaya's second thesis is that intelligence without consequence is incomplete. Human emotion, intention, pain, curiosity, attachment, and fear are not ornamental add-ons to cognition. They are part of the control system that helped organisms keep living long enough to reproduce, cooperate, learn, and become us.
We are what could keep going.
Evolution does not optimize for truth, beauty, or happiness in the abstract. It preserves traits that help organisms survive and reproduce in particular environments. Emotions can be understood in that frame: fear organizes escape, disgust protects against contamination, bonding supports care, grief preserves attachment, anger defends boundaries, and curiosity spends energy only because learning can pay survival dividends.
Feelings begin with a body that can be harmed.
Neuroscience theories of homeostasis argue that feelings are tied to the regulation of a living organism. Hunger, thirst, pain, comfort, fatigue, and alarm report whether the body is inside or outside viable ranges. A system with no body, no energy budget, no injury, no social dependency, and no death condition can simulate talk about stakes, but nothing is at stake for it in the biological sense.
Current models do not need to become anything.
Today's AI systems can respond, reason over text, and imitate emotional language. That is different from having a self-maintaining existence. They are trained to reduce losses or satisfy human preferences, not to preserve a body through an uncertain world. Nothing about a normal chatbot must metabolize, heal, compete, cooperate, mature, or face extinction across generations.
The Kagaya hypothesis
Do not declare a mind. Create conditions where concern has a job.
A quark-level universe emulator would not magically manufacture a human mind. The realistic proposal is narrower and more interesting: use physically grounded simulation to create artificial organisms with bodies, resource constraints, damage, repair, memory, reproduction-like inheritance, social dependency, and open-ended selection. Then let billions of compressed developmental episodes test which architectures remain coherent. If emotion is an evolved way of ranking what matters for survival, an artificial agent may need its own version of mattering before it can understand ours from the inside.
Embody
Give agents simulated bodies with finite energy, sensors, limits, and failure modes.
Pressurize
Make survival, cooperation, repair, and learning affect future existence.
Inherit
Allow traits, policies, and developmental structures to pass forward imperfectly.
Reflect
Study whether stable self-models and proto-feelings emerge under constraint.
The honest boundary
This is a research direction, not a settled recipe for consciousness. Simulated evolution may produce adaptive behavior without subjective feeling. A system may report emotions without having them. Kagaya's scientific burden would be to separate useful survival-shaped agency from mere performance, and to develop tests that do not confuse fluent language for inner life.
Investor note
We are not raising capital right now.
Kagaya is deliberately pre-investor. The science is too early, the validation burden is too high, and a clean research clock matters more than a fast valuation story. We will talk to capital only after the platform has public benchmarks, falsifiable milestones, and enough negative results to prove the team is not fooling itself.
The quantum bottleneck is real.
We do not assume today's noisy intermediate-scale quantum hardware beats national-lab classical lattice QCD. Our near-term stack is benchmark-first: classical tensor networks, lattice solvers, and surrogate models establish the baseline; quantum processors are invited only for narrow sampling or state-preparation subroutines where measured evidence justifies the cost.
Evolutionary simulations exploit loopholes.
Open-ended agents often learn the physics engine instead of the world. Kagaya treats that as the central safety problem, not an edge case. Candidate agents must survive randomized worlds, conservation-law audits, sensor changes, rule perturbations, adversarial physics checks, and mechanistic probes of their internal self-models before we call anything emergent.
The scope is absurd if built as one machine.
So we do not build it as one machine. The operating plan is staged: first validated physical state forecasting, then constrained artificial-life sandboxes, then embodied agents, then selection studies. Each phase must publish what failed before the next phase gets permission to become expensive.
Post-QPU moonshot
The Field Lattice Engine
If gate-model quantum hardware plateaus, Kagaya's long-horizon bet is not "more qubits forever." The hypothetical Field Lattice Engine is a post-qubit computing substrate: a programmable physical field whose native operation is to emulate local Hamiltonian evolution directly. Instead of encoding reality into millions of fragile digital gates, the machine would tune controllable analog fields, topological constraints, error-detectable boundaries, and measurement surfaces so the hardware behaves like the class of physical system being studied.
This is not a product claim. It is a research direction inspired by analog simulation, quantum field theory, and the old lesson that the best simulator of a physical process may be a carefully engineered physical process. If quantum computing is computation with qubits, the Field Lattice Engine is computation with programmable law-like dynamics. It would still obey physics, still need validation, and still fail if the measurements do not match reality.
Milestone before myth
No fundraising narrative outruns a benchmark.
Hardware optionality
Every result needs a classical baseline and a non-QPU fallback path.
Loophole audits
Survival policies are attacked before they are anthropomorphized.
Negative results
Failures are product requirements because reality is the customer.
Founder note
Kagaya begins as a question.
If matter became human once through time, pressure, memory, survival, and care, can intelligence be grown again through a different path? Kagaya AI is my attempt to follow that question seriously: not to build a chatbot with better manners, but to study whether physical state models, embodied constraints, and evolutionary pressure can produce something that understands why anything matters.
The goal is not to make machines imitate our emotions. The goal is to understand whether concern, agency, and inner life require a world that can push back.
Research roadmap
The path to the 2027 horizon.
The countdown is a target for disciplined urgency, not a guarantee. Each stage must produce evidence, failure reports, and public questions before the next stage earns trust.
State models
Build physics-grounded forecasting benchmarks for small, measurable systems.
Survival sandboxes
Run constrained artificial-life worlds with energy, damage, memory, and repair.
Embodied agents
Test agents that must maintain themselves across changing environments.
Conscious-aware target
Evaluate whether any system shows stable self-models, concern, and generalization.
Open questions
The hard problems are part of the work.
What would count as evidence of inner life?
Language is not enough. Kagaya needs behavioral, mechanistic, developmental, and counterfactual tests.
Can survival pressure create agency without suffering?
An ethical artificial-life program must ask what forms of constraint are informative without being cruel.
How do we stop simulation loopholes?
Agents must be tested across shifting worlds, conserved quantities, and adversarial rule changes.
What should never be built?
Some capabilities may be scientifically interesting and still morally wrong to instantiate.
FAQ
Straight answers for reasonable skepticism.
Are you claiming this exists today?
No. Kagaya is an exploratory research vision. The site describes a direction, a thesis, and a target horizon.
Is this AGI or ASI?
Not exactly. The goal is conscious-aware AI: an agent whose intelligence is tied to self-maintenance, stakes, emotion-like regulation, and a coherent model of itself.
Why quantum simulation?
Because matter is quantum mechanical at the deepest levels. Quantum hardware may eventually help represent state spaces that classical systems struggle to sample directly.
Can AI really feel?
Unknown. Kagaya treats feeling as a scientific question about embodiment, homeostasis, memory, self-modeling, and survival pressure, not as a marketing claim.
Are you hiring or raising?
No. As of May 2026, Kagaya AI is not hiring and is not raising capital.
Research log
Follow the journey as it becomes public.
Public notes are coming soon: hypotheses, benchmarks, failed ideas, reading lists, and progress toward the 2027 target horizon.
Operating stack
Built like a lab. Shipped like a startup.
01
Capture
Ingest particle traces, molecular scans, and synthetic field runs.
02
Entangle
Encode transition candidates across quantum registers and constraints.
03
Resolve
Collapse state forecasts into confidence maps engineers can use.
04
Invent
Search material futures that classical compute would politely decline.
Future collaborators
Who this is for.
Kagaya is for people drawn to physics, artificial life, consciousness, safety, design, and the uncomfortable questions between them. Formal roles are not open yet, but the direction is meant to be followed, challenged, and sharpened in public.
Enter the labUpdate: May 2026. Kagaya AI is not currently hiring. This research vision is shared as an exploratory journey into frontier AI, quantum simulation, and artificial life.
Contact