What is happening here?
Most of modern machine learning was found by experiment, not derived from theory. ENKAIDU is building the missing layer: a framework grounded in physics and mathematics where computation, learning, and inference are derived, not engineered.
A frontier intelligence lab
A research institution that derives computational architectures from mathematical and physical theory, then builds them. Derivation and engineering are one continuous program.
The physics beneath learning
Thermodynamics constrains every computer. Energy methods already sit, mostly unrecognized, inside the strongest model families. Wave dynamics can carry computation. We work on the mathematics that joins these facts into one framework.
A new computational foundation
The objective is a theoretical and engineering base for a different computational paradigm, and the systems and tools that follow from it. The research program states the tests this has to pass.
How the laboratory operates
The research program defines the questions. The doctrine defines the methodological standards: what assumptions the lab will not make, what evidence it requires, and how results are expected to pass into working systems.
-
INo architecture is sacredModel families are hypotheses, not axioms. Any architecture must be revisable when theory or experiment no longer supports it.
-
IIExplanation before optimizationEmpirical performance alone is not an explanation. Results should be grounded in formal structure, measurable constraints, or identifiable mechanisms.
-
IIITheory must pass into machineryA theoretical result that cannot inform the design of architectures, inference procedures, or computational systems is regarded as incomplete.
-
IVSystematic developmentThe objective is not a sequence of publications or product cycles. It is a sustained program that develops theory, models, software, and infrastructure over extended timescales.
Scientific motivation
Transformers, diffusion models, and their variants were discovered empirically and then scaled. They work, but no derivation says they are the right shape: statistical learning theory can analyze parts of their behavior after the fact, and it did not produce them. ENKAIDU starts from the position that computation and learning are physical processes governed by entropy, energy, and wave dynamics, and that correct architectures can be derived from these laws instead of guessed at and tested. If this derivation succeeds, the resulting systems will not be better versions of current ones. They will be different in kind.
Start from physical law, not architectural convention. Thermodynamics, statistical mechanics, and information theory impose hard constraints on any learning system. These constraints define the structure from which architectures are derived.
Derive, do not guess. Energy landscapes, wave equations, entropy functionals, tensor structures. These are not metaphors. They are the substrates from which model classes and inference procedures are constructed.
Build what the theory produces. A derivation is finished when it ends in something that runs: an architecture, an inference procedure, a prediction that a numerical experiment can check. If the physics is correct, the engineering follows.
A field that scales what it cannot derive has not yet found its theory.
Four pillars
The program is organized into four pillars: the substrate, its mathematics, its physical constraints, and the higher-order structure above them. Each is listed with one of its current open questions. The full program, including methods and the sequence of tests it must pass, has its own page.
Wave and energy-based computation
Computation carried by wave propagation, energy minimization, and entropy production, treated as a paradigm of its own. Open question: which inference operations interference and relaxation can implement directly, and at what cost.
Mathematics of intelligence
Complexity theory, information geometry, and topology applied to the structure and limits of learning and inference. Open question: what geometric structure a model family needs for learning inside it to be tractable.
Physics of computation
Thermodynamic costs, reversibility bounds, and the statistical mechanics of learning and inference in physical systems. Open question: how far practical inference sits above the Landauer limit, and what closing that gap would require.
Self-reference and higher-order structure
Formal self-reference, reflective computation, and higher-order structures as requirements for systems that reason about their own operation. Open question: which fixed-point structures permit a stable self-model.
Recent notes
Dated working notes from the program: positions, decisions, readings, and dead ends, published as they are written. Entries are added when there is something to record, not on a schedule.
Who we look for
Researchers in mathematical physics, information theory, and computational science, and engineers who can build the systems this program requires.
The work spans formal theory, numerical experiment, and systems engineering, and much of it sits between those labels. Useful candidates either operate across more than one of these areas or go deep enough in one to move it. When you write, include something you have finished: a derivation, a proof, a simulation, a system that runs. Finished work counts for more than pedigree.