Questions

A small research lab working on one question: can the design of a learning system be calculated from physics, specifically from entropy, energy, and wave dynamics, instead of being found by trial and then scaled up?

The work is theory first. Derivations get tested in small numerical experiments, and anything that survives gets built and run on a real task. The research program states the question precisely, lists the published results it builds on, and sets out the tests the idea has to pass.

Enkidu, in the Epic of Gilgamesh, is made out of clay. The motto on the front page, ex luto ad astra, means "from mud to the stars". The point of both is straightforward: we treat intelligence as something that comes out of ordinary physical matter and can be explained at that level, with no extra ingredient required.

The glyphs 𒂗𒅗𒄿𒁺 spell the name in Sumerian. Cuneiform was pressed into wet clay, and it is the oldest writing we can still read. We liked it as a mark, and the clay connects it to the name.

Today it is a small lab. What we want it to become is an organization that does both the science and the engineering, where the things we work out get built, and where the systems, software, and knowledge accumulate in one place over a long time rather than being scattered across short projects.

It is not set up as a product startup, and nothing on this site is written for a funding round. The research program describes where the work actually stands, including what has not been done.

Two precedents. Thermodynamics turned engine design from a craft into a calculation, and electromagnetic theory produced radio.

Scaling what already works is being done well by groups with far more money than us, so there is little point in us competing there. The reason to work on foundations instead is a pattern that has held before: a field gets its engineering power once it gets its theory. Information theory took communication from rules of thumb to provable limits.

Machine learning today has powerful systems and recipes that work, and not much that explains why. Existing theory mostly analyses what has already been built rather than saying what to build next. Some architectures do have principled parts, convolution being the clearest case. What is missing is a derivation that runs all the way from a physical starting point to the structure that actually runs. So there are two outcomes here and both are worth having. If designs can be calculated, the systems that follow will be different in kind from what we have now. If someone proves they cannot be, that proof is itself a foundational result about the limits of the field.

Legg and Hutter, 2007. The measure is "in no way anthropocentric" and its value "is not computable".

We work as though there is one. Legg and Hutter wrote a mathematical definition of intelligence for arbitrary machines in 2007 and were explicit that it owes nothing to human beings in particular. The catch, which they state plainly, is that its value cannot be computed. It depends on Kolmogorov complexity, which is roughly the length of the shortest program that would produce a given thing, and which cannot be computed either.

So there is a definition and nobody can work out its value, which is more useful than it sounds. It means the question worth asking is not "what is intelligence", which we cannot settle, but which physical systems get close to it, how close they get, and what it costs them in energy and time. Those are questions a laboratory can actually work on.

To be clear about the size of the claim: we are not saying the Legg and Hutter version is the right one. We are saying a definition that owes nothing to human beings is possible at all, which is the part that changes how we work.

No, and the distinction matters to how we work. The brain is the one system we all agree is intelligent, which makes it tempting to treat as the specification, and we treat it as data instead.

A brain is one implementation that evolution arrived at under constraints no engineer would choose: build from cells, run wet and warm, never switch off, repair while running. Some of what it does follows from physics and would constrain any system doing the same job. The rest is biology working around its own limitations. Copying the whole thing means copying both, and nobody yet knows where the line falls. Finding that line is the fifth area of the research program.

Two of its properties are hard to ignore. It runs on about a fifth of the body's energy, so intelligence at that budget is demonstrated rather than hoped for. And it does not learn the way we train our systems: backpropagation needs a separate backward pass and access to the forward weights, and neither fits the brain. How it does assign credit for a mistake is an open question in neuroscience.

No. The waves here are classical ones, the kind water, sound, and light make. Superposition and interference are ordinary properties of those, and the mathematics around them is statistical mechanics and information theory rather than quantum mechanics. No quantum hardware is assumed, required, or claimed anywhere in this work. If a result ever does depend on quantum effects, it will say so plainly.

The phrase "artificial intelligence" now describes a market and a funding climate more than a research question, and most of what it covers is not what we work on. So we use narrower words, because our claims are narrower: learning systems, inference procedures, computational substrates. Each of those means something specific, which is the point.

No products, no benchmark claims, and no breakthroughs.

Three things exist. There is a written research program that says what would prove it wrong, a set of dated lab notes published as they are written, and formal groundwork with small numerical experiments in progress.

Technical results will be published when they survive scrutiny, along with whatever you need to rerun them yourself. Until that happens, the notes are the honest record of where things stand.

Write to research@enkaidu.com and attach something you have finished: a derivation, a proof, a simulation, or a system that runs. A background in mathematical physics, information theory, or systems engineering helps. We will read the finished work first, and it counts for more than where you studied or worked.