The vision
Intelligence,
the way nature
affords it.
A leaf does not train its venation. A crystal does not descend a gradient. Nature builds working structure from local rules, in one growing season, on ambient energy — and it builds it everywhere, for everyone, without asking a datacenter for permission.
Siddha exists to find out how much of machine intelligence can be made that way: grown from data in closed form — frugal by construction, transparent by construction, sovereign by construction — and to say, in public and with numbers, exactly how much cannot.
The fork in the road
One path got all the traffic.
The field chose, almost unanimously, the path of scale: bigger architectures, more gradients, more megawatts. It has produced marvels — we use them, we measure against them, and our own benchmarks print their victories at full size.
But a second path was left almost unwalked: let the data construct the model directly. No loss function, no backpropagation, no epoch zero. Our first architecture on that path — the Reverse Synthetic Network — already matches trained baselines on the work most organizations actually need: classifying, routing, triaging. The vision is to walk the path properly, and map it honestly.
Intelligence became industrial
The dominant recipe — fix an architecture, fill it with random numbers, push gradients through it millions of times — works. But its appetite is industrial: GPU-months, megawatt-hours, cooling towers, capital fences. The recipe itself decides who gets to make minds.
The meter runs on retraining
Data drifts weekly; budgets retrain monthly. Most organizations live with stale models not because intelligence is hard but because the loop that produces it is priced like a foundry run.
Opacity became normal
We accepted that a model is a sealed object — millions of weights nobody can point at. Audit, provenance, and explanation were retrofitted as paperwork instead of being properties of the construction.
None of this is an argument against trained frontier models — they do things synthesis provably cannot. It is an argument that the other half of the economy of intelligence deserves a builder.
What we hold to be true
The model is already in the data. Most of what we call “training” is paying, in megawatts, to rediscover statistics the data would have told us for free — and the part that isn't, the genuinely emergent part, deserves to be named precisely instead of marketed vaguely.
That second clause is the discipline. Our research found the exact boundary — similarity synthesizes; composition must be learned — and the vision is built on respecting it, not wishing it away.
The long arc
From seed to forest.
Four stages, each funding the next. The first is measured and shipped; the last is a research bet we state plainly as one.
Closed-form synthesis is real and competitive for discriminative work: 87.0% on 20 Newsgroups against a tuned SVM's 86.1%, zero gradient steps, about a minute of CPU. The Lab on this site grows working models in your browser tab. The benchmark protocol — bits-per-byte against a measured GPT-2 — keeps us and the field honest.
Siddha Forge: labeled data in, deployed classifier out, in seconds — priced per synthesis, not per GPU-hour. Alongside it, the research the reviewers asked for: noisy-data robustness where the interconnection layer should finally pay off, larger-scale vision (MNIST and beyond), and synthesis as a fixed floor and fallback expert beside trained models.
Models grown where the data lives, as a norm rather than a workaround: clinics, courts, factories, panchayats — synthesis engines embedded on their own machines, auditable because every weight is a deterministic function of their own records. Procurement that accounts energy per model. Sovereignty by construction, not by policy promise.
The hard wall our own research published: closed-form statistics encode similarity, not predictive composition — generation plateaus at the n-gram ceiling, ~1.9× the bits of a trained GPT-2, and more data does not fix it. The horizon question is whether any gradient-free construction can cross that line: learned-but-frozen representations, hashed count structures, compositional priors we haven't found. We may fail. We will publish either way — that is the bet's integrity.
If we are right
What the world looks like when this works.
Not a chatbot demo — a change in who can make working models, and where they live.
The triage router is synthesized on the ward laptop from the hospital's own records. Nothing leaves the building. When admission patterns shift in monsoon season, the model is regrown before the morning round.
The defect classifier rebuilds itself every shift as the product mix changes — on the edge box bolted to the line, for the energy cost of a light bulb minute.
Every model in the paper reproduces on a student's laptop in the time it takes to read the abstract — because the models are deterministic functions of the data, not artifacts of a cluster nobody else has.
Document routing runs on a model the office grew itself, from its own files, on its own machine — inspectable feature by feature when a citizen asks why.
Every vignette is an extrapolation of a measured property — synthesis speed, determinism, inspectability, CPU-only builds — not of capabilities we haven't demonstrated.
The vow that makes it credible
We publish the walls we hit.
Every grand AI vision asks for faith. Ours asks for the opposite: check us. The generative ceiling, the digit benchmark we lose, the warm-start that didn't work — they are on the research page at the same size as the wins, with the scripts to reproduce them.
A vision disciplined by published negative results is the only kind that compounds: every wall we map is a wall no one in the field has to rediscover, and every claim that survives is one you can build on.