Borrowing DNA’s architecture, not its chemistry

Biological DNA solves a problem every complex system eventually faces: how do you store an enormous amount of structure in a compact, reliable, self-correcting form — and let it build something far more complicated than itself? The AI Genome Project — code-named Kritrima Tvasta(कृत्रिम त्वष्टा, “the artificial fashioner,” after Tvashtr, the Vedic craftsman who shapes all forms) — is Venta Labs’ research layer, asking whether that same architecture has anything to teach us about how AI models are structured, versioned, and passed from one generation to the next.

To be precise about the problem, because it is easy to overstate: models today do inherit. Distillation moves behavior from a large teacher into a smaller student. Continued pretraining resumes from an existing checkpoint for roughly half the compute of a fresh run. Each generation trains on synthetic data written, filtered, and ranked by the last one. Inheritance is not missing.

What is missing is addressability. None of those mechanisms let you name a single capability, lift it out, and put it somewhere else. You cannot meaningfully diff two checkpoints, attribute a behavior to a region of the weights, or ask for “this model, but with that model’s skill at X.” Capability transfer is real, but it is wholesale and lossy — never selective.

So this page does two things. First, it explains real human DNA clearly — the chemistry, the pairing rules, the reason it twists — grounded in established biology. Second, it asks what an addressable genome could mean for a model. The second half is deliberately framed as a vision, not a finished product — we say exactly where the science ends and the idea begins.

The four ideas worth actually understanding

Before the metaphor, the biology. These four ideas are the ones the rest of this page leans on.

From a biological genome to an “AI genome”

Here’s the conceptual leap, stated plainly: could an AI model’s parameters, structure, and behavior be represented in a form with the same properties DNA has — compact, self-describing, versionable, and auditable? Below is how the vocabulary might map. This is a working hypothesis, not an existing system.

What already exists — and where it stops

This intuition is not new, and the honest version of this project starts by saying so. The closest existing work is task arithmetic: a capability acquired by fine-tuning can be captured as a task vector, the difference between the tuned and base weights (Δθ = θ_tuned − θ_base). Those vectors can be added, negated, and scaled to compose or remove behaviors with no retraining at all. Model merging, LoRA adapters, and model soups all build on the same premise — that capabilities have some modular structure in weight space.

That it works is encouraging: it suggests the modularity is really there. But it comes with a hard constraint — task vectors are only composable between models that share an identical architecture and pretraining run. The moment width, depth, tokenizer, or data mix changes, the arithmetic stops being meaningful. Which is precisely the case we care about: carrying a capability across generations, not within one.

What we’re actually exploring

The long-term question behind this project: could a model’s capabilities be described in a compact, inspectable “genome” — something you could diff, version, partially recombine, or audit, the way you can with a DNA sequence — and could that description survive being moved into a model of a different shape?

We don’t have that solved, and we’d rather say so plainly than overclaim. What we have is a vocabulary borrowed from the biology above, an honest read of where the existing methods stop, and three questions we think are worth pursuing.