The short list of computational primitives that nearly all frontier-lab work is assembled from: what each is, the parameters it accepts, when to reach for it, where it fails, how heavily it is used, and what it pairs with. Every sheet closes with a reading of the primitive through a plant, so the open slot at the end, the one the Boquila exercise was trying to fill, has a shape.
Grouped by the question each family answers. Most real systems take one primitive from each of the first four rows and bolt them together; the substrate underneath almost everything is gradient descent.
Estimated share of frontier-lab production systems (the large language, multimodal and reasoning models shipped in 2025–26) whose pipeline depends on each primitive. These are my estimates from public technical reports, not measured figures; treat the ordering as reliable and the digits as rough.
The mechanism, not the inventory: a corpus goes in on the left, an answer comes out on the right, and each primitive is placed at the stage where it actually does work. Solid arrows are the main path; dashed arrows are the tuning and search loops that run beside it.
All seventeen, grouped by family around the center. Faint lines are the pairings from the sheets below; hover any primitive to light up what it combines with, click it to jump to its sheet.
Read a row as "this family, used with that one." Native means one is routinely built on the other; common means a well-trodden combination; rare means it happens, but you would have to argue for it.
Every sheet above shares a set of quiet assumptions. A leaf that grows itself into the shape of its neighbor breaks each of them. Listing the inversions is the fastest way to say what a plant-derived primitive would have to be, in language a lab would accept.
| The canon assumes | The plant does instead | Nearest existing sheet |
|---|---|---|
| A model separate from the body it controls | The body is the model; the trained structure is the organism | 07 Cellular automata |
| A global objective and a central error signal | Local-only signals; every cell sees its neighbors and the light on its own surface | 07 Cellular automata · 01 (local rules only) |
| Reversible parameter updates | Irreversible growth; a leaf cannot be un-grown, only abandoned | 03 Bayesian optimization |
| Cheap samples, many of them | Costly committed samples, few of them, paid for in carbon and water | 03 Bayesian optimization · 02 Evolutionary |
| Sensing, then acting | Sensing by growing; the action is also the sensor | 13 Active inference |
| The ability to simulate forward before committing | No forward model; hedge structurally by keeping modules redundant | none (MCTS has no plant analogue) |
| Fast discrete steps | Slow continuous developmental time; the clock is the season | 05 Recurrence · 13 State estimation |
Shape matching under local-only sensing with irreversible, resource-limited growth.
Neural cellular automata for the local-rule body, evolutionary search to find the rule, Bayesian optimization for the budgeted sampling, active inference for sensing-as-acting.
Not a metaphor but a benchmark: a target silhouette, a light field, a cell population, a carbon budget, and a score. Ship the environment; the biology is the ground truth that keeps it honest.