An institution that invents its next intelligence

The beam grows thicker where it has been bent. Along its underside, living cells lay down material while other cells remove tissue from places carrying little load. After several weeks in the test frame, its shape fits the forces that pass through it. Then the apparatus reverses the load.

One specimen adapts. The other breaks. Before the reversal, their scans were almost indistinguishable.

The institute has a model of how these materials bear weight and another of how their cells respond to strain. Neither explains the difference. One research group thinks the surviving beam contains a structure the scanners cannot resolve. Another thinks the two populations of cells have acquired different responses to stress. Those explanations call for different machines.

The next apparatus can hold the load constant while changing the chemical signals reaching a small region of the beam. Building it brings the institute's cell engineers together with the team that measures deformation. They want to know whether they can change how the specimen repairs itself without first changing its shape.

An investigation with a working lifetime

Cells retain effects of past stress. The institute has to predict what happens when thousands of those responses interact. Growing tissue changes the forces on its neighbours and the movement of nutrients through the beam, altering what the next cells do. A repair that strengthens one region can start a process that weakens another weeks later.

Robotic laboratories follow specimens through different histories of loading and growth. The institute keeps competing explanations alive while their predictions diverge. A result from cell biology changes which mechanical measurements matter; a mechanical observation sends the cell engineers back to an assumption about signalling. The research crosses disciplines faster than any one specialist could master them, but often has to wait weeks for a physical answer.

While the specimens grow, researchers derive consequences of tentative explanations, design interventions that would separate them and examine whether an apparent contradiction is caused by the apparatus. Their working understanding includes why a promising approach was set aside and what new observation would make it worth pursuing again.

During one wait, the model coordinating the signalling experiments is replaced by a stronger one. Its first decision is to turn down a proposal to repeat the scanner study. The approach was set aside because its observations could not distinguish the competing explanations; a more capable reader of the same scans would face the same problem. The successor instead develops the chemical intervention with the cell engineers. When its results overturn an assumption about signalling, the institute withdraws the experiments that depended on it and keeps the rest of the programme running.

Designing how a material changes

The chemical interventions narrow the explanation. A specimen's response depends on which cells have already entered a growth programme, even when its visible shape resembles that of another specimen. But simply adding cellular history to the existing model makes it too cumbersome to use. The institute could collect more detail indefinitely without gaining a practical way to design a beam.

A collaboration between its biologists, mechanics researchers and model designers develops a different object of prediction: the repairs a region is capable of making, and how each repair changes the options available to its neighbours. This description omits most molecular detail. It retains the distinctions needed to predict when a local response will spread, stop or make later adaptation impossible. Experiments test where that simplification breaks.

The change gives the institute something it previously lacked: a way to design the response to an unfamiliar history of stress. It can ask which initial arrangement of cells and support material leaves a structure able to recover from several different injuries. The laboratories can now search for materials with useful future behaviour, rather than selecting only for strength at the moment of testing.

The people directing the institute face a choice. They could concentrate its resources on qualifying a particular material for manufacture. There is a plausible commercial path, with customers waiting for structures that need fewer replacements. Instead, they commit a substantial part of the institute to developing a general capacity to design adaptive matter. That decision postpones revenue and gives the scientific organisation a harder question to pursue.

The wrong kind of improvement

The institute begins changing the intelligence responsible for its experiments. It trains a new research model on the interventions that exposed the beam's behaviour, including the explanations they eliminated. It also teaches the model to propose experiments in the new language of possible repairs. A task that once required prolonged negotiation between several teams becomes something the model can work with directly.

At first, the successor appears markedly better. It finds useful chemical interventions quickly and spends less time on unproductive measurements. Then an independent team gives it a material whose failures are dominated by fine structural differences. The new model keeps asking about cellular signalling. Its predecessor's slower, broader investigation finds the fault first.

The training had preserved a successful habit too aggressively. The institute had taught a specialist to repeat its own breakthrough and mistaken that for better scientific judgement. It returns the broad work to the earlier researcher and keeps the new model on chemical interventions, where it is faster.

The model designers retrain it on episodes where the first plausible mechanism proved wrong, and hold back materials whose failures belong to other mechanisms. In the next comparison, the retrained model faces two specimens that grow at different rates after the same chemical signal. Before proposing another signalling experiment, it asks whether nutrients reach the cells equally and orders a flow measurement alongside the mechanical scan. One specimen contains a constriction. The next trial varies nutrient supply while leaving the signal unchanged.

It replaces the broader researcher only after matching it on the held-back materials and on the work the institute already depends on. The institute also accounts for the cost of developing the replacement. That test becomes the standard for every later version: a successor must carry its predecessor's responsibilities as well as improve on its latest breakthrough.

What the successor can ask

Years later, a descendant institute works on membranes that change their permeability as the water passing through them changes. Their cells must keep useful channels open, resist damage and regulate their own growth. A successful local response can redirect flow and create a problem somewhere else. The institution recognises a question its predecessors learned how to investigate in the beam.

It begins with the acquired ability to reason about interacting repairs. Researchers adapt the representation to fluid flow and build an apparatus that can vary pressure and nutrient supply independently. They know which kinds of observation could expose a missing mechanism. They also inherit the caution earned by the failed research model: a familiar pattern is grounds for a test, not permission to assume the explanation.

Most of the models involved in the original beam experiments are gone. Their experience remains active in the descendant's specialist models, instruments and ways of arranging an investigation. A fresh collection of frontier models could read the same papers and laboratory notebooks. It would still have to develop much of this judgement through work. The descendant can spend that effort on the new problem.

The growing institute also changes its organisation. Some collaborations that began as temporary exchanges become new scientific specialties. Other groups dissolve when their methods become routine enough to build into tools. The institution develops the capacity to choose among these arrangements, drawing on the consequences of earlier reorganisations.

Intelligence at the scale of an institution

There is no single model to which all this progress belongs. The ability to design adaptive matter resides in a continuing system of researchers, learned representations and encounters with the physical world. Its experiments improve its models. Those models help invent instruments that expose new phenomena. The resulting science changes which kinds of researcher the institution needs and how it can develop them.

This is a route to system-level superintelligence: an organisation that becomes capable of inquiries beyond the reach of any of its components. The claim requires more than long operation or a large population of agents. The institution must acquire abilities, carry them through changes in its membership and use them to improve the production of further abilities. Its descendants have to reach questions that their founding intelligence could not formulate well enough to investigate.

In a later laboratory, a test structure contains several kinds of living material. Some regions carry weight; others move water or regulate the conditions under which the structure grows. A machine applies a load the assembly has never carried. Along its underside, cells add material. Elsewhere, channels widen to keep the growing tissue supplied. The assembly changes shape, and keeps carrying the load.