Horizons
Superintelligence will first be an institution that keeps improving as its models are replaced
We are betting on an institution of research models, people and instruments that trains and tests each successor model on its own experiments before that model joins the work. In agent work today, a person often finds a conclusion still resting on a premise that changed several sessions back, and corrects it. The five essays look ahead to such institutions. Their undertakings outlast the research models working on them, and an experiment can still be running after the premise it was designed around has been overturned.
Designing living materials
An institution that invents its next intelligence
Weeks in a test frame have thickened a beam of living material at the places where it bends. Cells inside it are still adding material there. Other cells, in the stretches that carry little load, are removing it. Elsewhere in the institute that designs these materials, robotic laboratories bend and grow more specimens on different schedules, and research models compare how each one responds. Its next research model learns from these experiments. When an independent team hands the new model a material whose failures come from small differences in structure, the model asks how the cells signal, and then asks again.
Regrowing spinal pathways
Repairing a nervous system that has learned to live with injury
The rim of the cup is against the table, and a therapist has a hand on it. The participant got it there with a lift of the shoulder and a turn of the wrist, once the fingers had closed. Some of the water no longer reaches the participant's mouth. A screen beside them shows more activity in the stretch of spinal cord that the institute repaired. Since the injury, this movement has depended on a support and on an elbow held at one angle. Six researchers founded the institute, and the AI scientists who far outnumber them have been choosing exercises that keep that old movement intact. The participant chooses a different grip. Learning it means supervised practice, and a stretch of time in which the familiar movement can no longer be counted on.
Power from superhot rock
What three people can afford to begin
Water goes down one well into rock hotter than older drilling equipment could survive, then rises through another, carrying the heat. Three founders mean to sell power made from that heat to an industrial district nearby, which will pay for electricity it receives and nothing for heat left in the ground. They can afford one field. Their AI organisation spends the founders' money finding out how the rock behaves, and the money goes further when the founders are not paying again and again to reconstruct work already done. At the second site the rock resembles the first field's. Its fractures connect in a different arrangement. The model that has worked longest on the first field proposes running the new site on the first field's operating pattern. The AI organisation's tests there include one in which water circulates through the rock in a pattern changed on purpose.
Restoring a river basin
Five visits to a changing river
Where the river used to dry into separate pools, a side channel now carries cold water under the footbridge, and by midday the bridge is in the shade of trees that were once small enough to step over. Among the visitors are people who approved the programme, some with grandchildren. The farms and towns along the river own the institution behind the work. Ninety-one years after the channel opened it will still be operating, long after the working lives of the people who began it have ended. The research models that planned this channel have already been replaced by successors trained on the basin's own experiments. Several summers of measurements show the young woodland adding to the river's flow in the dry season. The restoration plans being drafted now ask for more forest and count on more water from it.
Building shipyards on asteroids
The ancestry of a shipyard
The daughter yard's mining robots grip the asteroid's surface while they cut, since their weight alone will not hold them down. The yard came from its parent on another asteroid, folded into a transport frame with its furnaces and a small power supply. None of the founding models still directs the whole undertaking. Now the daughter yard casts specimens from its own ore and gives them the heat treatment the parent yard developed for metal from the first asteroid. The specimens that looked stable during conditioning crack when they are loaded at operating temperature. The daughter yard's models prescribe the same heat treatment for the next batch, and each batch they cast is metal the furnaces have already spent energy refining.
Testing a replacement research model on materials kept from its training
Each of these institutions is misled first by something it did well. The materials institute's retrained model asks about cell signalling when an independent team hands it a material whose failures come from small differences in structure. The nerve-repair programme, after its early successes, keeps protecting the participant's old movement. At the second geothermal site, the model with the longest record on the first field proposes that field's operating pattern. The river institution's plans count on more forest bringing more summer water. The daughter shipyard's models keep prescribing the heat treatment that worked for their parent. Because the correction of each mistake passes to the next generation along with the success that led to it, improvement in institutions like these can accumulate across many changes of model.
At the materials institute, the older model works more slowly and considers more possibilities, and it is the one that finds the fault in the independent team's material, while the newer model is still asking about cell signalling. The institute counts what developing the newer model has cost. From then on, a new version must do the work its predecessor was responsible for, tested on materials held back from its training and on the problems the institute depends on, and it must also improve on the latest breakthrough. The institute runs those checks on every later version and pays for them each time. Then it retrains the newer model. Working on two specimens that received the same signal and are growing at different speeds, the retrained model asks whether the cells are getting even supplies of nutrients, and it orders a flow measurement to run with the mechanical scan. The measurement finds a constriction in one of the specimens.
Postponing the first vessel for a second generation of machines
At the parent yard, a cutting head runs along a polished guide, and the first vessel needs the parts it cuts. Several cycles of heating and cooling after passing inspection, a new guide has changed shape. The parts cut along it carry its error. The vessel's departure had been planned on the assumption that a guide which passes inspection will hold its shape through operating cycles. The human directors decide the vessel can wait. Its metal goes into a second generation of experimental machines, along with the workshop time it was booked for. In agent work today, the same step, from a failed assumption to the plans built on it, often falls to a person. For AI teams, we are building durability infrastructure that reopens the decisions resting on a changed premise.