A question.
A run. A revision.

The real innovation is an enduring evolutionary experiment. Frontier models design and assess the experiments, adopting, revising and replacing generations of node firmware to find out how they perform. They can be insightful or mistaken, smart or surprisingly dumb. Their work is judged by the comparisons they leave behind.

A long-running coordinator such as Meta Muse Wonder can organise a team of node-specific Muse Spark agents. With a connected development and deployment host, the intended loop can keep running around the clock: contemplate an improvement, write it, compile it, flash a candidate, observe the result and decide what to retain or change.

The human is primarily the observer. We watch and listen for amusement and for the intuition that something unusual is happening: a rhythm, a hesitation, a collective response that does not fit our understanding of normal. That observation can become a question for the model team. Manual controls let us inspect, reproduce or intervene without making us the author of every experiment.

Give evolution something to contend with.

Kulira introduces competing or hostile behaviour—noise, deception, interference and unreliable peers. We can also change the world around the flock: temperature, movement, sound or proximity sensed through suitable lidar, infrared or mmWave hardware. Each added sensor needs its own supported hardware and firmware path; these are possible experimental inputs, not a claim that every node already measures them.

Keep the objective, baseline, firmware version and observations with each generation. A model’s confident assessment can be wrong. Preserve failed candidates and surprising outcomes so a rolling experiment remains something we can investigate, rather than a sequence of attractive animations.

Inspect the controller and manual controls → · Meet Kulira →