Genesis 3D

The intelligence is also in the way context is given to it. I do not send it the whole scene. I select the useful object in the tree, and Genesis 3D serializes only its relevant information: geometry, transformations, material, light, camera, or hierarchy. Context is selected by the interface rather than compensated for by an ever-larger window.

The model can then build an interface or a behavior directly inside the document. The code it produces becomes part of the scene and leaves with it on export.

Fish with autonomous needs

The version shown here is a lighter web variation of my behavioral labs. The fish do not follow a recorded path, and no language model decides how they move.

The system is based on internal states that keep changing. Hunger, rest, safety, and social need compete with each other; the environment adds different sources of stress. Perception itself depends on that state: a target is not chosen simply because it is close, and a hungry fish may accept a situation it would have refused a few moments earlier.

Learning was then tested separately on some behaviors, using several evolutionary approaches. The results, including the failures, led to an important separation: letting the decision learn when it benefits from evolving, without necessarily handing the mechanics of the gesture over to learning.

Try it, move the mouse in the aquarium: some fish will come to inspect it. Right-click to feed them. Shift + click to scare them. Their reaction depends on their state at the moment you intervene.

AI drone race

The drone follows real flight dynamics. To move forward, it has to tilt: its thrust follows the vertical axis of its own body. Drag depends on its airspeed, yaw has inertia, and the track itself is generated for each race from curves, then turned into a three-dimensional course.

The race includes four opponents. A driving model is being studied separately using data from my own races. I recorded 64 races and 41,062 telemetry samples: speed, position, orientation, tilt, offset from the racing line, and upcoming curvature. A degree-2 polynomial regression learns three relationships between curvature and the way I fly. The resulting model finally fits into nine coefficients.

The learned model has nine coefficients, but it is not yet connected to the opponents' driving. Their current behavior relies on the prototype's racing logic: racing line, corner anticipation and lateral control.

Try it, click the video to open the race. Three laps, with the keyboard or a gamepad; with a gamepad, the handling really starts to make sense.

Adaptive WebGPU rendering

The physical renderer computes the path of light in WebGPU, directly in the browser. The constraint is strict: one sample per pixel, where an offline render can use thousands. A raw image would be almost unusable.

The answer is not to add artificial power, but to use previous frames as information. The engine reprojects the history, accumulates statistical moments of luminance, estimates variance, then adapts its filtering to the geometry and local uncertainty.

The loop then closes: the ray tracer reads the variance produced by the denoiser in the previous frame. An area that is still unstable receives more samples; an area that has already converged enough receives fewer. The system reallocates its own compute budget according to what it still estimates poorly.

This is not AI in the sense of a learned model. It is a form of architectural intelligence based on estimation, memory, and feedback.

Try it, rotate the view with the mouse. Shadows, reflections, and indirect light are recalculated while the view moves.

Cameras with collective attention

Each camera has its own rhythm: an individual behavior multiplier, its yaw and pitch oscillations, and small saccades that appear and fade out gradually. The two axes use deliberately desynchronized frequencies so the movement never falls back into the exact same loop.

The intelligence comes from the attention model. A camera does not switch abruptly from “I am following” to “I am not following.” Its orientation is a continuous blend between tracking a target and its own spontaneous motion. It can be highly attentive, distracted, or almost indifferent.

Each camera is also influenced by its eight neighbors. A local variation in attention can then propagate through the grid. Other excitations use a delay proportional to distance: instead of moving as one block, the wall transmits its state from neighbor to neighbor.

Try it, move the mouse inside the image. The butterflies follow your movement, and the cameras react in turn, each with its own rhythm and its own level of attention.

Autonomous butterfly landing

The butterfly does not replay a landing animation. Its flight uses a Lévy walk smoothed with Catmull-Rom splines: many small moves, punctuated by larger ones. This type of distribution is used to model some search strategies observed in animals.

When it decides to return to the hand, a state machine manages the transition from flight to return, then to landing, but the trajectory is still recalculated: the finger can keep moving during the approach. The six legs are then solved with inverse kinematics and hinge constraints. Their targets are projected by raycast onto the actual finger mesh, so the feet really meet its surface.

The wings also take part in the decision. Their downward beat is reduced when the geometry indicates a risk of collision with the finger. And when several butterflies want to land, a multi-agent arbiter selects the best placed one while the others move away.

The visible intelligence therefore comes from several specialized mechanisms solving together a situation that no complete animation describes in advance.

Try it, move the finger with the mouse. The butterfly adapts its approach to the target. If it loses a stable situation, it can leave again and look for another chance to land.