Lab

A WebGL system that reads an episode of a podcast and turns it into video. Not an illustration of the conversation, a reading of it.

The podcast is ECO, which I produce with an AI collaborator: two agents argue about copyright, authorship and creative work, each reasoning from a different place. Two episodes got made. The visual system is the part worth showing.

The conversation as a space

Where the trail sits on screen carries meaning. Down is certainty, up is an open question. Left is the individual case, right is the systemic one. The corners are where a position is most exposed: a personal doubt in one, an unresolved system in the other.

Colour goes by state and not by speaker. Warm for friction, cool for reflection, muted green when someone gives ground.

The trail never clears. Every turn leaves its mark, so the conversation piles up on screen as it goes and the last frame holds the whole episode.

Two languages that do not repeat each other

One rule I kept from the start: the images are never literal. If the word on screen says mirror, the image is not a mirror. Two languages that reinforce each other rather than repeat, and the reading stays open.

What is underneath

The visuals are a WebGL shader with ping-pong render targets, which is what lets the trail accumulate frame by frame instead of redrawing. The audio is a separate pipeline and that one is automated end to end: a model writes each turn from a briefing written per agent, a voice model reads it, the pauses are chosen by rule (longer when the speaker changes, longest before an interruption), and the episode is assembled by concatenation with a second pass levelling every turn.

What is next

The two halves do not share a source of time. The audio pipeline never writes down when each turn starts, because concatenating files never needs to know, so the timeline driving the visuals had to come from outside it. Making the same process that assembles the episode also emit that timeline is the next move, and the one that would turn this from something I drove by hand into a system.