I only experience things in latent space now.

Concept

"I only experience things in latent space now" is a personal mediation on how the algorithm and I mutually reconstruct and deconstruct each other. It uses the parable of Indra's Net (from Hindu and Buddhist folklore) to build a structure that simulates the algorithm, an infinite net in which we all are reflected through one another. By voluntarily participating in this algorithm now, I pick up one of the jewels and fracture the net permanently. My childhood memories bleed through old video diffusion models that try to make sense of them while simultaneously abstracting and generalizing them.

Clips of South Asian "brain rot" and internet ephemera are passed through machine learning video inference models, like depth, attention, and uncertainty, to reveal how these recommendation algorithms see the world. Everything I experience now is mediated through this algorithm, as is everyone I perceive, and at the same time, through my interactions with it, I constantly shape this model too. These systems first adapt to just your age, location, and gender, then slowly become a fractured and infinite reflection of you and the world around you. Eventually there is nothing left outside of Indra's net.

Process

The footage, clips from LAND and South Asian internet ephemera, goes through four models, each reading it for a single property: SEA-RAFT for optical flow, Video-Depth-Anything for depth, CoTracker3 for point tracking, DINOv3 for self-supervised features. Each one throws out everything it wasn't built to see, so depth keeps distance and drops color, while flow keeps motion and drops whatever holds still.

All of it runs on one 8GB laptop GPU, which decided most of the architecture. The four models need incompatible CUDA and Python builds, so each runs as its own subprocess instead of an import, and flow and tracking can't be resident at the same time, so the flow pass finds where motion actually is, seeds the tracker from those points, then unloads itself before the tracker loads. One flow inference feeds seven visualizations. Nine visualizations per clip, about twenty minutes of GPU time each, and the results come back as video, cut against point clouds and environments built in TouchDesigner and Unreal Engine.

Exhibitions

Screened at Sydney Opera House, 28 May 2026, as part of Shortwave × Soft Centre (Vivid Live). Reworks footage from LAND (2025). Score by Cash. Runtime 6:06.