Working Paper · 2026
Embodied Delay: Translation and Autonomous Media Practice
Manuscript under journal submission · Documents Heron, an autonomous media installation, 2025–2026
Abstract
Heron is an Electron-based media installation that frames translation as an embodied, contingent, and delay-structured practice across biological navigation, artificial intelligence inference, and autonomous media playback. Rather than optimizing for accuracy or responsiveness, the work stages translation as recalibration, pause, and temporal misalignment. A live narrative engine generates text on a fixed cadence, while a library of pre-generated video loops persists as an asynchronous visual past. The system distributes agency across offline video synthesis, remote storage, periodic local synchronization, and live interpretation, allowing infrastructural constraints such as latency, buffering, pricing, and outages to surface as material conditions of meaning. Drawing on embodied and enactive cognition, interpretation, articulation, and representation occur across distinct but interdependent layers, where latency functions as an adaptive response to uncertain signals. By making delay durable and perceptible, Heron challenges assumptions of immediacy in computational media and proposes temporal suspension and autonomy as productive strategies for revealing the hidden labor and limits of contemporary AI systems.
Keywords
translation · embodied cognition · enactive cognition · generative systems · autonomous media · latency · seamful design
About the Work
“If a lion could speak, we would not be able to understand him.” Wittgenstein’s remark is not about vocabulary — it is about shared forms of life. Contemporary AI systems speak with increasing fluency while inhabiting no body and no environment; their relation to the world is infrastructural, shaped by probabilistic inference, buffering, token limits, pricing, and uptime. Heron takes this condition as its subject, staging a continuous, unresolvable act of translation between a live, environmentally aware text engine and a fixed visual past that was generated days or weeks earlier under entirely different conditions.
The installation runs autonomously and refuses responsiveness: it does not adapt to viewers, learn from feedback, or recalibrate toward engagement. It persists — generating text, looping video, and logging its own delays and fallbacks as an accumulating archive of autonomous computational life. Drawing on research into avian magnetoreception, the work borrows a biological insight: when signals become ambiguous, organisms do not interpolate aggressively. They wait. Latency becomes a form of embodied intelligence.
Extended measurement of API latency during continuous operation revealed no inference failures or timeout events — delay in Heron is architectural rather than compensatory, shaping interpretation without threatening continuity. By foregrounding pause and infrastructural constraint as productive conditions rather than technical shortcomings, the work opens space for artists and technologists to engage generative media not as tools for optimization, but as autonomous processes shaped by maintenance, uncertainty, and temporal negotiation.
System at a Glance
- Visual layer — a fixed library of 29 one-minute video loops generated offline with Google Veo, organized in five thematic layers and sequenced by Fisher–Yates shuffle
- Narrative engine — live text generated by Claude on a 15-second tick from recent memory, local weather conditions, stochastic directives, and a curated corpus
- Periodic fallback — every fifth tick bypasses the language model and injects a direct corpus line, modeling an organism reverting to learned routes
- Vocalization — OpenAI text-to-speech articulates finalized text downstream of interpretation, introducing no semantic variation
- Infrastructure — remote video storage with hourly local synchronization on a continuously running Electron application; latency, outages, and absence surface rather than crash