The first full day at the 2026 Association of Internet Researchers conference in Mexico City starts for me with a session on sonic social media, where time is amplified by sound. We start with an introduction by Elena Pilipets and Marloes Geboers, who begin by reminding us about the ambient nature of sound, which was played with by the pioneers of ambient music and theorised by various scholars focussing on the ambient commons, the role of sound vibrations, and the distribution of sound across physical and digital space.
But the first paper is by Sal Hagen, whose focus is on TikTok. There have been a slew of new concepts to capture online sociality in the latest social media platforms, with a strong focus on the role of algorithmic curation in platforms like TikTok and its ‘For You’ feed, but this is too simplistic: such ‘vibey’ feeds might be algorithmically generated, but there’s a great deal more going on. There is also a substantial amount of work exploring the role of artificial intelligence here, but this is often focussing on the technological underpinnings rather than on the connection of AI vectors and experiential vibes.
If vibes and vectors are related, then it should be possible to vibe with the clusters of TikTok’s latent space of videos – but this requires a comprehensive slice of that latent space for its analysis. This can be done for TikTok, actually, as it has a number of standard ID patterns for its video IDs, and this enables the systematic gathering of a large number of videos present in the latent space at the same time.
Once gathered, these videos can then be analysed through an embedding approach (and dimensionality reduction using UMAP), producing a first overall clustering of video types and themes. It’s also possible to explore the incidence and persistence of these videos over time, their similarity to a given reference video, and the evolution of TikTok video styles over time. This reverse-engineering does not actually study TikTok’s systems, but it is likely that TikTok is increasingly drawing on similar embeddings as well in its algorithmic curation processes.
