Stable JEPA could auto-chunk time and only invoke the predictor on surprises
MarcJSchmidt · x · 2026-11-23
MarcJSchmidt offers a technical observation on stable JEPA architectures: instead of feeding physical state in at a fixed rate, you can train the encoder to auto-chunk observations and emit a latent only upon certain events, effectively compressing multiple boring state observations along the time axis into one latent.
Implications:
- Natural compression/tokenization emerges without hand-designed temporal granularity.
- Predictor inference overhead drops substantially: the predictor is called not at fixed frequency but only when the encoder decides something unexpected or "expensive" happened.
Essentially merges event detection and representation compression, with direct relevance to embodied/world-model inference cost optimization.
Related event: Stable JEPA Can Auto-Segment Time by Events, Cutting Inference Cost(2 posts)→
More from Embodied
- ModRetro Chromatic: Codex Edition handheld ships with a programmable cartridge — Yamapama · 2026-11-23
- Dad builds a working 3D-printed Pokedex with ESP32 and local YOLO11n, all via Hermes — Teknium · 2026-11-23
- Tesla Optimus plays piano as humanoid dexterity hits a new milestone — dioscuri · 2026-11-23
- Tesla FSD V15 roadmap: full-route nav learning and Grok voice control coming — PTrubey · 2026-11-23
- Dev 3D-prints robot main shell: 20h51m job, 833g of filament at 83% done — jakedahn · 2026-11-23
- Figure shows off F.03 humanoid robots at Formula 1 — adcock_brett · 2026-11-23