· 3 items
Latent Space ↗
[AINews] Fal’s H3 Max Live breaks the infinite videogen barrier
- What
- Fal released H3 Max Live, breaking the infinite video generation barrier.
- Why
- Enables generating decent video faster than real-time playback.
- Impact
- Test for real-time video agent output.
arXiv ↗
Stride-k Subsampling: Train-Free Audio Token Reduction for Whisper
- What
- Stride-k subsampling introduces train-free audio token reduction for Whisper.
- Why
- Reduces ASR encoder representation redundancy without retraining.
- Impact
- Speeds up voice agent transcription pipelines.
arXiv ↗
A Universal Context-Reuse Layer for Cross-Model KV Sharing
- What
- A universal context-reuse layer enables cross-model KV sharing in LLM serving.
- Why
- Eliminates redundant prefill compute across shared context inputs.
- Impact
- Cuts inference costs for multi-model voice apps.