Interested in a Guest talk?

This session explores the evolution of voice assistant systems from turn-based interactions to fully asynchronous, full-duplex assistants that can listen, respond, and perform background tasks simultaneously. We dive into the Real-Time Venus architecture, which separates fast real-time conversation from a backend capability system for complex reasoning, tools, APIs, and web searches. The session also covers control-token-based action scheduling, speech playback management, and video memory optimization using motion-based frame filtering and MMR retrieval. Finally, we examine benchmark results, including the system’s robustness to noise and continued speech, along with current challenges such as interrupt responsiveness, tool-call errors, and unnecessary task delegation.

This session focuses on Efficient Video Sampling (EVS) and how it reduces the enormous token cost of processing long videos in multimodal LLMs. We explore how video frames are divided into patches, grouped into temporal cubelets/tubelets, and compared using cosine dissimilarity to identify informative regions, while the first frame acts as an unpruned reference anchor. The session demonstrates how quantile and Top-K selection can remove redundant static regions while preserving motion and information-rich content, with Top-K providing precise control over the final token budget. We also examine how EVS can reduce a 512-frame video from ~141K to ~42K tokens, along with its limitations, including Amdahl’s Law and the finding that EVS performs better in vision embedding space than raw RGB space.

This session explores Nemotron-3-Nano-Omni and its Efficient Video Sampling (EVS) technique for making long-video understanding more efficient. We examine how the 30B MoE multimodal model combines vision, audio, and language processing, and why long videos create massive token, memory, and compute requirements. The session dives into how EVS uses anchor frames, temporal tubelets, and patch similarity to identify redundant visual information and retain motion-rich regions. We also explore supporting optimizations such as dynamic resolution, 3D tubelet fusion, and audio-video interleaving, which can reduce token counts significantly. Finally, we discuss EVS's speed improvements, end-to-end limitations, and the accuracy trade-offs of aggressive pruning, particularly for fast-moving videos.

This session explores MIRA, a multi-player interactive world model from General Intuition that learns to predict and generate future video across four simultaneous player perspectives. We dive into how MIRA combines DINOv3 visual features, player actions, temporal information, and latent diffusion to maintain a consistent shared world across multiple views. The session also covers how the model is evaluated using FVD, FID, Action Recoverability Rate (ARR), and game-state probing, along with its current limitations such as physics inconsistencies, object memory failures, and incorrect game states, and discusses future applications in robotics and autonomous driving.

This session introduces the fundamentals of image and video diffusion models and then dives into NVIDIA’s SANA family of models (SANA, SANA 1.5, and SANA Sprint). We explore how SANA makes high-resolution image generation faster and more memory-efficient using Linear DiT, Deep Compression Autoencoders, kernel fusion, model pruning, and distillation. The session also discusses how these optimizations enable smaller models to achieve high-quality results at significantly lower computational cost, while covering the key ideas behind single-step diffusion generation and setting the foundation for a future session on video diffusion models.

November 1,  2025

October 18,  2025

August 9, 2025

July 12, 2025

June 28, 2025

May 31, 2025

April 18, 2025

April 4, 2025

March 21, 2025

March 7, 2025

Feb 14, 2025

Jan 24, 2025

Jan 10, 2025

Dec 13, 2024

Nov 29th, 2024

Nov 1, 2024

September 13, 2024

August 23, 2024

August 9, 2024

July 26, 2024

July 12, 2024 

June 28, 2024