Simon Mo, CEO of Inferact and lead maintainer of vLLM, discusses how open-source AI, exemplified by vLLM, transformed into critical infrastructure. The conversation highlights the technical complexities of serving LLMs, the evolving economics and licensing of open-weight models, the need for control over guardrails, and the rapidly disappearing capability gap between open and proprietary AI.
Jeff Dean discusses the evolution of AI, drawing parallels between Google's past breakthroughs (like fitting search in RAM and the origin of TPUs) and current challenges. He emphasizes that AI is becoming an energy problem, driving the need for specialized inference hardware. Dean highlights 'context engineering' and multi-agent systems as crucial for long-running, complex AI tasks, and offers advice for startups on finding niches where they can outperform larger entities by focusing on specific domains, data, and models. He stresses the importance of clear specifications for agents and 'taste' in problem selection, encouraging founders to question fundamental assumptions and automate the scientific method to build 'AI that builds AI.'
Nathan Lambert and Florian Brand discuss the accelerating open model landscape, focusing on the surge in Chinese models like Kimi K3 and GLM 5.2. They explore the reasons behind China's progress, the evolving US ecosystem, the cybersecurity implications of open-source bans, and debunk common misconceptions about distillation, particularly challenging Ben Thompson's recent claims. The episode concludes with predictions and a frontier model tier list.
Ziv Ilan from NVIDIA details how latency in video diffusion models can be drastically reduced to achieve real-time generation. He presents a layered approach combining dynamic quantization for memory and speed, chunk-based caching to skip redundant denoising computations, and, most critically, step distillation—training models to achieve high-quality output in significantly fewer steps. These techniques, packaged in the open-source FastGen repository, offer additive performance gains, enabling real-time video on a single Blackwell B200 GPU.
Sander Dieleman from Google DeepMind provides a behind-the-scenes look at the key components of training large-scale diffusion models for audio-visual data. The talk covers the entire pipeline, from the critical role of data curation and latent representations to the mechanics of diffusion, network architectures, sampling with guidance, and advanced control signals.