AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Latent Space Newsletter: AI News - Fal's H3 Max Live breaks the infinite videogen barrier

· Latent Space

The Latent Space newsletter reports on Fal's H3 Max Live, an autoregressive continuous version of H3 Max that allows for continuous, audience-steerable video generation, achieving real-time factor 1 at 768p in early preview. The newsletter also highlights Meta's Muse Code SDK for custom agents, DeepSeek V4 Flash Vision weights, and the strong performance of GLM-5.3 Flash and Qwen3.8-Flash-Next on agentic cost/performance benchmarks. Tencent Hunyuan's Hy4 Preview is noted as a 770B MoE model with over 1M context, showing gains in coding and agent stability. Further sections cover Hermes Agent's feature release for multi-agent workflows, the evolving DeepSeek Harness, and context management as a research frontier with WikiSkill/SKILL.state and Tencent's ContextPilot. Inference, compute, and AI infrastructure discussions include Apple hardware as a potential bottleneck for computer-use RL, Together AI and HUMAIN's 250MW Saudi data center for open models, and fragmentation in inference specialization and serving architecture. Runway introduced Solaris, an "Interface World Model" for generating interactive interfaces, and LeVJEPA offers a compute-efficient route to temporal representation learning.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.