Today in Generative Media

Senators propose AI copyright bill; Netflix uses Runway ML; AI is getting less reliable

News and Opinion

• Exclusive: Hawley and Blumenthal introduce AI protection bill (Axios)

• Netflix using startup Runway AI's video tools for production - Bloomberg (MSN)

• Grok and Groupthink: Why AI is Getting Less Reliable, Not More (Time)

• AI Hiring Favors Women Over Equally Qualified Men, Study Finds (Newsweek)

• The campaign to make it illegal for ChatGPT to criticize Trump (Platformer)

• Only Critical Thinking Ensures AI Makes Us Smarter, Not Dumber (Forbes)

• Inside an AI Story Lab: Student Voices - Featuring Lorenzo Fedon's Uncertainty (Elletra Fiumi on Substack)

• Why my p(doom) has risen, dramatically (Gary Marcus on Substack)

Software

• Act-Two is now available via the Runway API, allowing you to bring our most advanced motion capture directly into your apps, products, platforms and websites. (LinkedIn)

• Trained a Kontext LoRA that transforms Google Earth screenshots into realistic drone photography (Reddit) [RealEarth-Kontext]

• Can open-data models beat DINOv2? Today we release Franca, a fully open-sourced vision foundation model. Franca with ViT-G backbone matches (and often beats) proprietary models like SigLIPv2, CLIP, DINOv2 on various benchmarks setting a new standard for open-source research (X)

Research

• FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait (project page)

• Scalable Permutation-Equivariant Visual Geometry Learning (project page)

• Cameras as Relative Positional Encoding (project page)

• Review of Feed-forward 3D Reconstruction: From DUSt3R to VGGT (arXiv)

• Generalist Forecasting with Frozen Video Models via Latent Diffusion (arXiv)

• Generative AI at the Crossroads: Light Bulb, Dynamo, or Microscope? (Board of Governers of the Federal Reserve System)

Misc

• I am no longer chairing defenses or joining committees where students use generative AI for their writing (Lizzie Wolkovich, UBC)

• For anyone having a conviction that prompting in JSON is the reason why they get better results: it’s not. It’s literally because you structure your prompts better and create sections with clear goals. I brought up something similar when sharing my findings in VEO2: structuring makes it also much easier to fix. It’s not really about the JSON itself. (X)

• "[video game] as a community theater production" may be one of the most delightful Veo 3 Fast prompts (X)

• Testing how RunwayML Act Two handles motion transfer on input video. (X)