Tokens by Michael Brenndoerfer, Week #11

Tokens

Fal Cracks Real-Time Infinite Video Streams

Today's issue: Fal makes infinite video faster than real time, Nvidia circles Hugging Face, and Brussels drops the hammer on ChatGPT.

Welcome to #11 of your weekly AI newsletter.

Don't miss the next one 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

No spam. Unsubscribe anytime.


Smiling cartoon computer

Main Story

A developer staring intently at live interactive streaming video while completely ignoring traditional slow video generation pipelines.
When your video generation is finally faster than your binge-watching speed.

Fal Breaks the Infinite Video Speed Barrier

For years, AI video has been a test of human patience. You typed a prompt, stared at a loading bar, went to grab an espresso, and returned to a four-second clip of a plastic dog melting into a couch. That batch era just hit a wall. Fal rolled out H3 Max Live, a heavily post-trained and optimized build of MiniMax's H3 video foundation model that runs 35x to 50x faster than the official baseline endpoint.

By crossing the faster-than-realtime threshold, the system synthesizes video frames faster than the human eye can consume them. Instead of generating isolated clips that stop after a few seconds, the model sustains continuous, unbounded video feeds that react to dynamic inputs without buffering.

Developers immediately wired up live prompt steering. Creators are now running infinite interactive video streams guided on the fly by chat prompts, prompting Fal to launch a dedicated interactive streaming service to support the new wave of real-time applications.

  • 50x Acceleration: Post-training and inference optimizations cut generation latency below playback speeds.
  • Infinite Generation: Removes rigid clip length limits, allowing models to stream continuous video feeds without interruption.
  • Dynamic Steering: Allows users and chat interfaces to alter scenes, camera angles, and subjects in real time.

Bottom Line: AI video is graduating from static clip generation to real-time interactive media, turning what used to be a rendering queue into a dynamic visual engine.

Source: www.latent.space

Don't miss the next one 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

No spam. Unsubscribe anytime.


Smiling cartoon star

The Feed

  1. Nvidia is reportedly buying Hugging Face for $12.9 billion. The deal values the hub at an enormous revenue multiple, locking down the premier open-source AI ecosystem just as major model labs ramp up custom silicon efforts.

  2. The European Commission designated ChatGPT as a Very Large Online Search Engine, giving OpenAI four months to meet strict systemic risk rules under the Digital Services Act or face fines reaching 6 percent of global revenue.

  3. Runway unveiled Solaris, a real-time Interface World Model that streams dynamic websites and UI applications entirely out of generated pixels without touching HTML or CSS, offering a rich playground for computer-use agents.

  4. Tencent open-sourced Hy4 preview, an ambitious 770B parameter mixture-of-experts model activating 49B parameters per pass. It sports a 1M token context window aimed squarely at complex coding and enterprise productivity tasks.

  5. DeepSeek released open weights for DeepSeek-V4-Flash-Vision-Exp under an MIT license. The 305B multimodal model packages vision encoders onto its MoE core across 168GB of Safetensors files with ready vLLM serving support.

  6. Over 100 tech giants sent an urgent warning to critical infrastructure sectors about rapid advances in automated cyberattacks. Defenders are urged to deploy AI-driven tooling before cheap automated exploits overrun older setups.

  7. Sony, EMI, and Warner Chappell sued Anthropic over alleged torrenting of copyrighted songbooks and lyrics, citing internal staff chats praising shadow libraries while Anthropic maintains its training practices are fair use.

  8. Skild AI launched S1, a general-purpose robotic brain capable of learning 10-minute physical tasks from a single prompt video. The omni-bodied foundation model operates across quadrupeds, static arms, and humanoid hardware.

  9. Perceptron AI dropped Isaac 0.5, an open-weight 36B embodied foundation model scoring 97.2% on LIBERO manipulation. Its training laws show that broad video datasets can cut expensive physical teleoperation needs by 210x.

  10. SparkLLM open-sourced Spark X2.5-4B and 1.7B, bringing native 1M token context windows to compact edge models. Pretrained on domestic compute, the lightweight checkpoints integrate smoothly into llama.cpp and Ollama runtimes.

  11. Ant Group quietly uploaded UI-Venus-2-9B to Hugging Face with no announcement. The Apache-2.0 vision-language GUI agent boasts near-SOTA scores across OSWorld and WebVoyager while running standard vLLM serving endpoints.

  12. EuroHPC JU signed a 387.8 million euro contract to build LUMI-AI in Finland. Powered by AMD Instinct MI430X GPUs and EPYC chips, the green installation will expand European AI compute capacity tenfold by late 2027.

  13. Hugging Face subsidiary Pollen Robotics crossed $4 million in sales for its $399 Microduck robot. Powered by Rockchip edge silicon, the open-source platform has seen overwhelming demand push deliveries into 2027.

  14. AWS rolled out Agent Registry into general availability, offering enterprises a central searchable catalog to track, govern, and discover internal AI agents and custom tools while curbing unmonitored workflow sprawl.

  15. Google launched Gemini 3.5 Transcribe, an audio-to-text model engineered to parse dense technical jargon and background noise into cleanly formatted prose across live streaming feeds and batch enterprise analytics.

See you next week!

Michael

Smiling cartoon mobile phone

Do you want to be mentioned in the newsletter, or do you have feedback?

Get in Touch
Smiling cartoon computer

Get Tokens in your inbox 📬

Everything that mattered in AI this week, in five minutes. Occasionally hilarious.

Join 5,000+ readers · No spam, unsubscribe anytime.