Seedance Makes A Splash, Nvidia's AI-Guided Chip Designs, Helping Robots Not Forget

ByteDance just handed a serious video-generation tool to an audience most AI labs would kill for. Seedance 2.0, which debuted in China earlier this year, is now live inside CapCut's paid tier across Southeast Asia, Latin America, Africa, the Middle East, parts of Europe, Japan and the US. Feed it text, images, audio or up to three video clips, and it spits out four-to-fifteen-second shots with synced dialogue, ambient sound and camera moves dictated by the prompt. It can also edit existing footage, extend clips, or splice a reference subject into a new scene, generating multiple shots in one pass instead of stitching separate clips together.

The benchmarks are close but flattering. On Arena AI, Seedance 2.0 edges Alibaba's HappyHorse-1.0 by a hair in both text-to-video and image-to-video, though both scores are still marked preliminary. On Artificial Analysis, HappyHorse-1.0 actually leads three of four categories, with Seedance 2.0 winning only the audio-synced image-to-video slot. ByteDance itself admits the model struggles with detail stability, distorted audio and rendering text cleanly. And the launch wasn't clean either — a clip circulating with Tom Cruise and Brad Pitt likenesses prompted six major Hollywood studios to demand ByteDance stop training on copyrighted material, a dispute that's still unresolved.

What actually separates ByteDance from the pack isn't the leaderboard position, it's distribution. CapCut reportedly counts 736 million monthly mobile users, second only to ChatGPT among consumer AI products. OpenAI, meanwhile, is walking away from this fight entirely: Sora's daily active users fell from around a million at launch to under 500,000, while the app reportedly cost roughly a million dollars a day to run, and OpenAI is now shutting it down and redirecting compute toward coding and business tools. Chinese developers are filling the vacuum fast — Alibaba's HappyOyster for 3D environments and Tencent's open-sourced Hunyuan 3D both landed within days of each other.

On the hardware side of the AI world, Nvidia's chief scientist Bill Dally laid out, in a GTC conversation with Google's Jeff Dean, just how much AI is already baked into designing Nvidia's own chips. His team of roughly 300 researchers uses AI across five stages of chip design, and the results aren't subtle. NVCell, a system pairing a genetic algorithm with a reinforcement-learning agent, redesigns the 2,500 to 3,000 reusable layout cells needed every time Nvidia shifts to a new manufacturing process — work that used to tie up eight engineers for about ten months now finishes in an overnight run on a single GPU, matching or beating human results on area, power and speed.

A sibling system called PrefixRL goes further, designing the arithmetic circuits inside GPU math units. Dally called the resulting layouts "bizarre," but they're 20 to 30 percent better than human-designed equivalents — one 64-bit adder came out 25 percent smaller than what standard industry tools produce. Nvidia also fine-tuned LLaMA 2 models on its own internal design documents to build ChipNeMo and BugNeMo, tools that answer engineering questions and summarize bug reports using a fraction of the parameters a general-purpose model would need. Still, Dally was clear that prompting an AI to design a whole GPU end-to-end, then heading off to ski for a few days, remains a distant goal — verification alone is still the longest step in the pipeline, and nobody's cracked how to compress it fully with AI yet.