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

Andrew Ng used his latest newsletter to take a swing at one of AI's favorite doom narratives: the idea that AI is about to gut the job market. His argument is straightforward. Software engineering is the field most exposed to AI coding tools, yet hiring in that sector hasn't collapsed. The US unemployment rate sits at a stable 4.3%. Ng thinks the panic is being manufactured, partly by AI labs that benefit from sounding terrifyingly powerful, and partly by companies that would rather blame layoffs on AI than admit they overhired during the pandemic's cheap-money years.

While that debate plays out in op-eds, the actual technology keeps moving fast, and two stories this week show what real AI progress looks like on the ground. ByteDance pushed its Seedance 2.0 video generator into CapCut, the editing app used by roughly 736 million people a month, putting a serious video-and-audio generation model directly into the hands of a massive global user base. The model handles multi-shot sequences, lip-synced dialogue, and stereo audio generated alongside the video itself, and it's already trading blows with Alibaba's HappyHorse-1.0 for the top spot on independent leaderboards. It's also landed ByteDance in a fight with six Hollywood studios after a clip surfaced using the likenesses of Tom Cruise and Brad Pitt.

Meanwhile, OpenAI is shutting down Sora, reportedly because it was burning about a million dollars a day while user numbers slid from a million to under half a million. That contrast — one company retreating from consumer video generation, another doubling down by bolting it onto an app with hundreds of millions of existing users — says a lot about where the economics of this technology actually work right now.

On the infrastructure side, Nvidia's chief scientist Bill Dally laid out just how deep AI has burrowed into the company's own chip design process, in a talk with Google's Jeff Dean at GTC. Nvidia now uses reinforcement learning systems like NVCell to redesign thousands of tiny circuit layouts every time it shrinks transistors for a new manufacturing process, work that used to take eight engineers about ten months and now runs overnight on a single GPU. Another system, PrefixRL, produces arithmetic circuits that are 20 to 30 percent more efficient than human-designed equivalents, including a 64-bit adder that uses a quarter less chip area than standard tools produce. Nvidia even fine-tuned LLaMA 2 models on its own internal chip documentation to help engineers query specs and generate design code.

Dally is careful to note that full, prompt-to-finished-GPU design is still far off. But the direction is unmistakable: AI isn't replacing chip engineers wholesale, it's compressing months of grunt work into overnight runs and finding circuit layouts no human would have guessed at. That's the same theme running through the video-generation race and Ng's jobs argument — AI reshaping how work gets done, rather than simply eliminating the need for people to do it.