DeepSeek V4 Changes the Math on Frontier AI
DeepSeek just released V4 with fully open weights. One trillion parameters, Mixture-of-Experts architecture, and it cost roughly $5.2 million to train.
That number should make every AI lab uncomfortable. We are talking about frontier-level performance competing with Claude Opus 4.6 for a fraction of what US labs spend. The cost curve keeps bending the wrong direction for closed-source business models.
GPT-5.4 Thinking Beats Humans at Desktop Tasks
OpenAI's GPT-5.4 Thinking variant scored 75% on OSWorld-Verified, officially surpassing human-level performance on desktop task benchmarks. This is the test where AI agents use actual operating systems to complete real work.
I have been skeptical about desktop agents for a while. But 75% is no joke. That is better than most humans doing the same tasks under time pressure.
Netflix Drops Its First Public AI Model
Netflix released VOID on Hugging Face. It stands for Video Object and Interaction Deletion. Basically removes objects from video seamlessly.
Smart move. Netflix has mountains of proprietary video data, and releasing a niche tool builds goodwill with the open-source community without giving away the crown jewels.
Quick Hits
Google shipped Gemma 4 under Apache 2.0 with models from edge to datacenter. Microsoft released 3 MAI models for text, voice, and image generation. OpenAI acquired TBPN, the tech news show that pulled $5M in ad revenue last year. And Macy's AI shopping assistant is driving 4.75x more spending per visit. Retail AI is quietly printing money.