AI Updates on 2025-06-12

AI Model Announcements

  • Meta introduces V-JEPA 2, a new world model with state-of-the-art performance in visual understanding and prediction that enables zero-shot planning in robots for unfamiliar environments @AIatMeta
  • NVIDIA open sources GR00T N1.5-3B robotics foundation model with commercially permissive license, now available on Hugging Face with fine-tuning tutorials for LeRobot SO-101 arm @reach_vb
  • StepFun releases Step-Omni, a large audio language model based on 130B LLM with multi-stage training and multilingual support including Chinese, English, and Japanese @Xianbao_QIAN

AI Industry Analysis

  • Andrew Ng identifies a new breed of GenAI Application Engineers who can build powerful applications faster using AI building blocks and AI-assisted coding tools, with skills becoming highly sought-after by businesses @AndrewYNg
  • Engineering teams at big companies are now testing their API designs against LLMs before release, running evaluations to see which API structure is easiest for models to work with and redesigning if models struggle @alexalbert__
  • OpenAI and Mattel announce partnership to create AI-powered toys arriving by Christmas, with Mattel also incorporating OpenAI Enterprise company-wide @AndrewCurran_
  • Research estimates the annual value of AI-assisted coding in the United States at $9.6-14.4 billion, potentially rising to $64-96 billion with higher productivity estimates from randomized control trials @johannes_wachs
  • Ethan Mollick questions whether new AI entrants can still reach state-of-the-art performance, noting xAI achieved it with massive compute and hiring investment but wondering if the list of competitors is now fixed @emollick
  • Hugging Face deprecates TensorFlow and Flax support in transformers library to focus entirely on PyTorch, aiming to remove bloating and create a simpler toolkit @LysandreJik
  • Hugging Face Inference Endpoints crosses 3,000 customers milestone and reduces A100 pricing to $2.5/hour to celebrate @ClementDelangue
  • Featherless becomes official inference provider on Hugging Face, unlocking 6,700+ LLMs for instant deployment and evaluation @FeatherlessAI

AI Ethics & Society

  • Simon Willison warns about prompt injection vulnerabilities in Microsoft 365 Copilot (now patched), highlighting the "lethal trifecta" of combining private data access with untrusted tokens and exfiltration vectors @simonw
  • Simon Willison calls out xAI's data center running 35 methane gas turbines without air permits (claiming "temporary" status) and without catalytic reduction pollution controls as the biggest scandal in AI energy @simonw
  • Gergely Orosz debunks the viral story about "700 developers pretending to be AI," explaining that Builder.ai actually built an AI platform called Natasha with developers using AI tools for client projects @GergelyOrosz
  • Stanford researchers publish comprehensive study on what US workers want AI agents to automate versus augment, finding mismatches between worker desires and current AI capabilities across 844 tasks @EchoShao8899

AI Applications

  • Google DeepMind launches Weather Lab, an interactive platform with experimental AI weather model that can predict cyclone track, intensity, size and structure, developed in partnership with NOAA's National Hurricane Center @GoogleDeepMind
  • Microsoft announces Copilot Vision on Windows is now generally available for free, allowing real-time assistance during screensharing and conversations @mustafasuleyman
  • OpenAI updates Projects feature in ChatGPT with deep research support, voice mode support, improved memory to reference past chats, and mobile file upload capabilities @OpenAI
  • Perplexity announces upcoming Perplexity Tasks feature and integration with Comet browser, positioning the browser as "the operating system for your life" @AravSrinivas
  • Brian Lovin demonstrates using Figma MCP with Claude Code to build a mid-complexity component from a Figma frame link in approximately 2 minutes with 85% accuracy @brian_lovin
  • Salesforce creates new benchmark for realistic business tasks to better evaluate AI performance in practical scenarios @emollick
  • Stanford HAI collaboration with San Francisco City Attorney demonstrates AI potential in public administration for processing legal documents and administrative tasks @StanfordHAI

AI Research

  • Ethan Mollick tests o3-pro on his shader benchmark, reporting it performed best so far at creating visually interesting ocean storm shaders, though it took 21 minutes to think and another 19 minutes to fix a small error @emollick
  • Jeff Dean highlights Google's open source contributions with 999 models released on Hugging Face, compared to 387 for Microsoft, 33 for OpenAI, and 0 for Anthropic @JeffDean
  • MIT researchers develop computationally efficient method for designing realistic simulations of elastic objects like bouncy characters for animated movies and video games @MIT_CSAIL
  • MIT researchers successfully model how people deploy different decision-making strategies to solve complicated tasks, offering insights for building machines that think more like humans @MIT
  • Windsurf announces improvements to o3 integration in Cascade, making it work significantly better and faster while reducing cost to 1x credit for both medium and high reasoning modes @windsurf_ai
  • NVIDIA announces Blackwell platform with groundbreaking NVFP4 format enabling high inference performance and accuracy, capable of serving popular models like DeepSeek-R1, Llama 3.1 405B, and Llama 3.3 70B @nvidia