AI Updates on 2025-12-13

AI Model Announcements

  • OpenAI's GPT-5.2 exceeded a trillion tokens in the API on its first day of availability and continues growing rapidly @sama
  • Google rolled out an updated Gemini Native Audio model with higher precision function calling, better realtime instruction following, and smoother conversational abilities, now available to developers in the Gemini API @OfficialLoganK
  • Google launched Gemini 3 Pro with new capabilities for local search results integration with Google Maps, displaying photos, ratings, and real-world information in a rich visual format @GeminiApp
  • Sora released three new video generation styles: Handheld, Retro, and Festive, available to all users on web, iOS, and Android @soraofficialapp

AI Industry Analysis

  • Anthropic is reportedly in discussions with Google for a compute deal valued in the high tens of billions, with reports suggesting orders of $21 billion worth of TPUs to train larger models @AndrewCurran_
  • OpenAI and Disney deepened their partnership, with Disney receiving warrants to buy more OpenAI shares at current valuation, potentially creating stronger future ties between the companies @AndrewCurran_
  • China's Ministry of Industry and Information Technology reportedly issued guidelines prioritizing H200 GPU imports for companies capable of training models like Alibaba, Tencent, ByteDance, and DeepSeek, while restricting access for resellers and traditional enterprises doing inference @jukan05
  • Research on LLM pricing found short-run elasticity around 1, suggesting no immediate Jevons Paradox, but prices fell 1000x in two years while demand exploded, indicating the paradox occurs over time as firms gradually adopt AI at lower prices @emollick
  • Study estimates that ChatGPT led to a 6% differential increase in new startups between high-AI and low-AI adoption areas in China, demonstrating measurable economic impact on entrepreneurship @emollick
  • Gartner's credibility in AI analysis is being questioned after their AI coding assistants report ranked Amazon, GitLab, and GCP above Cursor while omitting Claude Code and OpenAI Codex entirely, with allegations that vendors pay for favorable rankings @GergelyOrosz
  • The AI coding assistants market shows dynamic competition with frequent leadership changes across different spaces, while many companies have not yet leveraged powerful AI models outside of coding and tech, often choosing cheaper options @emollick
  • Hugging Face is shipping 3,000 Reachy Mini robots worldwide, described as one of the largest AI robot shipments of the year, designed as an open-source DIY robotics platform for AI builders @ClementDelangue
  • GPT-4 level capabilities becoming 1000x cheaper in 2 years is critical for near-term economic impacts, as current dirt cheap AI capabilities suffice for many useful applications that most people are not fully leveraging @RishiBommasani

AI Applications

  • OpenAI adopted Anthropic's skills mechanism in both ChatGPT and their Codex CLI tool, with ChatGPT now featuring skills for creating and manipulating spreadsheets, docx files, and PDFs in a new /home/oai/skills folder @simonw
  • ChatGPT's new PDF skill was used to create a detailed report on the year's Kakapo breeding season, taking 11 minutes as it iteratively rendered and fixed issues like special character rendering @simonw
  • Cursor shipped rapid design tool improvements including element selection without animations, blur slider rounding, backspace to delete elements, undo/redo shortcuts, and multi-element context selection @cursor_ai
  • Google launched Android Emergency Live Video, allowing users to share vital visual information with one tap to emergency services for faster situation assessment and life-saving guidance @sundarpichai
  • Users are increasingly turning to LLMs like Perplexity for recipe searches instead of Google, which returns endless text and ads before the actual recipe, demonstrating how AI search provides cleaner, more direct results similar to the early 2000s web @GergelyOrosz
  • Developer built autonomous agents using custom harness with multiple tools, GPT 5.2 for second opinions, 7.5k system prompt, and periodic context re-injection to solve weird, hard problems requiring long horizons @Suhail
  • GPT-5.2 created an interactive Excel spreadsheet for D&D monster combat simulation including special abilities after 60 minutes of thinking time, while Claude 4.5 Opus completed the task quickly but simplified by omitting special abilities @emollick
  • Claude 4.5 Opus demonstrated advanced lateral thinking by not only drawing a unicorn in TikZ but also compiling it in LaTeX, converting to PDF, then PNG, and delivering the final image with decorative elements @emollick
  • shadcn/create launched allowing developers to build customized shadcn/ui implementations by picking component libraries, icons, colors, themes, and fonts, with the config rewriting component code to match preferences beyond just theming @shadcn

AI Research

  • DeepMind released the first paper training robots with Veo-generated world models, achieving 0.88 correlation to real world success rates on 1600+ trials on ALOHA 2 bimanual robots and generalizing to out-of-distribution scenarios without real world hardware trials @deedydas
  • DeepMind released a Gemini Deep Research agent for developers via the Interactions API, enabling embedding of Google's most advanced autonomous research capabilities directly into applications @GoogleAI
  • Google Research and DeepMind introduced DeepSearchQA, a new open-source web research agent benchmark designed to test agents on complex web research tasks @GoogleAI
  • Google Research and DeepMind launched the FACTS Benchmark Suite, the industry's first comprehensive test evaluating LLM factuality across four dimensions: internal model knowledge, web search, grounding, and multimodal inputs @GoogleAI
  • Frontier AI models show surprisingly little divergence in abilities, prompt adherence, and other factors, with American closed source models, Chinese models, and French open models all performing very similarly to each other @emollick
  • Meta's computer use agents team leader resigned after 1.45 years of building CUA infrastructure, data pipelines, evals, and models from scratch to achieve frontier level computer use agent performance @kohjingyu