AI Updates on 2025-11-12

AI Model Announcements

  • OpenAI releases GPT-5.1 with improvements to instruction following, adaptive thinking capabilities, and customizable tone/style presets including Default, Friendly, Efficient, Professional, Candid, and Quirky options @sama
  • World Labs releases Marble, a spatial intelligence platform that enables users to create and edit persistent three-dimensional worlds, representing a groundbreaking step in building world models @theworldlabs
  • Weibo releases VibeThinker 1.5B, a reasoning model trained for just $7,800 that outperforms DeepSeek R1 on math reasoning benchmarks (AIME24: 80.3 vs 79.8) despite being 100x smaller @WeiboLLM

AI Industry Analysis

  • Anthropic announces $50 billion investment in American AI infrastructure, constructing data centers in Texas and New York that will create thousands of jobs @AnthropicAI
  • Microsoft unveils second Fairwater AI datacenter in Atlanta, connected via dedicated AI network to create an AI superfactory enabling real-time collaboration across states for training next-generation models @Microsoft
  • CoreWeave operates as a $50B business with only ten employees, generating $4.8B annual run rate by reselling Nvidia GPUs to just three customers: Microsoft (71% revenue), OpenAI ($11.9B committed over 5 years), and Meta ($14.2B committed over 6 years) @deedydas
  • Magic Patterns reaches $1M ARR with no employees, demonstrating new business models enabled by AI tools @snowmaker
  • Figma opens first office in Bengaluru, India, where 35 million Figma files were created in the past year alone @zoink
  • Global data center spending reaches $580 billion this year, exceeding oil supply investment by $40 billion, highlighting massive infrastructure shift toward AI @TechCrunch
  • Cybersecurity firm Deepwatch lays off dozens of employees, citing move to accelerate AI investment @TechCrunch

AI Ethics & Society

  • AI Now Institute releases report "Fission for Algorithms" exposing efforts to fast-track nuclear development to power AI, including using generative AI in nuclear licensing while weakening regulation @AINowInstitute
  • German court rules OpenAI violated copyright law by training language models on licensed musical work without permission, ordering damages @TechCrunch
  • OpenAI's CISO publishes letter fighting New York Times' request for indiscriminate access to 20 million user conversations, arguing for need for "AI privilege" similar to attorney-client privilege given sensitive nature of AI conversations @OpenAI
  • Spanish PM Pedro Sánchez at WEF 2025 demands end of online anonymity, calling for every social media account to be linked to EU Digital ID Wallet, raising concerns about digital surveillance @JimFergusonUK
  • Stanford HAI leaders emphasize importance of open science in AI, warning about risks of privatized AI knowledge including loss of cross-pollination of ideas, reproducibility, global participation, and talent pipeline @StanfordHAI

AI Applications

  • Microsoft announces Project Gecko, delivering affordable AI expertise to small farms in India and East Africa using small language models and speech systems, demonstrating culturally nuanced AI applications for underserved populations @MSFTResearch
  • Stanford researchers build AI training tool for PTSD therapists to practice written exposure therapy skills 24/7 before working with real patients, addressing gap between patient need and therapist training @StanfordHAI
  • San Jose Mayor reveals how AI is transforming city services, from optimizing traffic to translating public meetings in real time @NVIDIAAI
  • Waymo expands service to entire SF Bay Area Peninsula from San Francisco to San Jose, now taking riders on freeways @JeffDean
  • Google partners with Cassava Technologies to enable data-free access to Gemini App and 6-month extended trial of Google AI Plus in Africa @joshwoodward
  • ElevenLabs strikes deals with celebrities to create AI audio, expanding voice synthesis applications @TechCrunch

AI Research

  • Hugging Face releases FinePDF-edu, a high-quality dataset of 350B tokens across 69 languages filtered using Qwen3-235B and ModernBERT classifiers, outperforming previous pretraining datasets on benchmarks @Thom_Wolf
  • Google DeepMind research teaches vision models to better organize visual concepts hierarchically, making them more reliable at generalizing across different categories @GoogleDeepMind
  • Sakana AI demonstrates using Thought Cloning with YouTube videos to improve LLM reasoning capabilities @shengranhu
  • Stanford researchers create synthetic brain MRIs using generative AI to accelerate computational neuroscience and understanding of brain disorders @StanfordHAI
  • Research on LeJEPA introduces novel pretraining paradigm free of traditional heuristics, testing 60+ architectures up to 2B parameters across 10+ datasets with 95% correlation between training loss and test performance @randall_balestr
  • Ethan Mollick demonstrates different AI models show varying attitudes toward same tasks, with models rating viability of ideas differently based on their training, highlighting importance of understanding AI personality differences @emollick
  • Cursor AI data shows developers prefer models like Sonnet 4.5 and GPT-5 for planning tasks, with significant shifts in model preferences over six months @cursor_ai
  • Andrej Karpathy reports Tesla FSD v13 on HW4 delivers flawless neighborhood drives with smooth, confident performance handling complex scenarios including tight lanes, construction, tricky left turns, and autonomous parking @karpathy
  • Ashish Vaswani's ICCV25 talk reveals Tesla's approach of processing sensor streams over long contexts through large neural networks for end-to-end driving, representing complete Software 1.0 to Software 2.0 rewrite @aelluswamy