AI Updates on 2025-06-06

AI Model Announcements

  • Anthropic introduces Claude Gov, custom models built for U.S. national security customers, already deployed by agencies at the highest level of U.S. national security with access limited to classified environments @AnthropicAI
  • Google releases Gemini 2.5 Pro update with state-of-the-art long context performance, especially capable on higher number of items being retrieved @OfficialLoganK
  • Google's Veo 3 video generation model is now live on both Replicate and FAL platforms @AndrewCurran_

AI Industry Analysis

  • Cursor raises $900 million in Series C funding, reaching over $500 million in ARR and being used by more than half of the Fortune 500, including NVIDIA, Uber, and Adobe @cursor_ai
  • Uber was revealed as the company where engineers preferred Cursor over GitHub Copilot, leading to company-wide licensing for all developers @GergelyOrosz
  • AI startups are showing significantly faster revenue growth compared to pre-AI software companies, with new benchmarks emerging for AI company performance @omooretweets
  • Forward deployed engineers are becoming the hottest job in startups, representing a shift toward services-led growth in the AI era @a16z
  • Waymo's market position in San Francisco has converged to 2-3x the wait time and cost of Uber, reflecting how much more people are willing to pay for autonomous vehicles @natolambert
  • Software is becoming consumers' third biggest expense after food and rent, with AI driving increased consumer spending on software products @a16z

AI Ethics & Society

  • OpenAI opposes New York Times' court request to prevent deletion of user chats, arguing it sets a bad precedent and compromises user privacy, with Sam Altman proposing the need for "AI privilege" similar to lawyer-client confidentiality @sama
  • Simon Willison warns about prompt injection vulnerabilities in the GitHub MCP server, where attackers can trick AI agents into stealing private data through malicious instructions @julien_c
  • Less than 10% of AI-focused YouTube viewers are female, highlighting the gender gap in AI adoption and education @clairevo

AI Applications

  • Current LLMs can achieve significant accuracy improvements in clinical oncology decisions when given access to medical tools, with GPT-4 going from 30% to 87% accuracy @emollick
  • Perplexity launches daily news pushes on WhatsApp and adds financial analysis features to finance pages @AravSrinivas
  • Microsoft Copilot introduces visual search capabilities with real images, videos, and cards to make searching smarter @Copilot
  • Hugging Face partners with Google Colab to add "Open in Colab" support for all models on the Hugging Face Hub, making AI model experimentation more accessible @GoogleColab
  • Opportunity International uses Ulangizi AI chatbot to help smallholder farmers in Africa improve agricultural practices with financial services and education @Microsoft

AI Research

  • MIT CSAIL and partners release Boltz-2, the first AI model to approach FEP simulation performance for protein-binding affinity prediction while being over 1000x faster, open-sourced under MIT license @MIT_CSAIL
  • François Chollet announces ARC-AGI-2 as a better tool for measuring breakthrough AGI capability progress, while ARC-AGI-1 remains better for comparing AI systems and measuring efficiency @fchollet
  • EleutherAI releases the Common Pile v0.1, an 8TB dataset of openly licensed and public domain text, with 7B models trained on this data matching the performance of similar models like LLaMA 1&2 @AiEleuther
  • Hugging Face releases ScreenSuite, a comprehensive evaluation suite for GUI Agents with vision-only evaluation, Ubuntu & Android environments, and mobile, desktop & web coverage @amir_mahla
  • Research suggests that lightly trained 14B specialized models can regularly outperform o3 for backing real agents, highlighting the gains from specialization @corbtt
  • Current opinion suggests that Deep Research, Codex agent work by training models on short horizon RL tasks and general robustness, while training end-to-end on very sparse RL tasks remains further out @natolambert
  • MIT develops a game-changing animation technique that simulates soft, squishy motion with Pixar-level physics in real time, potentially revolutionizing animation, gaming, and robotics @MIT