AI Updates on 2025-10-23

OpenAI acquired Software Applications Incorporated, the maker of Sky, a natural language interface for Mac, to integrate their desktop AI experience into ChatGPT @OpenAINewsroom
Microsoft unveiled Mico, a new animated avatar for Copilot AI that brings back elements of Clippy as a friendly, customizable face for the chatbot @TechCrunch
Google announced advancements in Earth AI, bringing Gemini capabilities into Google Earth for instant object finding and pattern discovery from satellite imagery @GoogleAI
NVIDIA's Gr00t N1.5 cross-embodiment foundation model for robots is now available in LeRobot, featuring multimodal inputs and flow matching action transformer for action prediction @LeRobotHF
Google AI Studio introduced Annotate mode, allowing users to mark up UI with drawing tools and have Gemini action them directly in code @OfficialLoganK

Reddit sued Perplexity for allegedly engaging in industrial-scale scraping of millions of Reddit user comments, while Google pays Reddit $60 million annually and OpenAI pays about $70 million for training data access @AndrewCurran_
Executive Order 14319 requires LLMs to be ideologically neutral to qualify for government procurement, driving increased neutrality work among AI companies seeking government contracts @AndrewCurran_
Stability AI formed a strategic partnership with EA to co-develop transformative generative AI models, tools, and workflows for game development @StabilityAI
The Wall Street Journal reported that the Trump administration is considering taking equity stakes in quantum computing companies, similar to their approach with Intel @AndrewCurran_
Kensho Technologies, a $500M AI startup acquired in 2018, produced founders of six near-unicorn companies including OpenEvidence, Surge, Langchain, and Suno, demonstrating the value of joining startups with smart people @deedydas

Gergelyorosz identified em dashes as an AI smell in supportive messages for laid-off workers, noting that most people don't know how to type them manually, suggesting AI-generated content @GergelyOrosz
Yann LeCun argued that one cannot prove AI safety before building and refining AI systems, comparing it to turbojets which required actual construction and careful refinement for reliability @ylecun
Dileep George criticized the misinterpretation of Rich Sutton's Bitter Lesson, arguing that LLMs violate the principle by training on human discoveries rather than letting models discover independently @dileeplearning
Yann LeCun revealed that humanoid robot companies have no idea how to make robots smart enough for domestic use, requiring multiple breakthroughs beyond current capabilities @theneoniche

The Government of Jordan deployed an AI-powered learning assistant Siraj built on Replit to 1.6 million students and 90,000 teachers across public schools, with the pilot built in under a month by one person @Replit
Perplexity Finance now allows users to listen to earnings calls and will soon enable voice questions during audio streams @AravSrinivas
Microsoft introduced Copilot Groups for real-time collaboration, allowing teams to brainstorm, co-write, plan, or study together with AI assistance @satyanadella
OpenAI launched Shared Projects for Free, Plus, and Pro users, enabling collaborative work in ChatGPT with shared chats, files, and instructions @OpenAI
Sora is adding character cameos, video editing tools, enhanced social features, and Android app support, with trending cameos displayed in real-time @billpeeb
Meta AI's photo editing tools are now available in Instagram Stories, allowing users to describe what they want to add, remove, or change @TechCrunch
Microsoft Edge introduced Copilot Mode, an AI browser that meets users where they left off across tabs and completes multi-step actions @satyanadella

Berkeley AI researchers uncovered a Guess-then-Refine mechanism in LLMs, where early layers predict high-frequency tokens as guesses and later layers refine them as context builds @akshatgupta57
Berkeley AI presented Omni-Scan, a novel method for bimanual robot 360-degree object scanning and reconstruction using 3D Gaussian Splats @ZehanMa123
Hugging Face and Meta launched OpenEnv, a universal RL Environment interface providing frontier-grade reinforcement learning environments for the open-source community @_lewtun
NVIDIA's llama-embed-nemotron-8b achieved new #1 position on MTEB Embedding Benchmark Leaderboard, beating Gemini and Qwen3 with 69.46 average across tasks @TheAhmadOsman
Ethan Mollick observed that AI video generation maintains visual consistency across multiple clips better than audio consistency, noting that video can generate from last frames while having world-model properties @emollick