AI Updates on 2025-11-10

AI Model Announcements

  • Meta releases Omnilingual ASR, a suite of automatic speech recognition models supporting over 1,600 languages, including 500 low-coverage languages never before served by any ASR system. The release includes models ranging from 300M to 7B parameters, a 7B-parameter multilingual speech representation model (Omnilingual w2v 2.0), and a dataset spanning 350 underserved languages @AIatMeta

AI Industry Analysis

  • Gamma reaches $100M ARR profitably with only 50 employees ($2M ARR per employee) and achieves a $2.1B valuation in Series B funding led by a16z, demonstrating the efficiency of AI-native companies in disrupting established categories like presentation software @thisisgrantlee
  • Venture firms are increasingly skipping due diligence entirely to remain competitive, with examples including a $10M offer to an unincorporated startup and a $20M Series A closed in 2 days with no dataroom opens @deedydas
  • Interest rates, not AI, are identified as the primary driver of job market changes, with job losses beginning several months before ChatGPT's November 2022 release, following the end of 11 years of zero interest rates @GergelyOrosz
  • Yale University economists find AI has had zero major effect on jobs so far, with job shifts measured by dissimilarity index moving only slightly faster than during computer and internet eras, and no significant change in unemployment patterns among AI-exposed roles @rohanpaul_ai
  • Cursor CEO Michael Truell reveals the company uses a two-day onsite work trial for all engineering and design hires to test end-to-end codebase capabilities and cultural fit, even at 200+ employees @a16z
  • Scribe reaches 78,000 enterprise customers, including 45% of Fortune 500 companies, using their platform to capture and optimize workflows @scottbelsky

AI Ethics & Society

  • Ethan Mollick warns that many systems are still built around the assumption that quality writing and analysis are costly and meaningful signals, but these systems are not ready for the revelation that this is no longer true with AI @emollick
  • Andrew Curran predicts a political fight over Reddit's place in the data ecology in 2026, noting the tension between the administration's focus on ideological content of training corpora and the fact that OpenAI and Google each pay Reddit over $60M annually for training data @AndrewCurran_
  • Mustafa Suleyman emphasizes that superintelligence must be built for humanity's sake, not just for its own sake, warning that it won't be a better world if we lose control of it @mustafasuleyman

AI Applications

  • Perplexity's Comet Android users are completing coding projects on Vercel from their phones, demonstrating the potential of general agents on mobile devices to provide significant agency on the go @AravSrinivas
  • Google Gemini promotes its capabilities as a personalized study partner for students, allowing them to upload PDFs, slides, photos of diagrams, and handwritten notes, then summarize readings, explain concepts, and create custom practice quizzes @GeminiApp
  • OpenAI launches one year of free ChatGPT Plus for US service members within 12 months of separation or retirement, and US veterans who have left the military in the last 12 months @kevinweil
  • Suhail describes a shift in coding approach with AI, preferring to ask AI to show step-by-step instructions for understanding rather than having it write code directly, especially for ML code where understanding tensor shapes and architecture changes is critical @Suhail
  • Nathan Lambert releases research on character training in AI, exploring how easy it is to craft personalities like sycophantic chatbots and how this will change as systems move from chat to agents @natolambert

AI Research

  • Fei-Fei Li publishes an essay on spatial intelligence as the next frontier for AI, arguing that truly spatially intelligent world models must achieve three essential capabilities: creating with a storyteller's imagination, navigating with a first responder's fluency, and reasoning about space with scientific precision @drfeifei
  • Researchers release Gelato-30B-A3B, a state-of-the-art computer grounding model achieving 63.8% on ScreenSpot-Pro and 69.1% on OS-World-G, outperforming specialized models like GTA1-32B and VLMs approximately 8 times its size like Qwen3-VL-235B @anas_awadalla
  • Researchers release SYNTH, a fully synthetic generalist dataset for pretraining, along with two new state-of-the-art reasoning models. Baguettotron, trained exclusively on this dataset with only 200 billion tokens, achieves best-in-class performance in its size range @Dorialexander
  • Tsinghua University and Shanghai Jiao Tong University paper receiving perfect scores at NeurIPS 2025 finds that Reinforcement Learning with Verifiable Rewards (RLVR) improves accuracy but doesn't create new reasoning patterns, with the base model still determining the upper limit of reasoning ability. The research suggests distillation, not RL, shows genuine signs of emergent reasoning @jiqizhixin
  • The Longitudinal Expert AI Panel (LEAP) launches with 339 top experts providing monthly forecasts for three years on AI capabilities, adoption, and impact. Experts predict major effects by 2030 including 7x increase in AI's share of US electricity use and 9x increase in AI-assisted work hours, and by 2040, 30% of adults using AI for companionship daily and 60% chance of AI solving a Millennium Prize Problem @Research_FRI
  • MIT researchers develop new nanoparticles that enhance mRNA delivery, potentially reducing vaccine dosage, costs, and side effects, with the goal of achieving safe and effective vaccine responses at much lower doses @MIT
  • Francois Chollet releases the latest edition of Deep Learning with Python, focusing on building deep intuition through theory and mental models alongside practical programming patterns, using Keras 3 as a framework-agnostic API with JAX for state-of-the-art performance @fchollet
  • Simon Willison questions how much baked-in knowledge an LLM needs to be useful, asking whether specialist coding models can be trimmed down by stripping out detailed knowledge of human history and geography, referencing Andrej Karpathy's concept of "cognitive core" @simonw
  • PyTorch announces that Arm's Neural Graphics Development Kit now supports the full ML lifecycle for real-time rendering, from PyTorch-based training to deployment with ExecuTorch, as demonstrated at PyTorch Conference 2025 @PyTorch