AI Updates on 2025-11-02

AI Model Announcements

  • Alibaba announces Qwen3-VL can now run locally with Unsloth AI, offering fine-tuning and reinforcement learning capabilities via free notebooks @Alibaba_Qwen

AI Industry Analysis

  • Meta's AI spending is beginning to raise concerns among Wall Street investors about the company's financial commitments @TechCrunch
  • OpenAI CEO Sam Altman revealed the company is generating well over $13 billion in annual revenue and appeared defensive when questioned about how it will fund its massive spending commitments @TechCrunch
  • YouTube has become a $60 billion ARR business growing 15% year-over-year, accounting for 15% of Google revenue, with over 2% of all human waking time spent on the platform @deedydas
  • Individual releases of open AI models only matter in the short term as they become obsolete without continued releases, with the capability/cost improvement curve being steep @emollick
  • A key question remains whether Chinese labs and Mistral will continue releasing open weights models as economic costs and value continue to scale, since open source AI lacks the same value capture mechanisms as open source software platforms @emollick
  • The end goal of the open weights AI strategy remains unclear, as unlike open source software which captures value through services or hardware, value doesn't flow back the same way from open weights models @emollick
  • The tech job market is tightening, making degrees from top CS colleges and working at companies with top brands increasingly advantageous, with building up pedigree becoming more important than before @GergelyOrosz
  • As the tech job market tightens with more qualified candidates than open positions, hiring increasingly happens by pedigree from top schools or workplaces, though algorithmic interviews give those without pedigree a fair shot @GergelyOrosz

AI Ethics & Society

  • Humanity's biggest challenges won't be solved by AI thinking for 1000 hours alone, but by many collaborating humans with AI that understands their different skills, goals, and values to empower collective action @ericzelikman
  • Yann LeCun argues that scaling up transformer-based LLMs will not achieve human-level AI, stating there's no way to get a system that can invent solutions to new problems rather than just retrieve from gigantic memory @rohanpaul_ai
  • LeCun recommends abandoning LLMs for human-level AI in favor of joint-embedding architectures, energy-based models over probabilistic ones, regularized methods over contrastive ones, and model-predictive control over reinforcement learning @rohanpaul_ai
  • Skilled people wield AI tools better than unskilled users, with great coders producing better, cleaner, more organized code faster, while those without developed skills cannot verify if AI output is award-winning or garbage @Dan_Jeffries1

AI Applications

  • Google Sheets and Excel no longer have a learning curve thanks to AI assistance, with GPT-5 Pro being particularly effective at handling complex spreadsheet tasks @natolambert
  • The importance of learning to vibe code, AI engineer, and prompt is not because building products is trivial, but because making the thing should be commodified so time and creativity can be spent on figuring out the right problem, market fit, and commercialization @clairevo
  • With 12 minutes of thinking, GPT-5 Pro suggested repurposing a known drug to treat an untreatable food allergy, matching results from an unpublished peer-reviewed study, demonstrating the potential of LLM-driven scientific discovery @DeryaTR_
  • Code agents make building websites and dynamic content highly enjoyable, enabling rapid development of tools and repositories for content creation @natolambert
  • Odyssey-2 now streams 16:9 video on large screens, demonstrating an advantage of interactive video models where real-time generated video intelligently adapts to the screen, viewer, and input device unlike pre-recorded video @olivercameron
  • Odyssey-2 generates video instantly with less than a second latency after clicking start streaming, all available for free @odysseyml

AI Research

  • A revealing test prompt asks models to write a paragraph demonstrating capabilities across multiple dimensions then explain their approach, with Claude excelling at writing and GPT-5 Pro nailing intellectual tricks @emollick
  • Reinforcement learning enhances majority vote accuracy but not pass@k, boosting the probability of correct completions already in top-k without clearly enhancing overall model capabilities according to DeepSeekMath research @cwolferesearch
  • GPT-5 is clearly less sycophantic than Claude at this point, a development worth acknowledging @xlr8harder
  • The world's best language models are far better at intricate details of RL algorithms than at providing medical advice for pet illnesses, highlighting capability gaps @natolambert
  • Claude 4.1 Opus outperforms Claude 4.5 Sonnet according to user testing @natolambert
  • MIT researchers developed BoltzGen, a generative AI model that designs proteins and peptides of any modality to bind to different biomolecular targets, unifying design and structure prediction, freely available for unrestricted academic and commercial use @MIT_CSAIL
  • MIT researchers developed a method enabling artists to design realistic simulations of elastic objects like bouncy or squishy characters for animated movies or video games @MIT