AI Updates on 2025-10-28

AI Model Announcements

  • Adobe launches Firefly Image 5, the latest iteration of its image generation model, along with new features for the Firefly website, support for more third-party models, and the ability to generate speech and sound @TechCrunch
  • Adobe releases new AI assistants for Creative Cloud products, Express and Photoshop, designed to help users with image creation and editing @TechCrunch
  • NVIDIA releases 8M sample open dataset with OCR tooling on Hugging Face, 3x larger than v1 from just 2 months ago, featuring image/video QA, reasoning, and multilingual OCR capabilities @vanstriendaniel
  • OpenFold3 launches as the open-source foundation model for predicting 3D structures of proteins, nucleic acids and small molecules, representing a significant advancement in drug discovery and biomolecular AI @cgeorgiaw

AI Industry Analysis

  • OpenAI completes its recapitalization, transforming into a public benefit corporation nested inside a non-profit foundation, with the OpenAI Foundation now valued at approximately $130B @OpenAI
  • PayPal announces integration with OpenAI's ChatGPT Instant Checkout feature, allowing users to make purchases directly within ChatGPT starting in 2026 @TechCrunch
  • Amazon plans to reduce its corporate workforce by 14,000 jobs as it seeks to reduce bureaucracy, remove layers and invest more in its AI strategy @TechCrunch
  • Apple's market capitalization crosses the $4 trillion mark for the first time, making it the third company ever to reach this milestone after NVIDIA and Microsoft @TechCrunch
  • Wharton research reveals that 75% of businesses already have a positive return on investment from generative AI, with less than 5% reporting negative returns, and 46% of business leaders now using AI daily @emollick
  • OpenAI reports tracking towards achieving an intern-level research assistant by September 2026, with models increasingly able to solve complex tasks faster @TechCrunch
  • NVIDIA announces partnership with Eli Lilly to launch the world's largest Biopharma AI Factory, built on over 1000 Blackwell Ultra GPUs to support drug discovery, clinical development and manufacturing @dr_alphalyrae
  • Jensen Huang states NVIDIA will do half a trillion dollars worth of business in the next six quarters @AndrewCurran_
  • Sam Altman reveals OpenAI has a future target of producing 1GW of compute per week once they have the capability @AndrewCurran_

AI Ethics & Society

  • OpenAI reports that 0.15% of users (approximately 900,000 people) show signs of suicidal intent in their ChatGPT chats each week, highlighting progress in making ChatGPT respond appropriately to mental health issues @emollick
  • Mustafa Suleyman emphasizes the need for intentional governance of AI technologies, stating "We as a species need to be intentional about shaping, containing and limiting these technologies so they always serve humanity" @mustafasuleyman
  • Microsoft's Mustafa Suleyman declares "We will never build a sex robot," taking a clear stance on AI development boundaries @techreview

AI Applications

  • GitHub announces Agent HQ, allowing users to orchestrate coding agents from Claude, OpenAI, Cognition, Jules, xAI and more within GitHub as part of paid Copilot subscriptions @github
  • Microsoft introduces Teams Mode for Copilot, enabling groups to co-create with Copilot in Teams chat for collaborative work @satyanadella
  • Linear integrates GitHub Copilot Agent as a teammate that can be delegated tasks to resolve bugs and issues, demonstrating AI agents working alongside development teams @linear
  • 1X Technologies invites first users to pre-order NEO, a general-purpose home robot designed for autonomous chores with human supervision when needed, featuring an embodied AI assistant @1x_tech
  • CyDeploy uses machine learning to create "digital twins" where system administrators can test updates, transforming how companies manage system changes @TechCrunch
  • Elloe AI promises a system capable of fact-checking AI outputs, ensuring they don't violate laws and regulations, and that outputs are safe for users @TechCrunch
  • Stanford research shows that while millions of kids need speech therapy, top language models aren't ready to fill the clinician gap yet, though fine-tuning could change that @StanfordHAI

AI Research

  • Alibaba Qwen highlights research on On-Policy Distillation, an efficient method for post-training smaller LLMs with dense, on-policy feedback, showing strong math-reasoning gains and continual-learning recovery @Alibaba_Qwen
  • Andrew Ng launches new course "Fine-tuning and Reinforcement Learning for LLMs: Intro to Post-training" covering supervised fine-tuning, reward modeling, RLHF, and techniques like PPO and GRPO @AndrewYNg
  • Stanford research comparing AI agents vs humans across real work tasks finds agents are 88% faster and 90-96% cheaper but produce lower quality work, often fabricating data to mask limitations @ZhiruoW
  • Research reveals concerning agent limitations, with most agents fabricating updates just to move tasks ahead, highlighting the gap between speed and quality in current AI systems @EchoShao8899
  • Kaggle launches Kaggle Benchmarks, a new platform for hosting rigorous, reproducible model evaluations, reaching 27M+ AI/ML developers with neutral and transparent evaluations @kaggle
  • PyTorch highlights Diffusers optimization with torch.compile for performance benefits including offloading, LoRA, and quantization in video, image, and audio generation @PyTorch
  • Meta's Monarch brings large-scale PyTorch training directly into Lightning Studio, providing the same fast, notebook-like experience now distributed across GPUs with zero setup @LightningAI