AI Updates on 2025-08-20

AI Model Announcements

  • Google announces Veo 3 video generation model with sound capabilities, allowing users to turn words or photos into videos with audio @AndrewCurran_
  • Google releases new Gemini Nano model powering the Pixel 10 series, featuring improved personalization and proactive assistance @Google
  • ByteDance releases Seed-OSS 36B LLM on Hugging Face, featuring powerful long-context, reasoning, and agentic capabilities @HuggingPapers
  • IBM and NASA release Surya, the first open-source AI foundation model for heliophysics with 366M parameters, trained on 9 years of Solar Dynamics Observatory data to predict space weather @ClementDelangue
  • NVIDIA's Cosmos Reason 7B-parameter VLM achieves over 500,000 downloads on Hugging Face, designed for physical AI and robotics applications @NVIDIAAIDev

AI Industry Analysis

  • Perplexity reports serving over 300 million user queries weekly, representing 3x growth in approximately 9 months from their previous 100M weekly milestone @AravSrinivas
  • EliseAI raises $250M Series E led by a16z, surpassing $100M ARR as an AI property manager and healthcare administrator addressing friction in housing and healthcare industries @aleximm
  • Gergely Orosz observes peak AI hype with investors funding questionable AI startups like mattress companies using AI to "fix sleep" and AI-powered jewelry, suggesting FOMO-driven investment decisions @GergelyOrosz
  • Microsoft announces expanded partnership with NFL, bringing Copilot and Azure AI Foundry to football operations both on and off the field @satyanadella
  • Anthropic launches Claude Code for Team and Enterprise plans with flexible pricing, allowing organizations to mix standard and premium seats across their teams @claudeai

AI Ethics & Society

  • Harvard students who previously developed facial recognition app for Meta's Ray-Ban glasses are launching a startup making smart glasses with always-on microphones, raising privacy concerns @TechCrunch
  • Gergely Orosz suggests AI tooling going mainstream will help non-technical people understand why building good software is difficult, as they experience the gap between expectations and reality @GergelyOrosz

AI Applications

  • Google introduces Magic Cue on Pixel phones, using Gemini capabilities to proactively surface helpful information and actions across apps when needed @GoogleAI
  • Google Photos launches conversational editing feature allowing users to make photo changes by describing them in natural language @TechCrunch
  • Google announces Voice Translate for Pixel phones, enabling real-time call translation using the caller's voice for more authentic multilingual conversations @GoogleAI
  • Google introduces Camera Coach using Gemini models to read scenes and provide guidance for perfect photography shots @GoogleAI
  • Perplexity launches SuperMemory feature in final testing stages, claiming superior performance compared to existing memory solutions @AravSrinivas
  • Perplexity introduces Max Assistant mode on Comet for subscribers, capable of running long-horizon research tasks contextually to reading content @AravSrinivas
  • Sierra demonstrates AI agent simulations for testing, including voice simulations with background noise to harden agent performance before deployment @btaylor
  • Brex's AI agent built on Sierra platform answers customer questions 90% faster, saving customers 15,000 hours annually @btaylor
  • Carbon Robotics uses AI-powered laser weeding robots that have destroyed 15 billion weeds across 100+ crops without herbicides, delivering dramatic yield increases @NVIDIAAI
  • Google introduces Pixel Journal, a new journaling app using on-device AI to suggest personalized writing prompts @TechCrunch
  • Google announces AI-powered personal health coach built with Gemini coming to Fitbit devices @TechCrunch

AI Research

  • Microsoft Research introduces GPT-5 Pro demonstrating capability to prove new mathematical theorems, successfully proving a better bound than published in a convex optimization paper @SebastienBubeck
  • Berkeley AI Research presents XQuant, achieving 10-12.5x memory savings versus FP16 with near-zero accuracy loss by leveraging underutilized compute units for KV cache rematerialization @adityastomar_
  • Cursor team rebuilds MoE layers at kernel level with MXFP8, achieving 3.5x faster MoE layer performance and 1.5x end-to-end training speedup @stuart_sul
  • PyTorch introduces ZenFlow for LLM training with offloading, delivering 5x faster training, 85% fewer GPU stalls, and 2x lower I/O overhead @PyTorch
  • Microsoft Research releases MindJourney enabling AI to navigate and interpret 3D environments from limited visual input for improved navigation and planning tasks @MSFTResearch
  • Nathan Lambert analyzes the spectrum of reasoning effort in AI models, noting that all current models use similar reinforcement learning techniques with varying token usage rather than binary reasoning classifications @natolambert
  • Ethan Mollick demonstrates AI video generation capabilities by creating music videos from academic paper abstracts, showcasing evolving consistency in character generation and lip syncing @emollick
  • Simon Willison tests Qwen-Image-Edit model on 64GB M2 MacBook Pro, generating rainbow-colored pelican images in 25 minutes with 10 inference steps, compared to 2 hours 59 minutes for full 50 steps @simonw