AI Updates on 2025-08-11

AI Model Announcements

  • Meta FAIR's Brain & AI team won 1st place at the Algonauts 2025 brain modeling competition with TRIBE (Trimodal Brain Encoder), a 1B parameter model that combines pretrained representations from Llama 3.2, Wav2Vec2-BERT, and V-JEPA 2 to predict brain responses to movies @AIatMeta
  • ByteDance released Seed LiveInterp 2, a full duplex speech-to-speech model for realtime voice translation that's 3x faster than before with only ~3s lag and >70% correctness @deedydas
  • GLM-4.5V introduced as a breakthrough in open-source visual reasoning, delivering state-of-the-art performance among open-source models with a 106B-parameter MoE architecture @Zai_org
  • NVIDIA unveiled new Nemotron Nano 2 and Llama Nemotron Super 1.5 models for AI agents, plus Cosmos Reason vision language model for physical AI applications at SIGGRAPH 2025 @NVIDIAAI
  • Perplexity launched video generation with audio for Pro and Max subscribers, with Max users getting higher rate limits and enhanced quality @perplexity_ai
  • Claude now supports referencing past chats, allowing users to easily pick up from where they left off @claudeai
  • Google's Gemini Live now connects with Google apps, allowing users to share camera or screen for instant help @GeminiApp
  • Google released Deep Think for Ultra subscribers, showing strong performance in math and coding problems @GeminiApp
  • Ant Group released EchoMimicV3, a new talking head model based on Wan 2.1 1.3B @Xianbao_QIAN

AI Industry Analysis

  • OpenAI's GPT-OSS achieved over 5M downloads in under a week on Hugging Face with 400+ fine-tunes, outpacing DeepSeek R1's launch numbers and becoming the most-liked release of any major LLM this year @reach_vb
  • China's largest tech companies are on pace to spend 1/10th the capex of their American counterparts, potentially benefiting from open-source AI strategy where others pay for GPU costs @natolambert
  • NVIDIA and AMD agreed to give 15% of revenues from H20 and MI308 chip sales in China directly to the US Government as part of export license agreements @AndrewCurran_
  • Reid Hoffman explains OpenAI's strategy of immediately opening GPT-5 to everyone as a blitzscale bet to lock in massive network effects, despite higher serving costs, to reach their goal of 1 billion weekly active users by year's end @reidhoffman
  • Paul Graham notes that the two most impressive companies in the current YC batch are not working on AI, emphasizing that founders matter more than the industry when predicting startup success @paulg
  • Gergely Orosz observes that as AI interview helper tools become more sophisticated, companies will increasingly insist on in-person interviews to distinguish real candidate capabilities @GergelyOrosz
  • Mustafa Suleyman predicts that as AI models become commoditized, value will be added in the orchestration layer, coordinating multiple models to combine strengths rather than routing to just one best model @mustafasuleyman
  • Ethan Mollick suggests that when AI development plateaus, it may actually accelerate AI integration into daily life because it becomes easier to figure out what complementary products and services are needed @emollick

AI Ethics & Society

  • Sam Altman discusses the concerning attachment people develop to specific AI models, noting it feels different and stronger than previous technology attachments, and outlines OpenAI's responsibility in managing user relationships with AI to ensure long-term well-being @sama
  • Geoffrey Hinton warns that major cuts to National Science Foundation funding would be very bad for the future of the US @geoffreyhinton
  • MIT Technology Review reports on early-adopter judges using AI in their courtrooms, raising questions about AI's role in judicial decision-making @techreview

AI Applications

  • FutureHouse, co-founded by MIT alum, developed AI agents to automate scientific research steps including information retrieval, synthesis, chemical synthesis design, and data analysis, aiming to give scientists new tools rather than replace them @medialab
  • Ethan Mollick demonstrates Claude's creative capabilities by having it rewrite The Great Gatsby as "de-carcinized" (removing crab-like defensive behaviors), showing AI's ability to understand and execute complex literary transformations @emollick
  • Eugene Yan successfully teaches Qwen3-8B a new made-up vocabulary using semantic IDs, showing the model becoming bilingual in English and semantic IDs after 3,400 training steps @eugeneyan
  • Simon Willison notes that Qwen3-4B-Thinking became the first model to directly push back against his "pelican riding a bicycle" test, calling it "oddly specific and completely unrealistic" and demonstrating more assertive behavior @simonw

AI Research

  • OpenAI achieved gold medal-level performance at the 2025 International Olympiad in Informatics (IOI), placing 6th among humans and 1st among AIs, using the same IMO gold model without IOI-specific training, demonstrating that reasoning generalizes across domains @SherylHsu02
  • Alexander Wei from OpenAI emphasizes that their IMO gold model set a new state-of-the-art in internal competitive programming evaluations, showing reasoning capabilities generalize across mathematical proofs, competitive programming, and algorithmic problem-solving @alexwei_
  • Noam Brown highlights that OpenAI's IMO gold model being their best competitive coding model demonstrates the generalization of reasoning across creative, fuzzy, and precise reasoning tasks @polynoamial
  • Demis Hassabis discusses Google's plans for Genie 3, including user-generated content sharing and the convergence of Genie, Veo, and Gemini models into an "omnimodel" that can do everything @AndrewCurran_
  • Noam Brown analyzes research showing AI's economic impact may not appear in GDP because most benefits accrue to consumers rather than being captured in market prices, similar to email, Wikipedia, and Google Maps @polynoamial