AI Updates on 2025-07-24

AI Model Announcements

  • Alibaba releases Qwen3-Coder-480B-A35B, a 480B parameter MoE model with 35B active parameters achieving 70% on SWE-Bench Verified and 1M context length, potentially the best coding model yet @deedydas
  • Alibaba launches Qwen3-MT, their most powerful translation model supporting 92+ languages and covering 95%+ of the world's population, trained on trillions of multilingual tokens @Alibaba_Qwen
  • Tom Warren reports GPT-5 will launch in August with GPT-5-mini launching simultaneously in both client and API, and GPT-5-nano planned for API only @AndrewCurran_
  • OpenAI plans to launch an open source model before GPT-5, described as similar to o3-mini with reasoning capabilities @AndrewCurran_

AI Industry Analysis

  • Google processes over 980 trillion tokens monthly across their surfaces, doubling from 480 trillion in May, with Gemini app reaching 450M monthly active users @AndrewCurran_
  • Over 70 million user videos have been created with Veo 3, demonstrating significant adoption of Google's video generation model @AndrewCurran_
  • Safe Superintelligence (Ilya Sutskever's company) will exclusively use Google TPUs for their AI development @AndrewCurran_
  • Meta adopts novel approach of building weather-proof tents to house GPU clusters, enabling new data centers to come online in months instead of years @AIatMeta
  • Financial Times reports over $1 billion worth of NVIDIA chips have reached China in the last three months, including Blackwell chips, despite export controls @AndrewCurran_
  • China now has 5 frontier AI labs competing globally: DeepSeek, Alibaba Qwen, Bytedance, Hailuo, and Kimi, with rapid development pace at likely lower costs than US counterparts @deedydas
  • Research shows developers save the most time with AI tools through stack trace analysis and refactoring rather than code generation, based on DX research with 180 companies @GergelyOrosz
  • Forward-looking tech companies like GitHub and Shopify are hiring more interns because of AI, observing CS students use AI tools more fluently than before @GergelyOrosz
  • Jack Dorsey releases two apps in less than a week using AI tool Goose for rapid development, demonstrating the "vibe coding" trend @TechCrunch

AI Ethics & Society

  • President Trump's AI Summit comments on copyright suggest AI should be able to learn from content without paying for each use, comparing it to human learning and noting China doesn't follow such restrictions @AndrewCurran_
  • New government requirements state that to be eligible for agency contracts, an LLM must be developed with truth-seeking and ideological neutrality principles @AndrewCurran_
  • Ethan Mollick demonstrates that over 60% of older links from New York Times articles are now broken, suggesting only LLMs will "remember" much of the web's ephemeral content @emollick
  • Careful review of Humanity's Last Exam benchmark reveals many questions have incorrect "right" answers, highlighting ongoing challenges in AI measurement and benchmarking @emollick
  • François Chollet warns against the tendency to anthropomorphize AI systems that are not human, emphasizing the importance of understanding their true nature @fchollet

AI Applications

  • Perplexity launches Comet browser with AI assistant capabilities that can distribute itself and onboard new users, receiving positive reviews for its functionality @testingcatalog
  • Cursor releases Bugbot, which found over 1M+ bugs in human-written PRs in the past month, with over half being real logic issues that were fixed before merging @cursor_ai
  • GitHub launches Spark, a prompt-to-app platform for creating and iterating on React apps with user authentication and persistent storage @simonw
  • Figma releases Make to everyone, a prompt-to-app solution that allows users to create prototypes and publish to Figma Community @figma
  • Google introduces photo to video feature coming to Google Photos and YouTube Shorts @sundarpichai
  • Google launches virtual try-on clothes feature using AI technology @TechCrunch
  • Linear introduces Dashboards feature allowing users to create custom views to monitor key metrics @linear
  • xAI partners with Kalshi to bring Grok to prediction markets @xai

AI Research

  • Anthropic develops three AI agents for alignment auditing that can autonomously uncover hidden goals, build safety evaluations, and surface concerning behaviors, with their investigator agent winning 42% of auditing challenges @AnthropicAI
  • Google achieves gold-medal level performance in the International Mathematical Olympiad using an advanced version of Gemini with Deep Think mode @sundarpichai
  • Research introduces Rubrics as Rewards (RaR) framework using structured, checklist-style rubrics as interpretable reward signals for on-policy training, yielding relative improvements on HealthBench-1k @iScienceLuvr
  • Cameron Wolfe explains that reward models remain relevant in the age of reasoning models, as most systems still use both RLHF for human preference alignment and RLVR for verifiable reasoning tasks @cwolferesearch
  • Anthropic launches "AI psychiatry" team as part of interpretability efforts to research model personas, motivations, and situational awareness and how they lead to concerning behaviors @Jack_W_Lindsey
  • MIT scientists program living cells with logic gates like biological computers to detect and destroy cancer with precision @MIT
  • PyTorch demonstrates SmolLM3-3B running at 15 tokens/sec on Galaxy S22 using TorchAO and ExecuTorch for on-device deployment @PyTorch