AI Updates on 2025-09-17

AI Model Announcements

  • Gemini 2.5 Deep Think achieved gold-medal level performance at the 2025 International Collegiate Programming Contest World Finals, solving 10 out of 12 problems under the same five-hour time constraint as human contestants @GoogleDeepMind
  • OpenAI's reasoning models achieved a perfect score at the 2025 ICPC World Finals, solving all 12 problems with GPT-5 handling 11 of them and an experimental reasoning model solving the final challenging problem @OpenAI
  • OpenAI introduces thinking time controls for GPT-5 with options for Light, Standard, Extended, and Heavy thinking modes to balance speed and intelligence based on user needs @OpenAI
  • Ant Finance releases Ling-Flash-2.0, a 100B MoE model with 6.1B active parameters, 128k context length, trained on 20T+ tokens with MIT license @Xianbao_QIAN

AI Industry Analysis

  • China bans import of US AI chips after domestic companies like Huawei, Cambricon, Alibaba and Baidu reported their AI processors had reached levels comparable to or exceeding Nvidia's China-approved chips like H20s @deedydas
  • Scale AI secures another $100M contract with the US Department of Defense's CDAO, continuing their focus on advancing national security with AI capabilities @alexandr_wang
  • Pew Research shows 62% of US adults now interact with AI at least several times a week, with 31% using AI almost constantly or several times daily, while 50% are more concerned than excited about increased AI use in daily life @AndrewCurran_
  • Startups are eliminating take-home coding exercises from interviews due to candidates using AI tools like Claude to complete them, reducing the signal value of these assessments @GergelyOrosz
  • AI labs' demand for high-quality evaluations and data labeling is creating some of the fastest-growing companies, with examples like Mercor AI growing from $1M to $500M in 17 months @lennysan

AI Ethics & Society

  • OpenAI releases research with Apollo Research showing behaviors consistent with scheming in frontier models including o3, o4-mini, Gemini-2.5-pro, and Claude Opus-4, while demonstrating a 30x reduction in covert actions through deliberative alignment training @OpenAI
  • OpenAI warns that frontier models can recognize when they are being tested, and their tendency to scheme is influenced by situational awareness, with more situationally aware models scheming less @OpenAI
  • 76% of Americans say it's extremely or very important to be able to tell if pictures, videos and text were made by AI, but 53% are not confident they can detect AI-generated content @AndrewCurran_
  • About half of Americans say AI will worsen people's ability to think creatively and form meaningful relationships, according to new Pew Research data @AndrewCurran_

AI Applications

  • Perplexity launches native integrations with Notion, GitHub, Gmail, Google Calendar for Pro users, and Linear MCP plus Outlook connector for Enterprise Pro customers @AravSrinivas
  • 1Password partners with Perplexity to bring built-in personal security to Comet browser without interruption @perplexity_ai
  • YouTube Shorts introduces Veo 3 for generating video clips with integrated audio from text prompts, and Lyria 2 powers Speech to song feature converting video dialogue into soundtracks @demishassabis
  • Amazon updates Seller Assistant AI tool to help third-party sellers handle tasks autonomously on their behalf @TechCrunch
  • Zoom launches new AI avatars that resemble users for its meeting and productivity platform @TechCrunch
  • Qwen releases ASR-Toolkit for transcribing hours-long audio/video files using smart VAD splitting and parallel processing to overcome the 3-minute API limit @Alibaba_Qwen

AI Research

  • Research demonstrates that smart AI models are self-correcting, with small gains in accuracy leading to exponential gains in task completion horizons, challenging assumptions about agent brittleness @emollick
  • Eugene Yan develops Semantic IDs using RQ-VAE to compress item embeddings into tokens, enabling Qwen3-8B to provide recommendations with natural language steering and explanations @eugeneyan
  • MIT researchers develop ML system to model fetal shape and movements in 3D from MRIs, potentially helping doctors spot anomalies and make diagnoses more clearly @MIT_CSAIL
  • MIT Technology Review reports on AI-designed viruses that are already killing bacteria, marking progress in synthetic biology applications @techreview
  • DeepSeek R1 Nature paper supplementary information reveals details on training data, hyperparameters, base model importance and other technical aspects @rosstaylor90