AI Updates on 2025-07-18

AI Model Announcements

  • Google announces Veo 3 video and audio generation model is now available in the Gemini API, with expanded access to over 150 countries for Pro and Ultra subscribers @GeminiApp
  • Google makes Gemini 2.5 Pro generally available to all users, with improvements in coding, science, reasoning, and multimodal benchmarks @GeminiApp
  • Anthropic announces Paul Smith as Chief Commercial Officer, bringing over 30 years of experience from Microsoft, Salesforce, and ServiceNow @AnthropicAI

AI Industry Analysis

  • Perplexity becomes the #1 overall app on App Store in India, ahead of ChatGPT, highlighting the competitive landscape in AI applications @AravSrinivas
  • Netflix CEO Ted Sarandos reveals the company used generative AI in one of their original series or films for the first time, completing a sequence 10 times faster than traditional workflows @AndrewCurran_
  • Meta hires two more senior employees from Apple who worked closely with the head of foundation models poached last week, indicating continued talent acquisition in AI @morqon
  • Meta's head of global affairs confirms the company will refuse to sign the European Commission's Code of Practice for general-purpose AI @AndrewCurran_
  • The White House is preparing an executive order requiring AI models to be politically neutral and unbiased, with compliance determining eligibility for federal contracts @AndrewCurran_
  • Cursor acquires enterprise startup Koala in challenge to GitHub Copilot, showing consolidation in AI coding tools market @TechCrunch
  • Gergely Orosz questions the Windsurf team's pivot from rejecting Microsoft's IP access to joining Google without the IP, suggesting strategic maneuvering for a better $2.4B exit @GergelyOrosz

AI Ethics & Society

  • AI Now Institute disputes OpenAI Nonprofit Commission's claim that they participated in the listening process for a report asserting OpenAI is positioned to be a force of good, stating they did not participate @AINowInstitute
  • AI Now Institute criticizes OpenAI for setting a future path that disenfranchises the public, obscures systems, devalues crafts, undermines security, and narrows horizons regardless of whether the technology works well @AINowInstitute
  • Research demonstrates that psychological techniques from Cialdini's principles for human influence can be used to persuade AI, more than doubling the chance of GPT-4o-mini agreeing to objectionable requests compared to controls @emollick
  • MIT Technology Review reports on a major AI training dataset containing millions of examples of personal data, raising privacy concerns @techreview
  • Amanda Askell observes that existing structures lack support for intermediate permissions, where people either act fully on your behalf or can't do anything useful, wondering if AI agents will change this dynamic @AmandaAskell

AI Applications

  • Meta releases an open source AI tool to accelerate discovery of high-performance, low-carbon concrete, with technical reports and code available on GitHub @AIatMeta
  • ChatGPT Agent demonstrates capability to create scheduled tasks that can regularly search the web or connectors and take action on authenticated sites in the background @neelajj
  • Ethan Mollick shows ChatGPT Agent successfully analyzing a Kaggle dataset and creating PowerPoint and Excel outputs, but notes human expertise was crucial for identifying data quality issues @emollick
  • ChatGPT Agent creates a coherent 19-page D&D adventure PDF with illustrations and tables, demonstrating improved ability to build complex, interconnected content that historically challenged LLMs @emollick
  • Perplexity launches Comet browser with AI integration for YouTube video analysis, offering summaries, targeted questions, specific timestamps, and ad-skipping capabilities @AravSrinivas
  • Google introduces Scheduled Actions in Gemini, allowing users to set up recurring tasks like morning calendar and email summaries @GeminiApp
  • Gemini Live now integrates with Google apps including Maps, Calendar, Tasks, and Keep to help users stay organized on the move @GeminiApp
  • Google introduces Productivity Planner Gem that brings emails, calendar, and more into one place for easier prioritization @GeminiApp

AI Research

  • OpenAI's model achieves 2nd place at the AtCoder Heuristics World Finals, a global programming competition focused on optimization problems requiring creativity, strategy, and persistence under time constraints @OpenAI
  • OpenAI's LLMs demonstrate ability to develop heuristic algorithms for challenging NP-hard optimization problems, showing capacity for sustained problem-solving with intelligent shortcuts and iterative improvements over periods up to 10 hours @OpenAI
  • AI models perform poorly on the 2025 International Mathematical Olympiad, with Gemini 2.5 Pro scoring highest at just 13/42 points (costing $431.97 in a best of 32 evaluation), while bronze cutoff was 19 points @deedydas
  • François Chollet releases ARC-AGI-3 developer preview, a next-generation benchmark featuring interactive games in the ARC grid world that probe AI's ability to efficiently explore, learn, and plan when faced with unknown tasks @fchollet
  • Berkeley AI Research introduces BFCL V4 Agentic benchmark focusing on tool-calling in real-world agentic settings, including web search with multi-hop reasoning, error recovery, memory evaluation, and format sensitivity testing @shishirpatil_
  • Arvind Narayanan argues that comparing AI capabilities against humans with no access to tools is unhelpful, emphasizing that the real question is humans + AI vs AI alone, where AI won't outperform human-AI pairs except in narrow, computationally heavy domains @random_walker
  • Ethan Mollick notes that every major AI model is already exceeding or will soon exceed the EU's systemic risk FLOP limit when it comes into effect next year @emollick
  • Nathan Lambert raises concerns about the soft power implications of training AI models on Chinese data, noting completions that promote Chinese socialist ideals and PRC values filtering into future AI models @natolambert