AI Updates on 2025-09-15

OpenAI releases GPT-5-Codex, a specialized version of GPT-5 optimized for agentic coding, featuring dynamic thinking time allocation and ability to work independently for over 7 hours on complex tasks @OpenAI
Anthropic publishes the first comprehensive Economic Index analyzing AI usage patterns across US states and countries, showing people delegate complete tasks to Claude 39% of the time, up from 27% eight months ago @AnthropicAI
Holo1.5 achieves state-of-the-art UI localization and QA performance with 3x gains versus Qwen-2.5 VL, now available up to 72B parameters as a strong base for computer-use agents @laurentsifre

Alphabet joins Microsoft, Apple and NVIDIA in the $3 trillion market cap club, reflecting the massive market value being created by AI companies @AndrewCurran_
Perplexity becomes the fastest growing GenAI app on both Android and iOS platforms, demonstrating rapid adoption of AI-powered search tools @AravSrinivas
Companies with custom API chatbots are falling behind as major lab chatbots become more agentic, bringing together many tools in single interfaces with memory and projects @emollick
China investigates Nvidia's 2020 acquisition of Mellanox Technologies as trade tensions between the U.S. and China heat up over AI chip technology @TechCrunch
GPT-5-Codex already represents approximately 40% of Codex traffic and is expected to become the majority by end of day, showing rapid adoption of the new model @sama

Stanford researchers study the dangerous trend of kids using "undress" apps to create deepfake nudes of their peers, highlighting the impact of AI-generated child sexual abuse material @StanfordHAI
AI detection remains a complex policy problem requiring careful balance between false negatives and false positives, with even very good detectors being defeatable @emollick
Research highlights new risks in using LLMs for annotation in research, showing how researchers can "hack" their results through model selection and prompting choices @emollick

Tesla introduces Mūn, a new Grok-powered avatar personality for all Tesla vehicles as part of Elon Musk's plan to have AI avatars in every Tesla @AndrewCurran_
Google Gemini showcases creative applications of Nano Banana image generation, including pose changes with sketches, storyboarding for films, and creating 3D renderings from pencil sketches @GeminiApp
Perplexity partners with AICTE to provide training, resources, and 4 million free Perplexity Pro licenses to Indian engineering students as a preferred research and learning tool @AravSrinivas
DocWrangler, a mixed-initiative IDE for semantic data processing, receives Best Paper Honorable Mention at UIST 2025, addressing challenges in analyzing unstructured documents with AI @sh_reya
Tabracadabra system brings tab-to-autocomplete functionality to any textbox using a General User Model that leverages everything visible on a user's computer for context @oshaikh13

GPT-5-Codex demonstrates dynamic reasoning allocation, being 10x faster for easy queries while thinking 2x longer for complex queries that benefit most from additional compute @polynoamial
Research shows smaller models under 15B parameters benefit most from supervised fine-tuning, while larger 70B+ models perform better with reinforcement learning approaches @natolambert
Study finds that 4 trillion tokens is now considered a small amount of training data in 2025, demonstrating the massive scale requirements for modern AI training @chrmanning
MIT Media Lab's Cynthia Breazeal and alum Sam Rodriques are named to TIME100 AI 2025 list for their contributions to AI research and applications @medialab