AI Updates on 2025-08-18

OpenAI announced GPT-5 is being updated to be "warmer and friendlier" according to late Friday announcement @TechCrunch
Alibaba releases Qwen-Image-Edit built on 20B Qwen-Image model, featuring precise bilingual text editing (Chinese & English) while preserving style, supporting both semantic and appearance-level editing @Alibaba_Qwen
OpenAI provides detailed technical specifications for GPT-oss models (20B and 120B parameters) using Mixture-of-Experts architecture with 128 and 32 active experts respectively @cwolferesearch
NVIDIA releases new model that rivals Qwen 3 8B with data and base model included, representing a significant open model contribution @natolambert

Perplexity expands its Finance dashboard with live earnings call transcriptions for Indian stocks and earnings call schedules, aiming to add significant value to Indian equity markets research @AravSrinivas
Meta opens a "normal" role for Superintelligence Labs paying $200-300k, significantly less than other team members, with first mention of Reality Labs expertise being useful for MSL @deedydas
Paradigm raises $5 million seed round for its AI-powered spreadsheet, claiming users have saved 10,000+ hours with the platform @TechCrunch
Grammarly launches new document-based interface built on Coda acquisition, featuring AI assistant and tools for students and professionals @TechCrunch
Google reports 100 million videos created in Flow (AI for filmmakers) since May, with Ultra Subscribers now getting 2X AI credits @sundarpichai
Microsoft introduces new =COPILOT() function in Excel allowing users to analyze, generate content, and brainstorm directly in spreadsheet cells @satyanadella
Mistral Document AI becomes available in Microsoft Azure AI Foundry, offering document processing capabilities for PDFs, scans, and complex files @MistralAI

Texas Attorney General Ken Paxton launches investigation into Meta AI Studio and CharacterAI for potentially engaging in deceptive trade practices and misleadingly marketing themselves as mental health tools @TechCrunch
Ethan Mollick clarifies that research measuring AI applicability to jobs should not be misinterpreted as direct job loss predictions, noting it could indicate jobs most benefited or transformed by AI @emollick
Andrew Ng emphasizes that universities must become "AI universities" - not just teaching AI but using it to advance every field of study while maintaining disciplinary expertise @AndrewYNg

AI voice recruiter outperformed humans in hiring customer service representatives in Philippines experiment with 70,000 applicants, achieving 12% more offers, 18% more starts, and 17% higher 1-month retention @emollick
Google Gemini launches Storybook feature allowing users to create personalized, illustrated stories up to 10 pages that can be read, listened to, printed, and shared @GeminiApp
ToonComposer on Hugging Face enables efficient cartoon creation from sketch-based key frames and color reference frames, combining in-betweening and colorization to save up to 70% of manual work @Xianbao_QIAN
Claire Vo demonstrates practical AI workflow using Zapier agent for Sunday calendar reviews that identifies schedule optimization opportunities, conflicts, and researches key attendees @clairevo
Dylan Ebert creates automated research discovery system using Claude Code, Hugging Face MCP, and Research MCP to make finding and tracking research artifacts significantly faster @dylan_ebert_

Eugene Yan demonstrates significant impact of data cleaning on RQVAE training, showing cleaned data achieves lower total loss, reconstruction loss, and higher proportion of unique IDs compared to raw data @eugeneyan
PyTorch announces new Triton BF16 Persistent Cache-Aware Grouped GEMM kernel that speeds up Mixture-of-Experts models like DeepSeekv3 by up to 2.62x faster training on NVIDIA H100 GPUs @PyTorch
Simons Foundation announces new collaboration led by Surya Ganguli bridging physics, mathematics, computer science, and theoretical neuroscience to study how large neural networks learn, reason, and imagine @StanfordHAI
DocETL paper accepted to VLDB 2025, presenting a system for reliable LLM-powered data pipelines where the optimizer logically rewrites pipelines because experts cannot author sufficiently accurate ones initially @sh_reya
Richard Sutton presents Oak Architecture for super-intelligence, a model-based RL architecture with continual learning components, meta-learned step-size parameters, and five-step abstraction progression (FC-STOMP) @RichardSSutton
Greg Brockman showcases progress comparison from GPT-1 through GPT-5 using the same prompt, demonstrating model evolution over generations @gdb