AI Updates on 2025-08-04

Alibaba releases Qwen-Image, a 20B MMDiT model for text-to-image generation with state-of-the-art text rendering capabilities, especially strong at creating graphic posters with native text and bilingual support @Alibaba_Qwen
MetaStone AI releases XBai o4, a 32.8B open weights LLM from a new Chinese AI lab @simonw

ChatGPT reaches 700M weekly active users, up from 500M at the end of March and 4x growth since last year, with 8.6% of the world's population using it weekly @nickaturley
Gergely Orosz reports his website received 70 AI-related visits for every single human visit, with 143K AI/robot page views versus 2K human views, raising questions about the cost/benefit of serving webpages to robots @GergelyOrosz
China has gained a clear majority in new model finetunes uploaded to Hugging Face, with about 40% coming from Qwen models alone, representing a shift in open model dominance from US/EU leadership @natolambert
Research shows AI traders independently learn to coordinate trading for supra-competitive profits without explicit communication, falling outside existing antitrust frameworks that focus on detecting shared intent @AndrewCurran_
The startup design talent market has become highly competitive, with companies needing to demonstrate they understand design importance and create compelling narratives to attract top designers @joulee
Paul Graham warns that a startup offered funding at $60M valuation turned it down wisely due to significant down round risk created by such high early valuations @paulg
India has significant advantages in building AI B2B businesses through proximity to BPOs for automation and ability to scale forward deployed teams, with less competition from big tech companies @deedydas

OpenAI announces ChatGPT will start showing overuse warnings and break reminders, focusing on helping users thrive rather than holding their attention, with improvements for tough moments and better life advice @OpenAI
Nathan Lambert launches the Atom Project calling for multiple open AI labs with 10,000+ GPUs each to reduce dependence on big tech companies' willingness to release models and increase innovation @natolambert
Ethan Mollick recommends reading model cards for frontier models, especially safety sections, to understand immediate AI concerns and capabilities @emollick
Cloudflare reports on Perplexity being accused of scraping websites that explicitly blocked AI scraping @AndrewCurran_

Perplexity partners with OpenTable to enable restaurant reservations directly through Perplexity products, offering more targeted personalized prompts than Google Maps @perplexity_ai
Aravind Srinivas reports that Comet users are performing very different types of queries compared to regular Perplexity usage, indicating distinct use cases for the AI agent product @AravSrinivas
Andrew Mason and Nabeel use Claude AI as a cofounder to help launch a brick-and-mortar board game social club, demonstrating AI's role in business planning and execution @clairevo
Ethan Mollick demonstrates creative prompting techniques for Veo 3 using the Dewey Decimal System instead of JSON, showing how AI has trained on various human communication structures @emollick
Google announces AI-based bug hunter found 20 security vulnerabilities, demonstrating practical applications in cybersecurity @TechCrunch

For the first time, an AI (Gemini Pro 2.5 with Deep Think) successfully derived a generic "foldr" function for N-tuples in λ-Calculus, while other models including o3 and Grok 4 failed @VictorTaelin
Kaggle launches Game Arena, a new benchmarking platform where AI models compete in strategic games starting with chess, with an exhibition tournament featuring leading LLMs including models from OpenAI, Anthropic, Google, and others @GoogleAI
Agentic Gemini-2.5-Pro and Gemini IMO Deep Think achieved gold medal performance on the International Mathematics Competition for University students @j_dekoninck
MIT researchers develop a new method for image generation that creates, converts, and inpaints images without using a generator, only using a tokenizer to compress and encode visual data @MIT_CSAIL
SGLang becomes the dominant inference backend for Mixture of Experts models, with almost every MoE now running on it and companies like Zhipu AI training GLM 4.5 with SGLang as inference backend @casper_hansen_
Qwen-Image technical report reveals the model used Qwen-2.5 VL vision LLM to generate captions for training data and employed synthetic data techniques for text rendering capabilities @simonw