AI Updates on 2025-09-05

Alibaba releases Qwen3-Max-Preview with over 1 trillion parameters, claiming stronger performance than their previous Qwen3-235B-A22B-2507 model, now available via Qwen Chat and Alibaba Cloud API @Alibaba_Qwen
OpenAI announces conversation branching feature now live in ChatGPT, allowing users to explore different conversation paths @gdb
Moonshot AI releases Kimi K2-Instruct-0905 with 32B activated parameters out of 1T total, featuring enhanced agentic coding intelligence and 256K context window @AdinaYakup

OpenAI will have their own custom chips for the first time next year, co-designed with Broadcom for internal use only, with Broadcom securing $10 billion in orders from this mystery client @AndrewCurran_
Anthropic reaches $1.5 billion class action settlement with book authors over LibGen and PiLiMi datasets, paying approximately $3,000 per book in what's described as the largest publicly reported copyright recovery in history @AndrewCurran_
Top 3 out of 4 Productivity Apps in the US App Store are AI applications, with 2 from Google, 0.5 from Microsoft, and Perplexity being the only smaller tech company represented @AravSrinivas
OpenAI acquires the team behind Alex Codes, a popular tool for using AI models within Apple's Xcode development suite, in another acqui-hire deal @TechCrunch
Dot, a personalized AI companion, is shutting down after one year of operation, with the team expressing gratitude to users who built close bonds with the AI @jasonyuandesign
Claire Vo reports finally paying herself after nearly 2 years of building ChatPRD, emphasizing the value of building a healthy, bootstrapped business from day one rather than pursuing growth-at-all-costs strategies @clairevo

California and Delaware Attorneys General express concerns to OpenAI about ChatGPT's safety for children and teens, highlighting ongoing regulatory scrutiny of AI systems @TechCrunch
Common Sense Media reports that Google's Gemini falls short on children's safety measures, raising concerns about AI systems' appropriateness for younger users @TechCrunch
Warner Bros. sues Midjourney for generating AI images of Superman, Batman, and other copyrighted characters, highlighting ongoing intellectual property disputes in AI-generated content @TechCrunch

Perplexity launches Finance pages with future estimated revenues for individual American stocks, with Indian stocks support coming next week @AravSrinivas
xAI introduces PDF analysis features in Grok, allowing users to highlight sections and get explanations or ask specific questions about document content @xai
Microsoft partners with Woodland Park Zoo to test SPARROW, an AI system that sends wildlife data directly to the cloud for studying vulnerable Pacific martens @Microsoft
Figma Make becomes available to all higher education and bootcamp education accounts, expanding access to AI-powered design tools @figma
Isotopes launches a sophisticated analytics agent, co-founded by Arun Murthy, one of the creators of Hadoop who later joined Scale AI @TechCrunch
Sierra, a customer service AI agent startup, claims to have hundreds of customers including SoFi, Ramp, and Brex @TechCrunch

OpenAI publishes research explaining why LLMs hallucinate through a connection between supervised and self-supervised learning, describing key obstacles that can be removed to reduce hallucinations @adamfungi
Cameron Wolfe's Deep Learning Focus newsletter reaches 50,000 subscribers, highlighting key technical topics including reasoning models, AI agents, mixture-of-experts architectures, and LLM-as-a-Judge evaluation techniques @cwolferesearch
Hugging Face releases FineVision, described as the best free open dataset to train vision language models, containing 200 training sets condensed into 18B images across 9 subcategories @ClementDelangue
PyTorch explores FlashAttention in 3D through 2-Simplicial Attention, modeling the algorithm with hardware-aligned design and rewriting kernels in TLX (Triton Low Level Extensions) @PyTorch
Arvind Narayanan discusses the "false summit" phenomenon in AI development, where perceived milestones repeatedly prove to be intermediate steps rather than final achievements, leading to accusations that skeptics keep "moving the goalposts" @random_walker