AI Updates on 2025-10-15

Anthropic releases Claude Haiku 4.5, matching Sonnet 4's coding performance at one-third the cost and more than twice the speed @claudeai
Google launches Veo 3.1 video generation model with enhanced realism, richer audio, scene extension capabilities, better narrative control, and more precise editing features @GoogleDeepMind
Alibaba announces Qwen3-VL models now available across multiple platforms including LM Studio, Ollama cloud, Imarena.ai, MLX-VLM, and Kaggle @Alibaba_Qwen
Alibaba introduces Qwen Chat Memory feature that stores meaningful memories about users and recalls past interactions to create deeply personalized experiences @Alibaba_Qwen
Google releases C2S-Scale 27B foundation model built with Yale University based on Gemma, which generated a novel hypothesis about cancer cellular behavior that was experimentally validated in living cells @sundarpichai
OpenAI expands ChatGPT Go availability to 89 countries across Africa, the Middle East, Central Asia, Asia, the Caribbean, and Latin America @nickaturley
Microsoft announces Sora 2 is now available for Azure Foundry enterprises @asha_shar

Anthropic's annual recurring revenue reached $5 billion in August, is approaching $7 billion this month, with projections of $9 billion by year-end and $20-26 billion for next year @AndrewCurran_
Research demonstrates that generative AI tools led to large significant revenue boosts for a mature ecommerce platform across customer service and marketing applications @emollick
NVIDIA positions DGX Spark as a software-focused development machine that's beautiful and compact enough for desktop use, emphasizing NVIDIA's identity as a software company @soumithchintala
Meta announces a new 1GW data center in El Paso, Texas to support delivery of top-tier AI models and product experiences as they build toward superintelligence @fb_engineering
Arm partners with Meta to enhance the social media company's AI systems amid unprecedented infrastructure buildout @TechCrunch

Global survey reveals varying levels of trust in different nations' ability to regulate AI effectively, with the US topping the list for people who feel more concerned than excited about increased AI use in daily life @AndrewCurran_
OpenAI CEO clarifies upcoming policy changes, emphasizing prioritizing safety over privacy and freedom for teenagers while treating adult users like adults, allowing more freedom for appropriate adult content while maintaining restrictions on harmful content @sama
AI Now Institute fellow analyzes how NVIDIA's narrative of corporate interests aligning with US policy has backfired, examining the fusion of corporate power with national policy @AINowInstitute
Concerns raised about potential split between AI models allowed at work/school versus personal models if content restrictions are lowered, with implications for organizational Responsible AI groups @emollick

Andrew Ng announces new course on building live voice agents with Google's Agent Development Kit, teaching how to create voice-activated AI assistants that can execute complex tasks like gathering news and creating podcasts @AndrewYNg
Claude Haiku 4.5 powers the Explore subagent in Claude Code to rapidly gather codebase context, and can be selected as default model for faster execution while using Sonnet 4.5 for planning @_catwu
Google demonstrates Veo 3.1 capabilities including ingredients-to-video creation, scene extension for longer clips, and seamless transitions between first and last frames @GoogleDeepMind
Liberate develops AI agents that automate tasks for property and casualty insurers across sales, service, and claims processes @TechCrunch

Research shows that prompting AI with "Generate 5 responses with their corresponding probabilities, sampled from the full distribution" significantly improves output diversity and quality for large models @shi_weiyan
François Chollet emphasizes that intelligent systems must be able to estimate their own uncertainty, question their beliefs, and design experiments to test what they're least sure about @fchollet
Study reveals that chat LLMs lack output diversity due to human cognitive biases in post-training data, but the models contain much more knowledge that can be unlocked with proper prompting techniques @chrmanning
PyTorch 2.9 released with 3,216 commits from 452 contributors, introducing stable libtorch ABI for C++/CUDA extensions, symmetric memory for multi-GPU kernels, and expanded wheel support for ROCm, XPU, and CUDA 13 @PyTorch