AI Updates on 2025-10-13

Alibaba's Qwen3-VL-235B-A22B-Instruct achieves #1 position on OpenRouter for image processing with 48% market share @Alibaba_Qwen
Microsoft releases MAI-Image-1 model, ranking #9 on LMArena and striking a balance between generation speed and quality @mustafasuleyman
Google announces Gemini 2.5 Native Audio Thinking as the new leading Speech to Speech model, achieving 92% on Big Bench Audio benchmark and setting new state-of-the-art for native speech reasoning @sundarpichai
Google rolls out upgraded Video Overviews for NotebookLM with new visuals powered by Nano Banana image generation model and introduces "Brief" format for quick summaries @demishassabis

OpenAI announces collaboration with Broadcom for 10 gigawatts of custom accelerators designed by OpenAI, with Broadcom developing them after 18 months of joint work @AndrewCurran_
JPMorgan announces $10 billion direct equity and venture capital investments into US companies deemed critical to national security, citing concerns about reliance on unreliable sources of critical minerals and manufacturing @AndrewCurran_
Google announces $9+ billion investment in South Carolina through 2027 as part of continued investment in American AI innovation @sundarpichai
Grok's new Imagine version 0.9 represents a significant upgrade, with xAI's rapid development pace indicating the AI video app war is arriving sooner than expected @AndrewCurran_
Sora-level models will likely compete through exclusives and less censorship, with companies like Disney potentially granting cameo rights for character appearances in user-generated videos @AndrewCurran_
Developers who have built production software and have no AI lab affiliations are increasingly reporting that AI tools greatly help their own work, representing a significant shift in expert opinion @GergelyOrosz

Deloitte was held accountable in Australia for submitting work riddled with false AI citations, highlighting the need for accountability in AI-generated content @TechCrunch
California's SB 243 is designed to protect children and vulnerable users from harms associated with AI companion chatbots @TechCrunch
A censorship arms race is expected among AI video models, with unrestricted Sora-level models representing a significant milestone toward media singularity @AndrewCurran_
Theory of Mind for AI appears to be a skill independent of professional expertise, creating understanding gaps between experts who benefit from AI and those who don't @emollick

Microsoft showcases M365 Copilot partner integrations including ServiceNow for autonomous cross-functional processes, Snowflake for natural language data queries, and LexisNexis for legal document drafting @satyanadella
Microsoft launches Copilot Study and Learn Mode that adapts to learning preferences, provides guided assistance without giving away answers, and generates quizzes from uploaded content @Copilot
Salesforce announces upgraded Agentforce platform designed to help enterprises build and deploy AI agents @TechCrunch
MIT PhD student develops computer vision algorithms including "CODA" to help monitor vulnerable ecosystems and support wildlife conservation efforts @MIT_CSAIL
Anduril Industries unveils "EagleEye" helmeted computing system designed to turn soldiers into AI-augmented warfighters @TechCrunch
Stanford scholars are generating synthetic MRIs that could simulate neurological futures based on current habits, making brain aging predictions increasingly plausible @StanfordHAI

Andrej Karpathy releases nanochat, a minimal 8,000-line codebase for training ChatGPT clones from scratch, demonstrating that a functional LLM can be trained for as little as $100 in 4 hours on cloud GPUs @karpathy
Columbia CS Professor Vishal Misra argues that LLMs cannot discover new science because they compress the world into Bayesian manifolds and hallucinate when reasoning outside training data, with true AGI requiring the ability to create entirely new manifolds @a16z
Anthropic's Jack Clark maintains that current AI systems will continue to improve using existing architecture with no diminishing returns, bringing transformative change closer @AndrewCurran_
Research suggests AI water usage for all US data centers ranges from 50M gallons daily for cooling alone to 628M gallons including dam evaporation, significantly less than golf course usage @emollick
New LFM2 Japanese PII extractor with only 350M parameters achieves performance on par with GPT-5 in quality while being extremely fast @huggingface