AI Updates on 2025-10-14

Alibaba releases compact versions of Qwen3-VL in 4B and 8B sizes with both Instruct and Thinking variants, offering lower VRAM usage while retaining full capabilities and outperforming models like Gemini 2.5 Flash Lite and GPT-5 Nano @Alibaba_Qwen
NVIDIA announces DGX Spark, the world's smallest AI supercomputer built on Grace Blackwell architecture, integrating GPUs, CPUs, networking, CUDA libraries and NVIDIA AI software for agentic and physical AI development @nvidianewsroom

OpenAI announces purchase of 10 gigawatts worth of AI accelerator hardware from Broadcom, indicating massive infrastructure investment @TechCrunch
Walmart partners with OpenAI to enable direct product purchases through ChatGPT, allowing users to link accounts, browse items, and checkout within the chatbot @TechCrunch
Anthropic expands partnership with Salesforce, making Claude a preferred model in Agentforce for regulated industries and deepening integration with Slack @AnthropicAI
Perplexity becomes the number one app in Play Store in India across all categories and is now a default search option for Firefox users @AravSrinivas
Reducto raises $75M Series B led by a16z, processing over 1 billion pages and growing monthly volume 6x in just five months after Series A @aditabrm
Google announces first-ever Google AI hub in Visakhapatnam, India, combining gigawatt-scale compute capacity, international subsea gateway, and large-scale energy infrastructure @sundarpichai
Paradigm shift observed in AI from generalist LLM APIs toward companies training and running their own specialized models built on open source, with 1M new repos on Hugging Face in past 90 days @ClementDelangue
AI task length for autonomous agents doubles every few months according to METR evaluations, currently at 2 hours with potential for 2 days next year and 2 weeks in 2 years @a16z

AI Now Institute criticizes OpenAI's easily tricked guardrails, emphasizing need for robust pre-deployment testing before AI models cause substantial harm @AINowInstitute
OpenAI announces plans to relax ChatGPT restrictions in coming weeks, allowing more human-like personality and emoji use, with adult content for verified users coming in December as part of "treat adult users like adults" principle @sama
Anthropic shares initial policy proposals from economists and researchers exploring potential economic effects of powerful AI and policy responses @AnthropicAI
OpenAI establishes Expert Council on Well-Being and AI with eight members including mental health and technology experts to guide responsible AI development @OpenAI

Microsoft introduces Formula Completions in Excel where Copilot proactively suggests formulas based on sheet context when users type "=" @satyanadella
Microsoft integrates Copilot Vision into Moto devices through Moto AI experience, allowing users to show problems rather than just describe them @Copilot
Google demonstrates AI chip design capabilities through AlphaChip, envisioning future where AI methods automate entire chip design process and dramatically accelerate design cycles @AndrewCurran_
Gemini app showcases creative workflow combining Nano Banana for custom pet illustrations, Storybook for narrative creation, and Veo 3 for video animation @GeminiApp
Claude app demonstrates superior performance as personal assistant, particularly with Gmail and Google Calendar integration compared to other AI models @emollick
Developer reports merging 55 Devin PRs and 896 Cursor chats resulting in 16 merged PRs with zero downtime, demonstrating production-ready AI coding capabilities @clairevo
Coco Robotics works toward automating delivery robot fleet using millions of miles of collected data for autonomous navigation @TechCrunch

Karpathy releases nanochat, enabling LLM training in just a few lines of code, representing simplified approach to model development @simonw
Stanford researchers develop SuperDec, extremely compact 3D scene representation replacing millions of Gaussians with just hundreds of primitives, ideal for abstract reasoning and planning in 3D @FrancisEngelman
MIT physicists improve atomic clock precision by reducing quantum noise that obscures atomic "ticking," with applications for online transactions and GPS @MIT
Microsoft Research develops red-teaming protocol for testing and securing DNA biosecurity screening tools, addressing AI safety in biological applications @MSFTResearch
Stanford HAI researchers present projects including world model of human brain for personalized medicine, AI analysis of police body camera footage for transparency, and digital cell twins for drug response simulation @StanfordHAI