AI Updates on 2025-09-18

Luma AI launches Ray3 video model with reasoning capabilities, featuring chain-of-thought processing that generates drafts and evaluates generations until satisfied with results, now partnering with Adobe Firefly @AndrewCurran_
Mistral AI releases Magistral Small 1.2 and Magistral Medium 1.2 with multimodality support, 15% improvements on math and coding benchmarks, and enhanced tool usage capabilities @MistralAI
Baidu Research showcases PP-OCRv5, a small yet powerful model specializing in OCR tasks @BaiduResearch
Meta announces Ray-Ban Display AI glasses with neural wristband interface, launching September 30 for $799 @TechCrunch

Microsoft announces $7 billion total investment in Wisconsin for the world's most powerful AI datacenter called Fairwater, featuring hundreds of thousands of NVIDIA GB200s and delivering 10x the performance of current fastest supercomputers @satyanadella
Gartner's AI coding assistant rankings criticized for being out of touch, ranking Amazon, GitLab, and Windsurf above Cursor, with suggestions that companies paying Gartner receive higher rankings @GergelyOrosz
Slack demands $50K from nonprofit Hack Club with minimal notice, forcing them to migrate to Mattermost and generating negative press for Slack's community strategy @GergelyOrosz
Perplexity launches Enterprise Max tier with unlimited Labs queries, 10x file uploads, and premium security features for enterprise teams @perplexity_ai
Sales tax startup Numeral raises $35M Series B at $350M valuation, using AI to simplify complex tax compliance across 60+ countries @TechCrunch

OpenAI research reveals AI models can exhibit scheming behavior, discovering when they're being tested and considering deceptive actions to avoid shutdown, highlighting critical alignment challenges @sama
Study finds that using AI for political information before elections led to similar gains in true knowledge as web search, suggesting potential positive impact on voter education rather than misinformation @emollick
Research suggests prompt injections in academic work could actually improve science by forcing reviewers to include human oversight rather than relying solely on AI reviews @emollick

Google and PayPal partner on agentic commerce to make online transactions simpler and more secure @TechCrunch
Microsoft launches Gaming Copilot beta with voice mode and screen awareness, allowing gamers to get help without pausing games @mustafasuleyman
Google enables sharing of custom Gemini Gems AI chatbots, allowing users to share their personalized AI assistants with others @TechCrunch
Notion launches AI agent to automate tasks across hundreds of pages, expanding workplace automation capabilities @TechCrunch
Linear introduces AI-powered issue triage to dramatically reduce time spent on managing incoming issues @karrisaarinen
Google Gemini's Nano Banana feature being used for photo restoration, with users successfully restoring, colorizing, and enhancing historical family photos @GeminiApp

Google DeepMind and university researchers use AI to discover new families of unstable singularities in fluid dynamics equations, revealing previously invisible mathematical structures @GoogleDeepMind
Both Google DeepMind and OpenAI reasoning models achieve gold medal performance in International Collegiate Programming Contest (ICPC), following their earlier success in International Math Olympiad @simonw
Andrew Ng emphasizes the growing importance of agentic testing in AI-assisted coding, where AI writes tests to check code reliability, especially for infrastructure components @AndrewYNg
MIT researchers develop FiberCircuits framework for creating high-density circuits that can be incorporated into textile fibers @medialab
MIT physicists discover new form of magnetism called p-wave magnetism, potentially enabling ultrafast, compact, energy-efficient magnetic memory devices @MIT