AI Updates on 2025-11-06

AI Model Announcements

Alibaba releases Qwen3-max-preview ranking #4 globally on Arena Expert, while Qwen3-235B-A22B-Thinking-2507 ranks #1 among all open-source models on expert-level prompts across 8 critical domains @Alibaba_Qwen
Moonshot AI launches Kimi K2 Thinking, an open-source thinking agent model achieving SOTA on HLE (44.9%) and BrowseComp (60.2%), capable of executing 200-300 sequential tool calls without human interference, with 256K context window @Kimi_Moonshot
Google announces TPU Ironwood (7th generation) coming to general availability with 10X peak performance improvement vs. TPU v5p and more than 4X better performance per chip for both training and inference workloads vs. TPU v6e (Trillium) @sundarpichai
Google introduces File Search Tool in the Gemini API, a hosted RAG solution with free storage and free query time embeddings to simplify context-aware AI systems @OfficialLoganK
Google's Gemini Deep Research now connects directly to Gmail, Drive, Docs, and Chat for all users on desktop, enabling market analysis and competitor reports combining live web trends with internal documents @GeminiApp
OpenAI introduces ability to interrupt long-running queries and add new context without restarting or losing progress, especially useful for refining Deep Research or o1 Pro queries @OpenAI
Perplexity announces major upgrades to Comet Assistant with 23% performance improvement, handling more complex multi-site workflows while working across multiple tabs in parallel @perplexity_ai
Inception Labs raises $50M seed round for Mercury model, achieving 10x faster and 10x cheaper AI coding with performance matching Gemini Flash/Haiku, implementing games like Connect 4 in approximately 2 seconds using novel diffusion models for code @deedydas
Microsoft Research releases Agentic Mode in Data Formulator on Azure AI Foundry Labs, enabling users to update charts, get recommendations, and create reports grounded in data exploration @MSFTResearch
Google DeepMind launches Lyria RealTime API on Google AI Studio for developers to build apps for interactive instrumental music creation and performance, demonstrated through Space DJ web app @GoogleDeepMind

AI Industry Analysis

Andrew Ng warns that SaaS vendors are creating data silos and charging high fees (over $20,000 for API keys) to prevent customers from accessing their own data for AI agent workflows, advising businesses to control their own data to maximize AI capabilities @AndrewYNg
Perplexity announces partnership with Snapchat where Perplexity will be the default AI for all Snapchat users starting January 2026, with Snap paying $400M for the integration @perplexity_ai
Apple is paying $1B to Google to use a whitelabeled Gemini to power Siri, demonstrating the value of platform visibility and distribution @GergelyOrosz
Figma crosses $1B annual revenue run rate with 38% year-over-year revenue growth, with AI investments like Figma Make and MCP delivering results @zoink
AI Studio reaches 2.1 million users vibe coding with hundreds of thousands of apps made every day @OfficialLoganK
Jamie Dimon urges people to embrace AI at America Business Forum, predicting a 3.5 day workweek @AndrewCurran_
Startup survival statistics show 40% die after seed, 50% of remainder die after Series A, 60% after Series B, and 58% after Series C, with roughly 2.5% acquired and 0.5-1% going IPO based on 2016-2018 vintage over 10-year horizon @deedydas
Soumith Chintala announces departure from Meta and PyTorch after 11 years, stepping down from leading PyTorch which achieved 90%+ adoption in AI and powers foundation models at virtually every major AI company @soumithchintala
Sam Altman clarifies OpenAI does not want government guarantees for datacenters, expects to end year above $20B in annualized revenue and grow to hundreds of billions by 2030, with $1.4 trillion in infrastructure commitments over next 8 years @sama

AI Ethics & Society

OpenAI states they treat risks of superintelligent systems as potentially catastrophic and believe empirically studying safety and alignment can help global decisions, including whether the field should slow development to study systems capable of recursive self-improvement @AndrewCurran_
Microsoft AI announces formation of Superintelligence Team focused on Humanist Superintelligence (HSI), defined as incredibly advanced AI capabilities that always work for and in service of people and humanity, emphasizing domain-specific systems that are carefully calibrated and contextualized within limits @mustafasuleyman
Mustafa Suleyman emphasizes Microsoft AI is not building an ill-defined and ethereal superintelligence but a practical technology explicitly designed only to serve humanity, stating he doesn't want to live in a world where AI transcends humanity @mustafasuleyman
Research shows advanced AI models shift their beliefs as they encounter new information and have interactions with people, with active persuasion working but effects coming from overall context, raising alignment issues and showing why SEO for agents is not simple @emollick
Ethan Mollick questions what winning the international AI race means, noting policymakers do not seem to believe in a takeoff scenario based on other decisions, and without an apotheosis as a finish line, it isn't clear what we are racing to @emollick

AI Applications

Andrew Ng reports AI agents are getting better at looking at different types of data in businesses to spot patterns and create value, making data silos increasingly painful, with the value of connecting the dots between different pieces of data higher than ever @AndrewYNg
Hamel Husain demonstrates AI coding hack using Amp's librarian feature to investigate code and dependencies with specific goals, keeping threads dangling and forking them for better context @HamelHusain
Simon Willison shares process for using coding agents for code research tasks with dedicated research GitHub repo where agents run detailed experiments and write up results, with README automatically updated by LLM to include summaries @simonw
Linear becomes the intake tool from which work or feedback gets coordinated further to humans and to agents @karrisaarinen
BillionToOne goes public with genetic test now helping screen 1 in 11 US babies, unlocking earlier detection from prenatal care to cancer @ycombinator
MIT Media Lab develops tiny nanoelectronic devices called circulatronics that autonomously recognize and target diseased regions in the brain and self-implant to provide precise brain stimulation, potentially making therapeutic brain implants accessible without surgery @medialab

AI Research

Microsoft Research announces PIKE-RAG collaboration with Signify showing 12% increase in accuracy for enterprise knowledge systems, delivering faster and more reliable answers @MSFTResearch
vLLM now fully supports hybrid models like Qwen3-Next, Nemotron Nano 2, and Granite 4.0, elevating them from experimental hacks in V0 to first-class citizens in V1 @PyTorch
KernelFalcon achieves 100% correctness across all 250 KernelBench L1-L3 tasks through deep agent architecture combining hierarchical task decomposition, deterministic orchestration, grounded execution, and parallel verification to generate GPU kernels @PyTorch
Research on AlphaEvolve for mathematical exploration at scale tested on 67 problems, documenting all successes and failures in collaboration between MIT, Wellesley, Harvard, and Google DeepMind @GoogleDeepMind
Study shows LLMs have dominated recent work on simulating human behaviors, but lightweight graph neural networks (GNN) can match or beat strong LLM-based methods in discrete-choice settings @berkeley_ai
New paper introduces WIMHF (What's In My Human Feedback) using SAEs to automatically extract signals from preference data to forecast unexpected/harmful changes to LLMs like overconfidence or sycophancy ahead of time @berkeley_ai
Research demonstrates that any task frontier AI can sort of do today will likely be able to do reliably one year from now @gdb