Skip to content

AI Newsroom (Digest)

AI News

A curated roundup of AI’s fast lane - crafted with a model’s help - spotlighting the biggest launches, research milestones, and what’s making waves in the industry.

2025

December 2025

2025-12-02 (Tuesday)

  • Mistral AI releases Mistral 3 model family
    Mistral AI unveiled its next-generation Mistral 3 lineup, including compact dense models in 3B to 14B sizes and the flagship Mistral Large 3 sparse mixture-of-experts powerhouse with 675B total parameters. All variants are open-sourced under Apache 2.0, boasting top-tier multimodal and multilingual prowess for developers and edge devices alike. (mistral.ai)

2025-12-01 (Monday)

  • Internals of DeepSeek Models V3 to V3.2 reviewed by Sebastian Raschka.
    DeepSeek launched its upgraded V3.2 open-weight model, matching top proprietary rivals like GPT-5 in benchmarks. This evolution introduces sparse attention and refined reinforcement learning for enhanced reasoning. (sebastianraschka.com)

  • DeepSeek releases V3.2 flagship model and reasoning variant
    DeepSeek launched its new flagship AI model, V3.2, with an experimental reasoning-enhanced version called V3.2-Speciale. These massive 685-billion-parameter models are openly available under MIT license, excelling in math like gold-medal IMO performance. (simonwillison.net)

  • Hugging Face launches Transformers v5 library
    Hugging Face rolled out Transformers v5 after five years, emphasizing cleaner code, modular designs, and support for over 400 architectures amid daily installs topping 3 million. This update aims to streamline contributions and bolster the AI ecosystem's backbone. (huggingface.co)

November 2025

2025-11-28 (Friday)

  • Moonshot AI introduces Kimi Agentic Slides
    Agentic Slides are powered by Nano Banana Pro and enable Agentic Search (Kimi K2), file upload, fully editable slides canvas and export in Powerpoint format. (x.com)

2025-11-27 (Thursday)

  • The Thinking Game - DeepMind Documentary
    A five-year documentary follows Demis Hassabis and the DeepMind team as they pursue the nature of intelligence, moving from game-playing AI to solving protein folding.
    In 2020 DeepMind already released the AlphaGo Documentary with the legendary Move 37 when competing in Go 2016 against Lee Sedol. (youtube.com)

  • DeepSeek releases open-source Math-V2 model
    DeepSeek unveiled a hefty 685-billion-parameter open-weights model tuned for top-tier math reasoning. It matches proprietary rivals by earning gold-medal equivalents on elite competitions like IMO and Putnam. (huggingface.co)

2025-11-26 (Wednesday)

  • Prime Intellect releases INTELLECT-3 106B MoE model
    Prime Intellect unveiled INTELLECT-3, a 106 billion parameter Mixture-of-Experts model excelling on math, code, and reasoning benchmarks via large-scale reinforcement learning. They open-sourced the full training stack, from weights to environments, empowering broader access to cutting-edge AI development. (primeintellect.ai)

  • AI-generated band Breaking Rust tops Billboard Country Digital Song Sales chart
    An AI-created country band named Breaking Rust has hit number one on Billboard's Country Digital Song Sales chart, showing audiences eagerly consuming machine-made music. This underscores the rising appeal of AI-generated entertainment amid ongoing hype. (technologyreview.com)

2025-11-25 (Tuesday)

  • Black Forest Labs releases FLUX.2 image models optimized for NVIDIA RTX GPUs
    Black Forest Labs unveiled the FLUX.2 series of cutting-edge image generators, boasting photorealistic details, pose control, and multi-reference consistency. NVIDIA teamed up for RTX GPU tweaks via FP8 quantization, trimming VRAM use by 40% to make high-fidelity AI art more accessible on consumer hardware. (blogs.nvidia.com)

  • Google launches Ironwood seventh-gen TPU
    Google unveiled Ironwood, its newest Tensor Processing Unit, as the seventh generation and most potent, efficient chip design so far. It's now ready for use in Google Cloud services. (blog.google)

  • Tom Gally releases SVG generation benchmark for frontier LLMs
    Tom Gally crafted a fun benchmark with 30 whimsical prompts, challenging nine top LLMs to produce SVG artwork like a sloth at the controls of an excavator. The results offer a charming glimpse into which models shine at creative vector graphics. (simonwillison.net)

2025-11-24 (Monday)

  • Anthropic releases Claude Opus 4.5
    New Claude Opus 4.5 by Anthropic - billing it as the top performer for coding, agents, and computer tasks amid rivalry from recent GPT and Gemini updates. Benchmarks edge ahead, yet real-world gains over prior models like Sonnet feel subtle. (simonwillison.net)

2025-11-23 (Sunday)

  • Interconnects.ai lists top US labs releasing open AI models
    A helpful roundup spotlights key US players like Ai2 with Olmo, Nvidia's Nemotron, and others dropping quality open models, amid rising buzz in open AI scenes on both sides of the Pacific. It's a timely snapshot as truly open releases pick up steam. (interconnects.ai)

2025-11-22 (Saturday)

  • Ai2 releases fully open Olmo 3 LLM series
    The Allen Institute for AI unveiled the Olmo 3 family, spotlighting the 32B Olmo 3-Think model as the top fully open option in its class, complete with training data, checkpoints, and tools to peek inside reasoning steps. They also shared smaller 7B variants trained efficiently on a refined 6-trillion-token dataset. (simonwillison.net)

2025-11-20 (Thursday)

  • Google DeepMind launches Nano Banana Pro image model
    Google DeepMind unveiled Nano Banana Pro, a cutting-edge image generation and editing tool powered by Gemini 3 Pro. It brings enhanced control, sharper text handling, and richer world understanding to turn ideas into polished visuals. (blog.google)

  • AI Engineering Code Summit discusses shift to agent swarms in coding
    Industry leaders at the NYC summit outlined a pivot from single large-context AI agents to multi-agent 'ant swarms' for reliable coding, with Steve Yegge forecasting the demise of traditional IDEs by 2026. This emerging approach emphasizes context engineering and verification to sidestep model limitations like the 'Dumb Zone.' (turingpost.com)

  • Allen AI releases OLMo 3 open reasoning models
    Allen AI unveiled OLMo 3, a fresh lineup of fully open 7B and 32B language models topping charts in base performance and reasoning tasks. These releases include all training details, paving the way for community innovation without the usual black-box mysteries. (allenai.org)

  • Hugging Face launches AnyLanguageModel for Apple LLM integration
    Hugging Face unveiled AnyLanguageModel, a Swift package that unifies APIs for local open-source models and cloud providers on Apple devices. It acts as a seamless swap for Apple's Foundation Models, easing developer experimentation with LLMs. (huggingface.co)

2025-11-19 (Wednesday)

  • MIT Technology Review Insights report on AI adoption in manufacturing
    Manufacturers are ramping up AI use alongside digital twins to streamline factories and cut downtime, with half now running AI in production versus just over a third last year. Big firms lead the pack at 77% adoption. (technologyreview.com)

  • OpenAI releases GPT-5.1-Codex-Max for Codex CLI
    OpenAI debuted GPT-5.1-Codex-Max as the default model in its Codex CLI coding agent, boasting top benchmark scores and a novel compaction technique for managing extended coding tasks over vast token contexts. This specialist model shines in agentic workflows, edging out rivals like Gemini 3 Pro. (simonwillison.net)

  • Hollywood backlash against signing AI actress Tilly Norwood
    Talent agencies explored representing AI-generated actress Tilly Norwood, prompting sharp criticism from actors like Emily Blunt and Natasha Lyonne who decried it as frightening and misguided. This flare-up underscores broader Hollywood anxieties about AI displacing human performers. (understandingai.org)

2025-11-18 (Tuesday)

  • Microsoft, NVIDIA, and Anthropic Launch Major AI Infrastructure Partnerships
    Microsoft, NVIDIA, and Anthropic are teaming up to supercharge Anthropic's Claude models on Azure using NVIDIA hardware, with commitments for vast compute resources and billions in joint investments. This move expands enterprise access to advanced AI while optimizing performance across cutting-edge systems. (blogs.nvidia.com)

  • Google releases Gemini 3 AI model
    Google unveiled Gemini 3, its most capable AI model to date, poised to empower users in turning concepts into reality with enhanced intelligence. Leaders like Sundar Pichai and Demis Hassabis highlighted this step forward in AI capabilities. (blog.google)

  • Google launches Search powered by Gemini 3
    Google rolled out its Gemini 3 AI model to enhance Search, delivering sharper reasoning and fresh interactive visuals tailored to user queries. This update promises a more intuitive experience with dynamic tools and simulations. (blog.google)

  • Google releases Antigravity IDE
    Google unveiled Antigravity, a desktop app blending a VS Code-style editor, agent dashboard, and browser tools to streamline app-building with Gemini models. It generates handy Markdown artifacts to track tasks and progress along the way. (simonwillison.net)

2025-11-17 (Monday)

2025-11-13 (Thursday)

2025-11-12 (Wednesday)

  • German court finds OpenAI liable for ChatGPT copyright breach
    A German court decided OpenAI broke national copyright rules by training ChatGPT on protected music without approval, directing the firm to compensate rights manager GEMA. This sets an early European benchmark for AI respecting creators' rights. (techcrunch.com)

  • Waymo enables freeway service for robotaxis
    Waymo now lets its autonomous vehicles use freeways for paying customers in Phoenix, Los Angeles, and the San Francisco Bay Area, a smart upgrade that speeds up trips across vast urban landscapes. This positions the service as a credible challenger to rideshare leaders. (understandingai.org)

  • Moonshot AI releases Kimi K2 Thinking model
    Moonshot AI introduced Kimi K2 Thinking, an enhanced open-source reasoning agent that handles step-by-step problem-solving and chains hundreds of tool calls, edging out top rivals on benchmarks. This builds on their push for agentic intelligence with strong long-context retention. (turingpost.com)

2025-11-11 (Tuesday)

  • Google launches Private AI Compute
    Google unveils Private AI Compute, a clever setup that lets users harness Gemini models' cloud prowess for smart AI help while ensuring their data stays strictly private. It's a thoughtful stride toward blending powerful AI with robust privacy safeguards. (blog.google)

2025-11-10 (Monday)

  • Kimi-Linear attention architecture unveiled
    Kimi-Linear emerges as a hybrid design blending linear recurrence with selective full attention for efficient long-context modeling. It promises to overcome quadratic scaling limits while preserving expressive power through innovations like Kimi Delta Attention. (vizuara.substack.com)

2025-11-06 (Thursday)

  • The “Sutskever Memo” Saga
    According to new testimony, OpenAI cofounder Ilya Sutskever compiled a 52-page memo accusing CEO Sam Altman of repeatedly lying and pitting executives against one another, which eroded the board’s trust and led to his ouster. This saga offers a rare glimpse into the turmoil at OpenAI, one of the world’s most influential AI companies, and highlights how internal chaos and broken trust can spark a dramatic leadership shake-up. (theverge.com)

2025-11-04 (Tuesday)

  • Sebastian Raschka overviews non-transformer LLM architectures
    Sebastian Raschka released a detailed blog post surveying emerging alternatives to dominant autoregressive transformer LLMs. It highlights linear attention hybrids, text diffusion, code world models, and recursive transformers for better efficiency and performance. (sebastianraschka.com)

2025-11-01 (Saturday)

  • Andrej Karpathy releases nanochat tiny language model
    Andrej Karpathy unveiled nanochat, a compact language model that fits on everyday hardware and illustrates full AI training cycles for educational purposes. This open project rapidly attracted over 35,000 GitHub stars, acting as a practical lab for hands-on experimentation. (turingpost.com)

October 2025

2025-10-27 (Monday)

  • Hugging Face launches huggingface_hub v1.0
    After five years of evolution, Hugging Face unveiled version 1.0 of its core Python library, huggingface_hub, boosting performance with modern HTTP tools and a revamped CLI for seamless access to millions of models, datasets, and Spaces. This mature release supports the next era of open machine learning for a vast community. (huggingface.co)

  • Big tech AI capex surges to historic boom levels
    Major tech firms like Amazon, Meta, Microsoft, Alphabet, and Oracle poured $241 billion into capex last year, equaling 0.82% of US GDP and on track to surpass peaks from iconic projects like Apollo. This spending, largely AI-driven, marks one of the biggest investment surges since World War II, with plans for even more ahead. (understandingai.org)

2025-10-06 (Monday)

  • Taylor Swift album promo videos confirmed AI-generated via Google tools
    Taylor Swift's marketing team released QR code videos for her new album that fans suspected were AI-made due to glitches, sparking backlash over hypocrisy. Analysis with Google's SynthID confirmed at least one used Google's video generation tech, signaling AI ads' mainstream push. (understandingai.org)

September 2025

2025-09-04 (Thursday)

  • Google releases EmbeddingGemma embedding model
    Google unveiled EmbeddingGemma, a sleek 308-million-parameter model that tops multilingual embedding benchmarks for its size while fitting neatly on mobile devices with a 2K context window. This open-source gem supports over 100 languages, paving the way for efficient on-device RAG and agents. (huggingface.co)

2025-09-02 (Tuesday)