Ongoing AI stories, each with every update in order and who confirmed it.
On Oct 11, attention-galaxy runs SmolLM2-135M in-browser with live next-token and attention visualization.
On Oct 11, AuroraAI-Research released AuroraIMG-6M, a 6.82M-parameter text-to-image model at 64×64.
On Oct 11, Academa Labs open-sources ManimGX, a Manim-compatible Rust/wgpu engine claiming ~90× faster 3D renders.
On Oct 11, Sakana AI published MASS, where one model organizes copies of itself into research teams and trains on what worked.
On Oct 1, Cursor added GLM 5.3 and GLM 5.3 Flash to the editor.
As of Oct 11, QbitAI covers Open Academic Paper Gen, an open-source multi-agent workflow on GitHub that takes a topic through nine stages from literature search on OpenAlex, Crossref, Semantic Scholar and arXiv to a cited draft.
As of Oct 11, QbitAI reports that Kaiming He's team introduced VISTA, a visual-native harness that stores every raw frame an agent sees and lets existing multimodal models recall, crop and compare them during reasoning without extra training.
On Oct 11, Atticus is an open-source hold-to-talk voice assistant for macOS that runs on Claude Code.
As of Oct 11, in an Oct 8 Sources interview recorded at Snapdragon Summit, Qualcomm CEO Cristiano Amon said unnamed frontier AI companies are asking Qualcomm to enable phones to run at least 100-billion-parameter models continuously by 2028 for always-on agentic AI with personal context—more than triple the ~30B MoE on-device ceiling of Qualcomm’s new 2nm Snapdragon Gen 6 platforms.
As of Oct 11, OrcaRouter launched OrcaCyber Zero 1.5, a gated cybersecurity model with a 1M-token context window, per MarkTechPost.
On Oct 10, DigUp is an open-source Mac app for local multimodal file search with EmbeddingGemma 2.
As of Oct 10, QbitAI reports Siemens will present industrial full-stack demos at China International Industry Fair 2026, including a first end-to-end humanoid robot path from design to shop floor, 20-plus new products, and six embodied-intelligence factory scenes spanning flexible load/unload, packaging and palletizing.
On Oct 10, Recraft published a guide to building character reference sheets from one image in Recraft Studio.
On Oct 10, Midjourney plans to test a Midjourney MCP with a limited community.
On Oct 10, webAI released TwIL-LM3-Pro, a 3.66B formal-logic model built on Granite 4.2 3B.
On Oct 10, Bindwell built an 8,000 sq ft AI pesticide lab in 30 days and can validate a model checkpoint in about 72 hours.
On Oct 10, vegaml open-sources Vega, an 800M/4B physics-based typed decision model with 73k context and images.
On Oct 10, Talorys is a self-hosted personal AI agent that deployed on Cloudflare's free tier.
On Oct 5, a16z Top 100 consumer AI, 7th edition: ChatGPT leads Gemini 2x on web visits.
On Oct 9, H Company shows Holo4 building brick models and walkable worlds with HoloBricks and HoloBlocks.
Items older than 72 hours never go in Top stories and show their original date.