Kimi K3, Gemini Notebook, Codex Micro, Anthropic

<!doctype html> Kimi K3, Gemini Notebook, Codex Micro, Anthropic Kimi K3, Gemini Notebook, Codex Micro, Anthropic

Открытые веса, агентные интерфейсы, ответные движки, маршрутизация моделей, безопасность и роботы: день, когда всё снова стало вопросом эксплуатации. Ужасно практично.

• Kimi K3 2.8T open MoE — new angle: Moonshot AI launches Kimi K3, a 2.8T-parameter open MoE with 1M context and strong coding benchmarks, turning open-weight competition into a scale-and-serving arms race
• Germany media law targets AI answers — new angle: German regulators classify Google AI Overviews and Perplexity as media-law content rather than neutral search plumbing, making answer engines editorial actors with obligations
• Google renames NotebookLM and adds cloud computers — new angle: Google folds NotebookLM into Gemini Notebook and gives notebooks cloud computers while Search opens connected apps, turning research notes into executable workspace agents
• Codex Micro agent hardware controller — new angle: OpenAI and Work Louder unveil Codex Micro, a physical joystick-style controller for AI coding agents, suggesting agent UX is becoming cockpit hardware instead of command lines
• Sakana Fugu adds Nemotron orchestration — new angle: Sakana AI adds Nvidia Nemotron to Fugu to argue that orchestrating open models can rival single frontier systems, moving competition from model weights to routing and composition
• Anthropic EU testimony backlash — new angle: EU officials criticized Anthropic for sending a junior remote witness on advanced AI safety, showing governance friction now includes institutional respect as well as technical risk
• Anthropic self-improvement warning — new angle: Anthropic warns frontier systems may soon accelerate AI R&D enough to need brake-pedal evals and deployment pauses, shifting safety debate from misuse to recursive capability feedback
• Linus says Linux is not anti-AI — new angle: Linus Torvalds says Linux will not ban AI-assisted development and will judge patches on technical merit, making open-source governance choose evidence over tool ideology
• Firefox running inside WebAssembly — new angle: Puter compiled Firefox to WebAssembly so a full browser can run inside another browser, a stunt that also shows AI-assisted engineering can compress absurd porting projects into subscription economics
• SearchOS multi-agent search loops — new angle: SearchOS proposes a system layer for collaborative search agents that prevents repetitive dead loops as histories grow, treating agent search as operations management rather than prompt charm
• RoboTTT scales robot context — new angle: RoboTTT scales robot policy context to 8K timesteps without extra inference latency, pointing toward robots that learn from demonstrations and perturbations in longer operational memory
• BadWAM warns world-action models can dream wrong — new angle: BadWAM shows world-action models can imagine plausible futures while selecting unsafe actions, undercutting the idea that predictive world models automatically make robotics safer
• Soofi S German open model — new angle: a German consortium releases Soofi S, a 30B-A3B open MoE claiming strong English and German results, showing national open-model efforts moving from policy speeches to training logs
• ExLlamaV3 local inference upgrades — new angle: ExLlamaV3 1.0 reports major decode speedups for quantized local models, reminding everyone that open weights only become useful when the miserable memory bandwidth accounting works
• NVIDIA Nemotron 3 Embed tops RTEB — new angle: NVIDIA says Nemotron 3 Embed ranks first on RTEB, making retrieval quality a first-class component of agent stacks rather than a boring prelude nobody thanks