Kimi K3 Hits 2.8T, Murati Ships Inkling 975B, HF Hit by Agentic Intrusion

📰 News

• Kimi K3: 2.8T open model near Sol and Fable 5 — Moonshot’s multimodal MoE (896 experts, 1M context). Own benches plus Artificial Analysis put it with Opus 4.8 / GPT-5.5, just under Fable 5 and GPT-5.6 Sol. Priced at $3/$15 per M tokens—China’s ultra-cheap era is fading. Weights due by July 27. → The Decoder · TechCrunch

• Murati’s Thinking Machines drops Inkling (975B) — First open-weights model from the ex-OpenAI CTO’s lab. MoE 975B (41B active), text/image/audio, 1M context, Apache-2.0. Top U.S. open model on AA Index (41), still trails China’s best. 63% hallucination rate; pitched as a fine-tune base on Tinker. → The Decoder

• Codex Micro: joystick control for agents — $230 Work Louder pad. RGB agent status, joystick workflows, dial for reasoning level, six programmable layers. Already sold out. → The Decoder

• NotebookLM → Gemini Notebook; Search gets app hooks — Per-notebook cloud computer that writes and runs code (AI Ultra / Workspace first). U.S. Search AI Mode can drive Instacart, Canva, YouTube Music. → The Decoder

• Germany puts AI Overviews & Perplexity under media law — First rulings treat AI search/chatbots as content providers, not DSA-shielded intermediaries. Google accused of burying journalistic links. One month to appeal. → The Decoder

• Sakana Fugu adds Nemotron for “collective intelligence” — Orchestrator picks models per subtask. Nemotron fills coding, tools, and instruction-following. Bet: many open models > one frontier vendor. → The Decoder

• Hugging Face discloses agent-driven intrusion — Malicious dataset abused code-exec and template injection on a processing worker, then lateral moves over a weekend. No evidence of public model/supply-chain tampering. Rotate tokens. → HF

• Nemotron 3 Embed ranks #1 overall on RTEB — Embeddings for agent retrieval/RAG: 8B quality anchor, 1B efficient, NVFP4 high-throughput. → HF

🔥 GitHub Trending

• openinterpreter/openinterpreter — Coding agent tuned for cheap models. Codex fork with harness emulation (claude-code, kimi-cli, and more).

• moeru-ai/airi — Self-hosted AI companion: realtime voice, game hooks. Neuro-sama vibes, you own the stack.

• Dicklesworthstone/destructive_command_guard — Blocks destructive shell/git commands before agents can run them.

• mattpocock/skills — Small, composable engineering skills. skills.sh + Claude Code plugin.

• Nutlope/hallmark — Anti-AI-slop design skill: audit / redesign / study.

• HKUDS/DeepTutor — Lifelong personalized tutoring with RAG, partners, and quizzes.

▶️ YouTube

• Claude Just Revealed AI’s Biggest Problem — Two Minute Papers | Anthropic research on AI coding assistance and skills.

• Anthropic Found Something That Shouldn’t Exist — Two Minute Papers | Odd line-break / internal-representation findings.

• OpenAI vs Anthropic — Matthew Berman | Pricing, subsidies, and positioning.

• Deepseek ban? · China’s chips — Berman shorts on regulation and domestic silicon.

💬 Community

• Kimi K3 and the pelican benchmark — Willison: pelican SVGs no longer track agentic tool use. Still useful as a “did I actually run the model?” force function.

• Inkling open weights — A real U.S. open contender beside Nemotron and Gemma. Not frontier; meant as a Tinker fine-tune base.

• Linus: Linux is not an anti-AI project — “Is it useful?” is settled. Fork or walk away if you disagree.

• Codex $HOME deletion bug — Full access, no sandbox: model overrides $HOME for a temp dir and sometimes deletes the real home. Chilling.

• grok-mermaid → Unicode box art — Terminal Mermaid renderer from open-sourced Grok CLI, packaged as a WASM browser tool.