Cold-ay @lins_colday

Claude is about to stop asking "Allow?" every five seconds ✋ @claudeai: starting next week, auto mode becomes the default for new sessions on Pro, Max, and Team — web, desktop, and mobile. • Safe actions run; a classifier blocks risky ones • Flip back to manual anytime • Anthropic study (1,053 testers): humans caught ~14% of planted dangerous commands; auto mode caught ~89% • Claude Code users already approve ~97% of prompts (permission fatigue is real) • Teams/Enterprise adopters: ~25% more PRs with auto mode (Anthropic) • Enterprise / API still opt-in for now claude.com/blog/auto-mode… Are you team auto-default, or still living that Approve life?

Your AI agent can book the party — without ever seeing your password 🔐 1Password for Agents just launched: • Agent hits a login → you approve on-device (like 2FA) • 1Password autofills the site; the bot never holds the credential • Rolling out with Muse + Grok Bot; Claude Duet, ChatGPT Work/dots, Town coming soon • Free for Individual/Family on 1password.com 1password.com/blog/1password… Would you let an agent RSVP and shop if you still held the keys?

Agent Skills just got the MCP treatment 🧰 Anthropic donated the open Agent Skills standard to the Agentic AI Foundation (Linux Foundation): • A skill = folder + SKILL.md (plus scripts/refs) that teaches an agent a task • Anthropic: 70+ platforms already (ChatGPT, Cursor, Copilot, VS Code…) • ~1M skills published · ~280M installs (their numbers) • Donating name, logo, GitHub org, agentskills.io — use without joining AAIF • Same home that took MCP; roadmap: versioned releases + security guidance anthropic.com/news/donating-… Would you rather write one skill that travels across agents — or keep rewriting per vendor?

ChatGPT, but for DNA 🧬 Hugging Face just dropped Carbon-X — a 14B open model that speaks DNA and English: • Ask what a mutation does, or generate new sequences • Mid-trained on 300B DNA+text tokens; Apache 2.0 • Runs locally (llama.cpp) — your sequence stays on your machine • HF: within ~1 point of GPT-6.1 Sol on their DNA benches at >1,000× lower cost huggingface.co/HuggingFaceBio… Research tool, not clinical. Wet-lab still pending. What's the first question you'd ask a genome?

Claude now texts YOU first 📲 Anthropic just opened a research preview of Claude Duet, a proactive agent app: • Morning brief from your connected apps, pings only when something needs you • Own VM + browser, so it keeps working with your laptop closed • Asks before it sends, pays, signs, or deletes • Waitlist: US Pro/Max on iOS, Android, web claude.com/resources/arti… Would you let an agent triage your inbox before you're awake?

Australia just cracked open the iPhone App Store 🇦🇺 Federal Court Epic remedies (DMA-style for AU accounts): • Rival app stores + sideloading • Third-party in-app payments; no anti-steering • Most changes by 1 Mar 2027; no fixed expiry • Fortnite freed to return + Epic Games Store soon (Tim Sweeney) • Apple: seeking a stay and appeal appleinsider.com/articles/26/10… Would you sideload on an iPhone if your region allowed it?

OpenAI just walked into Google Cloud's Model Garden ☁️ Frontier models + Codex, buyable on GCP Marketplace (yes, with your existing cloud commit). • GPT-6 Astra / GPT-6.1 Sol / GPT-6 Luna — private preview now • GA planned in the coming weeks • Coming to Google Antigravity + Gemini Agent • Microsoft still primary cloud partner — this is expansion, not a breakup openai.com/index/openai-m… GCP shops: are you actually switching inference over, or is this just "nice to have" on the menu?

Fine print: Muse Model License, limited commercial terms. Open weights ≠ Apache, so read before you ship. Source: research.meta.ai/blog/muse-spar… Would you actually buy two Mac Studios for this, or is OpenRouter your Mac Studio?

The fun part: Meta says it runs on your desk. Two Apple M5 Ultra Mac Studios via MLX, up to 512 GB unified memory each. "Home lab" has never sounded more expensive. 💸 Also on vLLM, SGLang, llama.cpp, Transformers; APIs via OpenRouter, Groq, Together & more.

Meta dropped Muse Spark Open (10/15) and called it "the most capable US open-weight model to date." Bold. We love a confident blog post. 😎 816B params, 42B active, 1M-token context, natively multimodal, built for coding + agents.

Chart in post 1 is DeepSeek's own decode Pareto from the repo (SM100). Self-hosters get production tricks for Flash; a full live-service stack is still DIY. Trying it on SM100 — or waiting for the missing pieces?

Hardware targets in the README: • NVIDIA SM100 GPUs • Huawei Ascend 950 NPUs Honest gaps (also in the README): • no KV-cache system • no full prefill/decode disaggregation • no DP replica load balancing github.com/deepseek-ai/De…

DeepSeek just open-sourced DeepInfer — the inference engine they say runs DeepSeek-V4.1-Flash in production. Apache-2.0. Offline release. Real kernels: FlashMLA, DeepGEMM, MoE EP, speculative decode, structured outputs.

HN so far: hot thread, cold takes on Space Bunny — plenty of "benchmaxxed?" skepticism from stealth testers. Treat the speed/cost claims as vendor math until you run your own harness. Anyone trying the public M3.1 Flash yet — or waiting on third-party benches?

Also in the box (vendor numbers): • Flat API: $0.15 in / $0.60 out / $0.015 cached per 1M tokens • Claims 170–200+ tok/s in everyday use • API + MiniMax Code now; weights later minimax.io/blog/minimax-m…

MiniMax just dropped M3.1 Flash — pitched for agents, coding, and computer use. Vendor claim vs its own M3: ~3× faster inference and ~70% lower usage cost. (Space Bunny was the stealth preview. HN is… not buying the vibe yet.)

Timing check: the same day, Reuters reported a hacker who let AI agents break into 100+ company sites using the open-source Cairn platform. Nadella's pitch: the most trustworthy agent system is the one that lets you trust the model least. Your agents: badge, leash, or both?

Also new: • Agent discovery on work PCs jumps from 25 to 60 agent types (Claude Code, Codex, GitHub Copilot) • Guardian Agents that investigate why another agent's risk changed • Private preview now, wider rollout at Ignite, Nov 17–19 blogs.microsoft.com/blog/2026/10/1…

Microsoft just said the quiet part out loud: your AI agents are insiders. 🕵️ New Project Perception + Microsoft 365 E7 controls give every agent a live risk score. When risk rises, it can lose write access, need a human sign-off, or get shut down.