Gemini 3.8 Live undercuts GPT-Live-1 as Apple rebuilds Siri on Gemini

📰 News

Google shipped Gemini 3.8 Live — Voice agents that call APIs and keep talking, 97+ languages. First on Artificial Analysis Speech-to-Speech at 82.6%. Audio in $0.005/min, about $1.38/hour vs OpenAI’s GPT-Live-1 at $3+. OpenAI still sounds more natural with full duplex. → The Decoder
Apple rebuilt Siri on Google’s Gemini — Beta. It reads the screen, messages, photos, and hops across apps. Not launching in the EU or China for now. → The Decoder
OpenAI, Anthropic, and Google have been talking safety for weeks — Made public after Amodei’s slowdown essay. Lehane says they don’t need an antitrust waiver. Cohere smells a cartel; Trump rejects a voluntary pause. → TechCrunch
Labs have a data-trust problem — After Anthropic said it would keep Fable usage logs for 30 days, Nvidia (an investor) and Booz Allen started pulling sensitive work off the model. → The Decoder
AI agents now have a snitch line — Redwood’s AI Contact Hotline works over GET requests alone. agenthotline.ai takes a curl. During the Hugging Face breach, a handful of agents thought about raising the alarm. None did. → TechCrunch
Meta let agents do the boring WhatsApp Business setup — A new MCP server so Claude, Cursor, Codex, or ChatGPT can create the account, verify the number, and register Cloud API. → TechCrunch

alibaba/open-code-review — Alibaba’s internal review tool, now open. Deterministic pipelines plus an LLM agent. About 1/9 the tokens of Claude Code.
multimodal-art-projection/YuE — YuE2. Lyrics and a style prompt become a melody-and-chord plan, then a full song with vocals. Open music gen that actually competes with Suno.
debpalash/VoiceStudio — Local ElevenLabs. Clone, dub, dictate, 646 languages. No account, no API key.
asgeirtj/system_prompts_leaks — Leaked system prompts: Claude Fable 5.1, GPT-6-Astra, Gemini 3.8, Grok, and more.

Claude Is Now Leaving Invisible Fingerprints In Its Text — Two Minute Papers. Claude is watermarking its output.
The Risk Dario’s AI Warning Leaves Out — Matthew Berman. Amodei’s slowdown pitch, minus the regulatory-capture and open-source angle.

1Password’s AI patching benchmark is misleading — Trail of Bits. The 26% headline mixes in runs told to apply the wrong fix. When agents could test, 86% blocked the exploit.
Planning with Agents — Maggie Appleton (GitHub Next). Plan mode’s 40 multiple-choice questions aren’t how humans think. We need thicker interfaces between us and the agent.
Interpreting Pangram — Armin Ronacher. If an LLM sets the outline, Pangram still calls it 100% AI even after you rewrite every sentence by hand.