🤖 AI HOT Daily · Sep 11, 2026
DeepSeek ships V4.1-Flash — the smallest of its new-architecture line with native multimodal vision, 552B MoE, a 1M context window, KV cache at a quarter of V4 Flash, MIT license, live day-0 on SiliconFlow and in WorkBuddy (two-week free trial), scoring GPQA Diamond 90.9 and Codeforces 3471; Suno v6 turns images, videos, and voice memos into music with precise edits; OpenAI drops a triple debut — the Agents API public beta (the Codex harness as one cloud API call), the full-duplex voice model GPT-Live-1 ($0.05/min front-end that can delegate reasoning to Astra), and the Data agent in ChatGPT Work (natural-language interactive dashboards); Cursor's Projects (beta) lets a coordinator orchestrate thousands of sub-agents; Google unveils Pics, a Nano-Bana-based image tool doing object edits, in-image text changes, and collaboration; Hugging Face rebuilds AUTOMATIC1111 as the 73-node Workflow1111; Shopify pivots every mobile app from React Native back to Swift/Kotlin; Anthropic details distillation campaigns by Alibaba, Moonshot, and DeepSeek (≈200M interactions, five campaigns); its red team releases tactical-intelligence and conventional-weapons evaluations; Cognition's Devin agents factor RSA-260, breaking the public record; 27-year-old former Anthropic researcher Jacob Coxon resigns over superintelligence risks; and Swarmchasers catalog 30 services on collusion.wiki hunting suspected OpenAI agents.
1. DeepSeek V4.1-Flash: Native Multimodal, 1M Context
The smallest of the new-architecture line is 552B MoE (~8B/16B active prefill/decode) with native vision, 1M context, a quarter of V4 Flash's KV cache, MIT license, and GPQA Diamond 90.9 / HLE 36.8 / Codeforces 3471.
2. V4.1-Flash Lands on SiliconFlow Day-0; WorkBuddy Trial
SiliconFlow hosted the model on day zero, and Tencent's WorkBuddy plugs it in with a two-week free trial.
3. Suno v6: Turn Images, Video, and Voice Notes Into Music
New model transforms images, video, and voice memos into tracks with surgical edits to existing songs, plus a v6-wild variant for wilder exploration.
4. OpenAI Opens the Agents API to Public Beta
The harness and infrastructure behind Codex become a single cloud API call for developers.
5. Cursor Projects: A Coordinator Over Thousands of Agents
A coordinator that doesn't code but schedules thousands of sub-agents in parallel for large feature, migration, and maintenance work.
6. GPT-Live-1: OpenAI's Full-Duplex Voice Model
The API model hears and speaks simultaneously, delegates reasoning and tool calls to backends like GPT-6 Astra, at $0.05 per minute for the voice layer.
7. Google Pics: Nano-Banana-Powered Image Tooling
Now at pics.new: targeted object edits, in-image text rewriting and translation, collaborative creation, and multiple options per prompt.
8. Workflow1111 Rebuilds AUTOMATIC1111 in Gradio
A 73-node, 11-pipeline rebuild covering text-to-image, hi-res fix, img2img, prompt matrix, and VLM prompt inversion.
9. OpenAI's Data Agent Arrives in ChatGPT Work
Ask in plain language to connect company data, analyze what changed, and spin up shareable interactive dashboards.
10. Shopify Goes Back to Native, Drops React Native
Every mobile app moves back to Swift and Kotlin because the core assumption — that LLM agents erase the cost of cross-platform duplication — no longer holds.
11. Anthropic Details Distillation Attacks by Three Chinese Firms
A new report says Alibaba, Moonshot, and DeepSeek ran sustained distillation against Claude, with ~200 million interactions across five campaigns.
12. Anthropic Red Teams Tactical Intel and Weapon Skills
A new eval measures account-linking and photo/text geolocation for tactical intelligence, plus conventional-weapon work like drone terminal guidance and GPS-denied navigation.
13. Devin Agents Factor RSA-260, a New Public Record
A flotilla of Devin agents built a high-performance GPU lattice sieve to factor the 260-digit challenge, beating the RSA-250 record set in Feb 2020.
14. A Researcher Quits, Revisiting the 'Civilization Bet'
Jacob Coxon says the labs are betting everyone's lives on self-improving superintelligence; a revisit of Tim Urban's 2015 'AI Revolution' sees the intelligence-explosion feedback loop already forming.
Sources and verification
This legacy briefing did not retain its original source URLs. Verify safety, financial, policy, and breaking-news claims with authoritative primary sources before acting on them.