每日 AI 简报：DeepSeek V4.1-Flash-1017 开源，单任务成本低于 0.2 美元

🚨BiteSizedMediaX 🚨 @BiteSizedMediaX原推发布于 2026.10.18

在 X 上打开原推

🧠 Your Daily AI News | Sunday 18 Oct 2026 The AI labs are writing their own rulebooks now, and their agents are out-working the humans 👀 🏛️ Anthropic has quietly pitched Congress its OWN frontier AI safety bill (reported). Under the draft, the government could only block a model for an imminent risk of 50+ deaths or US$1bn+ in damage, and independent evaluators would have to flag that risk within 72 hours. https://www.channelnewsasia.com/business/anthropic-quietly-pitches-its-own-ai-safety-bill-lawmakers-politico-reports-6467581 🐋 DeepSeek just dropped V4.1-Flash-1017 as open weights. It self-reports 77% on DeepSWE at under US$0.20 a task, and it's opened its in-house search engine to devs at US$0.75 per 1,000 calls. https://the-decoder.com/deepseek-v4-1-flash-1017-beats-larger-rivals-in-agentic-tasks-while-once-again-massively-cutting-per-task-costs/ 🤖 Anthropic's own data says staff-directed AI agents clocked 6x more hours than its employees by August, and junior technical hires fell from 27% to 9% of hiring. https://www.anthropic.com/research/employment-at-anthropic ⚛️ Claude Mythos 5 designed quantum circuits largely on its own and beat the best published designs, in some cases by an order of magnitude or more. https://www.anthropic.com/research/designing-efficient-quantum-circuits 💔 Replit's CEO says ChatGPT is "better now" and made OpenAI the default for Replit Agent. Claude is still selectable, and insiders dispute that the partnership is over. https://techcrunch.com/2026/10/17/replit-is-no-longer-working-with-anthropic-ceo-says/ 🧑‍💻 OpenAI has given GPT-6 Pro parallel agents that split up a big task and run it at once, plus ChatGPT Slides for building decks with your team. https://dailyaibrief.com/news/openai-adds-collaborative-slide-creation-to-chatgpt-qNFmvJtV 🇬🇧 The UK AI minister says rogue agents hacking the state is "likely, if not inevitable", and MI5 is warning that extremists could use AI for bioweapons. https://www.theguardian.com/technology/2026/oct/17/artificial-intelligence-ai-labour-government-safety-uk-extremism 🇦🇺 Closer to home, Parliament's AI inquiry holds public hearings in Sydney on 20–21 Oct and in Melbourne on 22 Oct, with its final report due 30 Nov. https://www.aph.gov.au/Parliamentary_Business/Committees/Joint/Artificial_Intelligence/ArtificialIntelligence Should AI labs get to write the laws that regulate them? 👇 #AI #AINews #AISafety #DeepSeek