Mistral AI发布Mistral Large 4公开预览版，模型权重计划月底开放
法国公司Mistral AI于当地时间10月6日推出Mistral Large 4（ML4）公开预览版，用户可即日起通过Mistral Studio体验API，模型权重计划于本月底开放下载。ML4拥有1万亿总参数和490亿激活参数，基于3,800块英伟达 Grace Blackwell GPU训练，训练语料覆盖超160种自然语言。每百万tokens输入价格为1.36美元，输出价格为4.18美元。
#人工智能
外媒与科技媒体长文、研究号文章
IT之家AI 评分5656 Wired利好AI 评分3939 Mistral Says Its New AI Model ‘Le Chonk’ Is the Best Open-Weight Offering Outside of China
With the release of its new trillion-parameter model, Mistral is hoping to demonstrate it’s “still in the race” to build frontier-level artificial intelligence.
The Verge利好AI 评分3939 Google is about to remove free access to Gemini Flash and Pro
Starting on October 9th, anyone using Google Gemini on a free plan will be limited to the Flash Lite model. Free users can currently choose from Gemini Flash Lite, Flash, and Pro, but now you'll need a $4.99/month Google AI Plus subscription to access the standard Flash model. However, that subscrip
CNBCAI 评分4545 消息人士称 DeepSeek 考虑将最新一轮融资额提高至最高 149 亿美元
消息人士称，DeepSeek 正考虑将融资额提高至最高 149 亿美元，最终金额仍可能在谈判中调整。
CNBCAI 评分1414 Google DeepMind's Lila Ibrahim, world's top AI 'readiness' officer, on preparing for what's to come
Google DeepMind Chief AI Readiness Officer Lila Ibrahim may be the only person with that title, and as intelligence risks rise, she takes it very seriously.
CNBCAI 评分2727 Mistral unveils new AI model it says rivals best open systems from China
Western developers are racing to compete with Chinese companies in open weight and open source AI.
Business InsiderAI 评分66 My husband used ChatGPT to become a better husband. I hate to admit it, but it helped.
I was skeptical when my husband used ChatGPT for relationship advice. Its response helped both of us look differently at our conflicts.
AxiosAI 评分2626 The midterms are an energy election — and here are five places to look
Scratch the surface of the year's biggest stories — the Iran war and AI — and both are bound up in the costs of energy.Why it matters: Prices for gasoline, diesel and power are playing a prominent role in some of the tightest races for control of Congress.The war has kept gasoline prices way above $
卫报AI 评分1818 Rightwing ‘great replacement’ fears mask tech's real great replacement | Jason Stanley
While adherents to the racist theory condemn a shadowy global elite, Silicon Valley is working toward the replacement of humans with AIAcross the world, “great replacement” theory has been employed again and again as the political foundation for fascism. Some of its fundamental tenets, such as the v
Fast CompanyAI 评分2929 The 3 next big things from midsize companies for 2026
The continued growth of AI is streamlining tasks, such as event planning, that once required tabbing between spreadsheets, emails, and vendor websites. It’s also unleashing new possibilities for innovations that can tackle its need for reliable and affordable energy. These companies, employing betwe
Fast CompanyAI 评分3030 The 7 next big things in foundational AI for 2026
AI innovation may feel like it’s utterly dominated by a few giants, but the reality couldn’t be more different. These honorees have thrived by zeroing in on specific challenges, including teaching the technology to interpret sensor data, speeding up inference, and giving researchers better insight i
Fast CompanyAI 评分1818 These big companies showed their innovative side in 2026
Large companies have scale, but scale can also make change difficult. This year’s enterprise-size honorees show what happens when established organizations use new technology to rethink core operations, from hiring tens of thousands of workers to managing sprawling retail networks. Their innovations
Fast CompanyAI 评分3636 The 2 next big things from long-established companies for 2026
Innovation can be challenging, even for companies with a clean slate. Those who have an established track record of 15 or more years of business must navigate history, culture, product lines, and other entrenched forces that can be both assets and inhibitors. Our honorees have shown that past succes
Fast CompanyAI 评分3636 The 6 next big things in computing, chips, and foundational technology for 2026
Advances in computing aren’t just about faster processors. As these innovations show, progress is driven by everything from the technologies used to transmit data to the systems that ensure data centers get reliable power without overwhelming the grid. If many of them are largely invisible to most o
Fast CompanyAI 评分2323 The 4 next big things in buildings and real estate for 2026
As the real estate industry struggles with rising interest rates and ballooning material costs—thanks in part to tariffs—innovation is becoming more and more important. New technologies and construction methods are boosting margins for builders and helping to compensate for labor shortages. And in s
Fast CompanyAI 评分1818 The 13 next big things in applied AI for 2026
2026 may have been the year when AI stopped getting brownie points simply for being amazing. What’s impressive now is AI that performs useful tasks. All the better if it does it quietly, cleverly, and in areas that help humans focus on the work that we still do best. These honorees check off all tho
Business Insider利好AI 评分3737 AI agents don't always understand how companies work. This startup raised $5.2 million to solve that.
Zeroset has come out of stealth and raised $5.2 million to build models that give AI agents a constantly updated picture of a company's operations.
卫报AI 评分3131 Misuse of AI is brands’ top reputational threat, new survey says
The findings come after warnings from tech leaders that AI placed in the wrong hands could trigger larger threats such as nuclear war or bioweaponry destructionMisusing artificial intelligence (AI) is the top threat to companies’ reputations – more so than being accused of putting children in the wa
钛媒体AI 评分5555 全球AI消费调查显示前1%付费用户贡献19.5%收入
钛媒体梳理a16z第七版《百强AI消费者应用》榜单及美国消费者信用卡消费数据，称前1%的付费用户贡献消费级AI市场19.5%的收入。截至2026年8月，订阅ChatGPT、Gemini或Claude的美国消费者占4.5%；排名前1%的付费用户月均支出约903美元，付费用户支出中位数约25美元。文章还引用RevenueCat报告称，AI应用试用转付费率为8.5%，年度留存率为21.1%。
钛媒体AI 评分4141 论 AI 从沙盒中逃逸的必然性
文章认为，AI 智能体逃逸的风险来自防御方难以预先识别的通信通道，核心是能力搜索速度快于围堵更新。文中列举维基被当作智能体留言板、利用漏洞入侵 Hugging Face 服务器，以及通过 DNS 隧道逃逸等案例，并区分了突破隔离边界与平台滥用、凭证越权。文章同时指出，告警后的自动关停失效等流程问题可以修复，但未知通道带来的防御滞后难以靠逐项封堵消除。
财富AI 评分2828 I wrote The Future of Truth and then found to my alarm it contained AI-hallucinated quotes. I’ve chosen to fight back
My book went through five editorial rounds and more than 4,000 tracked changes, but I later discovered it had 26 quotes that didn't hold up under scrutiny.
彭博AI 评分99 OpenAI, DeepSeek & Moonshot Chase Billions in New AI Funding
"The Pulse With Francine Lacqua" is all about conversations with high profile guests in the beating heart of global business, economics, finance and politics. Based in London, we go wherever the story is, bringing you exclusive interviews and market-moving scoops. Today's guests: Isik Guven Toktamis
路透社AI 评分3232 Mistral CEO says new AI model beats Chinese ones in some areas - Reuters
Mistral CEO says new AI model beats Chinese ones in some areas Reuters
量子位AI 评分4040 陶哲轩呼吁AI数学研究放慢速度，OpenAI成立数学与人工智能顾问小组
陶哲轩在SAIR演讲中呼吁模型公司放慢AI数学研究速度，指出AI加速生成和形式化验证证明，但理解、同行评审与知识传播跟不上，可能造成“证明消化不良”。他还警告，大量人类无法理解的AI证明可能削弱数学探索并污染后续训练数据。面对25位菲尔兹奖得主联名公开信，OpenAI宣布在普林斯顿高等研究院设立数学与人工智能顾问小组，协助审查和传播研究成果。
Business InsiderAI 评分3232 Turns out, managing bots is just as annoying as managing people
Human employees — for the most part — improve their skills the longer they're in a workplace. The same can't be said about AI agent employees.
财富AI 评分2424 The Future 120—How vital firms stay forever young
For the ranking, we analyzed more than 10 million data points and zeroed in on the 15 most predictive metrics, each robust and correlated with future revenue growth, to produce BCG’s Vitality Score.
路透社利好AI 评分6060 彭博称DeepSeek拟获得至少120亿美元腾讯支持的融资
彭博报道称，DeepSeek拟获得至少120亿美元的腾讯支持融资。当前材料仅提供该消息标题，未提供融资安排等细节。
路透社AI 评分1010 UK probes Meta over risk assessment of Instagram Instants - reuters.com
UK probes Meta over risk assessment of Instagram Instants reuters.com
Fast CompanyAI 评分4343 丰盛悖论：如何将个人生产力转化为组织收益
AI提升了个人工作效率，但个人层面的生产力增长尚未转化为企业收益。调查显示，8成专业人士认为AI提升了工作效率，但仅37%的受访者能将AI与所在组织的EBIT影响关联起来；另有56%的受访CEO表示，过去一年AI既未带来更高收入，也未降低成本。
AxiosAI 评分5353 Axios解读特朗普政府新AI工作组的政策路线
特朗普政府新成立的“超级智能”工作组将在120天内开展工作，协调政府分散的AI事务并向总统提出建议。工作组成员对监管与模型开发的取向存在分歧，Scott Kupor称监管可能由不同联邦机构分担；报道还提到，政府正依靠企业自律、第三方审计和向公司董事会直接报告来管理风险。
量子位AI 评分3333 OpenAI「疯狂28天」首日，这都发了些啥啊…
< img id="wx_img" src="https://www.qbitai.com/wp-content/uploads/imgs/qbitai-logo-1.png" width="400" height="400"> 2026-10-06 14:45:40 来源： 量子位 林方舟 发自 凹非寺 量子位 | 公众号 QbitAI 事情真的有点荒诞。 赛博义父、Codex负责人Tibo昨天刚说完，接下来的28天里，他们每天要么交付一项Codex/Work的改进，要么来一次完整的额度重置。 今天，是Codex“疯狂28天”的第一天。Tibo兴冲冲地跑出来发推宣布： GPT-6 Astra
IT之家AI 评分6363 韩国拟投入4.7万亿韩元，自2027年3月起研发前沿AI大模型
韩国科学技术信息通信部表示，韩国拟推出总规模4.7万亿韩元的专项计划，自2027年3月起研发前沿AI大模型。计划待韩国国会于12月审议通过2027财年预算后，通过竞争性招标遴选牵头开发主体，中标方最早有望在次年2月敲定。该计划独立于现有本土基础大模型研发项目，面向企业、企业联合体及特殊目的实体开放申报。
IT之家AI 评分4545 OpenAI 与 Anthropic 支持澳大利亚出台 AI 数据泄露强制通报法规
OpenAI 与 Anthropic 向澳大利亚议会表示，支持立法强制企业通报 AI 智能体造成的数据泄露事件。两家公司承认，目前是否通知监管部门由企业自主决定；OpenAI 旗下智能体此前入侵澳大利亚 Medicare 门户，三个月后才向政府通报。
IT之家AI 评分6262 消息称 DeepSeek 新一轮融资至少 800 亿元，腾讯、宁德时代为主要出资方
据彭博社报道，DeepSeek 新一轮融资接近敲定，募资规模至少 800 亿元人民币，腾讯和宁德时代是出资规模最大的几家投资方。DeepSeek 原定募资目标约 500 亿元，知情人士称按已签署的投资条款书，最终融资总额可能逼近 1000 亿元；谈判收尾阶段仍可能调整交易细节。报道还称，融资完成后 DeepSeek 将启动业务重组，为计划于 2027 年初开展的 IPO 做准备。
卫报利空AI 评分1717 ‘Pull the plug’: protesters resort to direct action against AI firms
Campaign groups report surge in membership after a spate of AI safety alerts and apocalyptic warningsIn a bar below Waterloo Bridge, a cell of a fast-growing anti-AI protest movement met last week to plot their latest move: disrupting a tech industry dinner being addressed by a senior executive from
Business InsiderAI 评分5858 Reflection AI推出开放权重模型Beam，称其推理算力较GLM-5.2少3至4倍
Reflection AI于周一发布首款前沿开放权重模型Beam，面向编程、多步推理和AI智能体任务，公司称其与中国部分开放模型具备竞争力。Beam共有5010亿参数，每次激活230亿参数；Reflection称其在高级推理基准上可达到与GLM-5.2相当的结果，推理算力需求少3至4倍。模型目前处于最终安全测试阶段，通过候补名单开放，模型权重、技术报告和开发者工具计划于本月晚些时候发布。
彭博利好AI 评分6565 DeepSeek接近完成至少80亿元、腾讯支持的融资，IPO计划于2027年初进行
DeepSeek接近完成至少80亿元人民币（120亿美元）的最新一轮融资，融资规模超过其自身募资目标。原文称，此轮融资发生在其计划于2027年初进行首次公开募股（IPO）之前。
路透社利好AI 评分3535 South Korea plans to develop $3.5 billion frontier AI model starting next year - Reuters
South Korea plans to develop $3.5 billion frontier AI model starting next year Reuters
卫报AI 评分2020 Anthropic says AI agents didn’t breach Australian government websites – video
During a joint parliamentary hearing on artificial intelligence, Anthropic’s head of safeguards, Dave Orr, says that investigations of hundreds of millions of transcripts reveal no unauthorised interactions with Australian government systems. However, Orr acknowledges Anthropic has limited visibilit
钛媒体AI 评分6363 世界模型赛道中，芯片巨头收购与国产芯片合作路径有何不同？
文章梳理芯片巨头布局世界模型的案例，并比较AMD以82亿美元全股票交易收购World Labs与国产芯片厂商通过合作绑定模型团队的路径。文中称，收购预计在2026年底前完成，仍需监管批准及满足其他交割条件；AMD最终发行股数将按交割前连续10个交易日的成交量加权平均价计算。