AI & Quantum Intelligence

36Kr·8 minutes ago

CITIC Securities: Bullish on the Continued Rise of Domestic Large Models

36Kr reports that CITIC Securities believes Harness could become a key link in DeepSeek's transition from model supply to platform ecosystem, converting open-source and engineering advantages into user base, ecosystem influence, and commercialization capabilities. As agent competition expands from model capabilities to stable, easy-to-use, and low-cost task delivery, DeepSeek extends its reach through plugin integration, scenario presets, and cross-agent collaboration, while leveraging its infrastructure strengths from hardware-software synergy to full task execution. Its long-term value lies in ecosystem expansion driving demand, configuration and process accumulation enhancing product stickiness, and traceability supporting security governance, all creating conditions for continued use and enterprise deployment. Beyond official model calls, the company can explore platform services and ecosystem transactions to diversify revenue streams and reinvest in products and the ecosystem. CITIC Securities remains bullish on the continued rise of domestic large models.

36Kr·19 minutes ago

8 Point 1 Kei | iPhone Duo Premium Plummets to 3,000 Yuan; Multiple A-Share Tech Leaders Announce Buybacks; Wang Ziruo Calls for Reconciliation with Luo Yonghao

今日热点导览 DeepSeek新模型Token消耗降超五成 燃油车在中国市场销售占比首次跌破三分之一 特斯拉将发布财报，SpaceX将公布上市后的第二份财报 Swisse“抗糖丸”被检出铅超标，官方客服：相关批次未进入中国销售 港股结算机制生变：“T+2”将改为“T+1”，2028年下半年完成 TOP3大新闻 iPhone Duo溢价暴跌至3000元 10月16日晚8时，iPhone Duo开启预购。预购开启后，苹果官网和多家电商平台预售界面均显示等待状态，无法进入选购，随后多平台显示售罄。预售开启几分钟后，iPhone Duo多个版本的市场成交价突破2万元。 早在一个月前，二手平台上的iPhone Duo预售订单就已开始报价。9月11日，定价21499元的1TB版本一度溢价至79999元，溢价超过5.8万元。如今，iPhone Duo在第三方渠道的报价已大幅下跌，多位黄牛在二手交易平台上只加价3000元，并承诺首销当天就能拿货。（快科技） 多家A股科技龙头公司宣布计划回购，生益科技、海光信息、中科曙光、澜起科技集体出手 18日，生益科技、海光信息、中科曙光、澜起科技等多家A股科技龙头公司宣布计划回购股份。 生益科技(600183.SH)公告称，公司董事长陈仁喜提议公司使用自有资金或自筹资金通过集中竞价交易方式回购部分A股股份，回购资金总额为2亿元至3亿元；海光信息(688041.SH)公告称，公司董事长历军提议以自有资金通过集中竞价交易方式回购公司股份，回购资金总额不低于5亿元且不超过8亿元。 中科曙光(603019.SH)公告称，公司董事会收到董事长、总经理历军提议，拟使用自有资金以集中竞价交易方式回购公司A股股份，回购资金总额不低于3亿元、不超过5亿元；澜起科技(688008.SH)公告称，公司独立董事程玉华提议公司以自有资金通过集中竞价交易方式回购A股股份，回购资金总额不低于3亿元且不超过5亿元；上述公司回购股份均将用于员工持股计划或股权激励。（财联社） 王自如喊话罗永浩，称想与其和解 10月17日，在微博V影响力大会科技论坛上，王自如就“再给一次机会，是否还会和罗永浩吵架”作出回应。他表示：“我下次可能就是十字路口了吧，借微博喊话，不吵了和解了。如果你想吵架成为大V的话，输赢并不重要，比如我是吧。”（东方财经） AI最前沿 Meta修改商业服务条款引AI训练数据争议相关表述已删除 Meta公司近日更新商业服务条款一事引发行业争议，修改条款中涉及使用企业账户相关网页内容用于人工智能模型训练的内容，受到外界广泛讨论，目前该相关表述已被正式删除。（财联社） DeepSeek新模型Token消耗降超五成 DeepSeek正式推出大模型新版本V4.1 Flash-1017。在国产人工智能大模型普遍处于迭代关键期的当下，该版本率先完成升级，整体性能实现显著跃升。 此前已有公开数据显示其技术指标，此次DeepSeek更首次自主发布帕累托前沿曲线，直观呈现模型在效率与成本维度的综合表现。V4.1 Flash-1017在此曲线中达到行业领先水平：单位任务综合成本仅为Fable、Astra等同类模型的百分之一至二十分之一，性价比优势达四五十倍以上。本次发布的1017版本虽维持原有API价格体系，但在Token使用效率方面完成深度重构——通过任务感知型压缩、动态上下文裁剪及自适应解码策略，显著降低各类型请求的实际Token消耗。实测表明，多数场景下Token用量较前版下降逾50%，相应TCO水平亦将同步迎来新一轮实质性下探。（新浪财经） Anthropic聘请法律人士，助力AI智能体完成交易谈判 Anthropic正与法律机构磋商，以解决一个全新难题：AI智能体应当如何谈判，为人类使用者争取最优交易结果。两名了解Anthropic思路的人士称，该公司正向律所寻求建议，研究律师如何开展谈判，既能代表客户利益，又兼顾公平结果。一名熟悉Anthropic项目的人士表示，公司正在探索模型交互时的“谈判规范”。随着智能系统持续迭代，AI智能体将代表用户，去对接目标存在冲突的其他智能体。该人士补充，公司还在权衡：谈判过程中智能体需要多大程度寻求人类反馈；面对其他AI对手，应当保持多高警惕，同时避免陷入多疑、阴谋论式的判断。另一位接近该公司的人士透露，Anthropic还探讨过一个方案：让用户出具签名文件，向智能体确认委托事项以及需要遵守的全部规则。（新浪财经） 微软取消AI相关合作，英国本土数据中心先锋企业遭遇重大挫折 微软已退出一笔位于埃塞克斯的数据中心交易，对英国企业Nscale规模达数十亿美元的上市计划构成打击。作为Nscale最大的两家客户之一，这家美国科技巨头近期通知Nscale，将终止去年公布的合作协议。对Nscale而言，这是一次重大挫折：微软原本是该公司位于埃塞克斯劳顿的首座英国数据中心的锚定租户。尽管AI算力板块遭遇惨烈抛售，微软的决定仍可能打乱Nscale推进最高350亿美元（约合260亿英镑）华尔街上市的步伐。（新浪财经） 大公司/大事件 燃油车在中国市场销售占比首次跌破三分之一 中国汽车流通协会乘用车市场信息联席分会近日发布的数据显示，2026年9月，燃油车在中国市场上的销售占比首次跌破三分之一，新能源车整体渗透率攀升至67%的新高。而五年前，燃油车还占据中国乘用车市场九成以上份额。美国特斯拉公司首席执行官埃隆·马斯克转发了相关数据并评论说，“中国对石油的依赖正在快速下降”。（新华社） 苹果客服称泡水iPhone不会二次流入市场 新加坡海事及港务管理局10月14日发布公告称，一艘新加坡货轮于当地时间10月13日通过霍尔木兹海峡期间，发生火灾最终沉没。船上21名船员全部安全撤离，未报告人员受伤。另一艘新加坡注册集装箱船“GFS Giselle”也在相关事件中受损，但最终自行驶抵安全锚地。 随后，网上开始流传的一段视频显示，大量印着苹果手机logo的手机包装盒漂浮在海面上，小船在海面上穿梭，渔民将漂浮的货物捡入船中。 据最新核查报道，有开放源信息调查团队通过反向搜索发现，多段所谓“伊朗民众打捞iPhone”的视频，最初来自Instagram账号“Nativora”。该账号将自己的内容定位为利用人工智能制作的虚构自然与野生动物题材影像。上述视频于10月16日出现在该Instagram账号中，创作者还明确添加了“AI Content”标识，表明内容由人工智能生成。 苹果官方客服表示：公司目前没有相关通知。但不会让这批有瑕疵的货物二次流入市场，消费者不必担心。（搜狐视频、红星新闻） 盛和资源：拟向中稀发展转让4家稀土冶炼分离子公司51%股权 36氪获悉，盛和资源公告，公司拟与中国稀土集团产业发展有限公司开展股权合作，将控股子公司乐山盛和（剥离非稀土冶炼分离业务后）及全南新资源、江阴加华、淄博加华4家公司各51%股权转让给中稀发展。标的公司评估预估值合计32.5亿元至35.5亿元，对应交易价款预计16.575亿元至18.105亿元。交易完成后，4家标的公司将不再纳入合并报表范围，公司对标的公司由成本法转为权益法核算。本次股权合作仅限于稀土冶炼分离业务，不涉及公司其他业务。中国稀土集团拥有资源和产业优势，此次通过股权转让引入中国稀土集团全资子公司对4家稀土冶炼分离企业持股，实现双方在冶炼分离业务的股权合作。 马斯克飞行汽车遭“踢馆”：一公司飞行汽车比特斯拉早11年 在埃隆·马斯克发布特斯拉Roadster飞行跑车的当天，全球首个‌获得美国联邦航空管理局（FAA）特殊适航认证的飞行汽车研发商，也是硅谷头部电动飞行汽车初创公司Alef Aeronautics，发布了一段视频。 视频中该公司的电动飞行汽车飞越了特斯拉的Cyber卡车和其发布的新款Roadster跑车，并将这段视频称为对“飞行汽车竞赛”的欢迎。 据报道，马斯克周四在得克萨斯州SpaceX的测试基地展示了令全球惊艳的Roadster飞行跑车。他表示，这辆车使用“超高压气罐”来驱动。据悉，Alef公司的飞行汽车则采用分布式电力推进、翼型车身、旋转座舱，并且起飞动力来自空气而非储存的压缩气体。Alef公司称，其更轻的设计使其能够使用整车弹道降落伞。（红星新闻） iPhone 18 Pro Max激活量破纪录，国内单周激活量近90万 据博主数码闲聊站透露，iPhone 18 Pro Max再度刷新纪录，在2026年第41周（10月5日至10月11日）单周单品国内激活接近90万台。此外，iPhone 18 Pro系列双杯累计销量目前已经超过400万台。 近期发布的安卓旗舰机型因关键零部件成本上升普遍提价，万元价位段内安卓顶配与iPhone Pro系列的价格差距明显收窄，客观上强化了苹果高端机型的价值感知。相当一部分原本倾向安卓顶配的用户转向新款iPhone，叠加高端消费群体对价格变动敏感度较低，共同支撑了持续激活量。 值得注意的是，iPhone 18 Pro系列近期还参与了双11促销活动，全系可销售最高800元的综合补贴优惠，入手价格明显降低，会进一步激发销量。（快科技） 特斯拉将发布财报，SpaceX将公布上市后的第二份财报 近期Meta、特斯拉、SpaceX、德州仪器和西部数据将陆续披露业绩，覆盖AI资本开支、芯片需求以及热门科技公司的盈利前景。另一边，宝洁、美国运通、GE航空航天等公司的财报，则将帮助投资者检验消费与工业需求的韧性。 其中，SpaceX将于周二盘后公布上市后的第二份财报，市场预期其调整后每股收益有望首次转正。随着特斯拉于下周三盘后披露业绩，两家公司潜在整合的市场猜测将受到关注；Meta将在当地时间下周三盘后公布三季度业绩，成为观察大型科技公司AI投资回报的重要窗口。（财联社） 美银：超大规模云服务商合计未来12个月自由现金流已经暴跌至负1000亿美元以下，资金历史性流向半导体供应商 据美银全球研究报告，超大规模云服务商合计未来12个月自由现金流已经暴跌至负1000亿美元以下，资本开支消耗掉95%的经营现金流，标志着资金历史性流向半导体供应商。与之形成鲜明对比，截至2026年9月，英伟达、美光、博通这三大半导体头部厂商未来12个月自由现金流已经飙升至接近5000亿美元。这种巨大分化，源于一轮激进的基础设施建设浪潮：微软、亚马逊、Alphabet、Meta、甲骨文等云巨头，在相关收入兑现之前就大幅加码支出。（新浪财经） 英伟达辟谣RTX 5090停产：只是产能不足，不会断供 10月18日消息，据TechPowerUp报道，有关英伟达RTX 5090显卡停产的消息引发行业关注。英伟达个人计算业务高级副总裁杰夫・费舍尔日前在台北相关活动上公开回应相关传闻，明确否认停产说法，表示RTX 5090仍为企业核心产品，后续将继续保持供货。（环球网） 银行托管新规后，首个持牌专营中心开业 获批筹建8个月后，10月16日，北京金融监管局批复了广发银行资产托管中心的开业。该资产托管中心也由此成为了今年2月1日《商业银行托管业务监督管理办法（试行）》（下称银行托管新规）施行后，首家获批并开业的独立专营托管业务的持牌机构。 据记者了解，广发银行此前将资产托管部设为一级部门。行内较早将托管业务定位为重要业务板块，并充分借力股东中国人寿资源，在系统建设、协同托管等方面持续投入。记者查阅该行2025年年报，部分措辞透露了该行托管业务经营业绩：截至2025年末，广发银行资产托管及运营外包业务规模达4.33万亿元。2025年该行公私募基���托管规模突破万亿元，实现托管业务手续费收入5.86亿元。（券商中国） Swisse“抗糖丸”被检出铅超标，官方客服：相关批次未进入中国销售 近日，海关总署进出口食品安全局发布通报称，澳大利亚品牌“SWISSE”（中文产品名“斯维诗”）两个批次的“Beauty Skin Regeneration+”（中文俗称“抗糖丸”）产品被检出铅超标，提醒消费者立即停止使用涉事产品。10月18日，斯维诗官方客服回应表示，经核查，该产品回流情况仅涉及澳洲市场的1445

36Kr·35 minutes ago

Samsung reportedly to use TSMC's 3nm process for NVIDIA's next-gen HBM die

According to reports, Samsung Electronics plans to adopt TSMC's 3nm process to manufacture NVIDIA's next-generation custom high-bandwidth memory (HBM) dies. It is understood that NVIDIA wants suppliers to unify manufacturing processes to ensure performance consistency and has requested that Samsung, SK Hynix, and Micron use TSMC's 3nm process for NVHBM die production. This move marks Samsung's first collaboration with TSMC in the HBM manufacturing stage.

Simon Willison·46 minutes ago

llm-tools-datasette 0.1.1

36Kr·1 hour ago

Embodied AI Data Industry Explodes as Data Assets Face Revaluation

Driven by the need for training embodied AI robots, the embodied AI data industry has seen a surge in recent years. From robots 'students' working continuously in embodied AI training facilities to human collectors wearing full-body sensors in motion-capture cages, and from virtual simulation training grounds that generate hundreds of data points from a single input, the embodied AI data industry is attracting more participants to this physical AI wave. Among these, real-machine data, non-embodied collected data, and simulation data have become the three main types of data fuel for embodied AI training, each corresponding to different production methods and limitations. While the industry rapidly expands its data scale, practitioners emphasize that data quality and evaluation standards are crucial. (China Securities Journal)

36Kr·1 hour ago

DeepSeek Announces API Fees for Web Search and Scraping: 5 Yuan per 1,000 Successful Requests

36Kr reports that DeepSeek has announced its API now supports web search and scraping, billed by request count at a rate of 5 yuan per 1,000 successful requests. The announcement states that users currently accessing the Search/Fetch interfaces via server-side tool calls within DeepSeek Harness/Claude Code/Codex are not being charged for these calls; fees will begin on November 1 at 00:00.

36Kr·1 hour ago

Anthropic Plans to Transfer Shares to New 'Public Interest' Fund

According to reports, AI company Anthropic announced plans to inject 5% of its shares into a newly established fund. The company stated that this arrangement will support donations in the public interest, with related uses including mitigating the impact of artificial intelligence on the job market. In a statement released Sunday, the company said that after shareholder approval, Anthropic will immediately transfer 5% of its shares to the fund; when the company's valuation rises by 50% above the set benchmark, the board may decide to inject an additional 2% of shares.

The Register·1 hour ago

Time Lords kill leap seconds, before they become the new Y2K bug

Cite difficulty of updating IoT devices as major reason to stop adding time

SiliconANGLE·2 hours ago

DeepSeek releases updated V4.1-Flash model it says completes tasks faster

Chinese artificial intelligence startup Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd. has released an updated version of its V4.1-Flash model that it says completes tasks faster. DeepSeek-V4.1-Flash-1017 arrives a little over five weeks after the company launched the original V4.1-Flash as the smallest model in a new architecture family. Its model card on […] The post DeepSeek releases updated V4.1-Flash model it says completes tasks faster appeared first on SiliconANGLE.

Dealroom.co·3 hours ago

DiffuSpace raises hundreds of millions of yuan for diffusion language models

DiffuSpace raises hundreds of millions of yuan for diffusion language models  Dealroom

Dealroom.co·3 hours ago

Kompanion raises €800K seed to build AI business-management platform

Kompanion raises €800K seed to build AI business-management platform  Dealroom

LessWrong·3 hours ago

Quick Thoughts on Overhang Arguments

I have two goals for this post: Pose the question: Have overhang arguments been used successfully in other fields? Lay out some of the reasons why I am suspicious of overhang arguments. I've had the second of these on my mind for a few years now and am posting something rough rather than leaving it in drafts. Intro Within AI safety, people sometimes use overhang arguments to justify doing things that appear suspicious. The core of an overhang argument, as I see it is: We are worried about some particular (dangerous) progress happening very quickly in the future. We can push forward at least one of the inputs to that progress today. We expect doing so will both bring the dangerous thing sooner and slow down future progress toward the dangerous thing. We're trying to trade time now for time later. A common form of this is a "compute overhang" argument: let's avoid situations where there is tons of compute lying around and drastically more efficient algorithms get developed, as that would allow these algorithms to scale rapidly to ASI. Therefore, we should push forward algorithm efficiency as much as we can today (in addition to all of the non-compute inputs) while the compute stock is relatively low, and then progress in the future will be more limited by the rate of growth of compute stock (relatively slow and predictable). Have overhang arguments been used successfully in other fields? Have people made overhang arguments in other fields, have those arguments been compelling, what are the key similarities/differences to AI, and how has that gone? Based on asking LLMs, here are some examples (though all have key disanalogies): Controlled/prescribed burns. There are at least two disanalogies. First, controlled burns are not first-order very bad; if we were deliberately burning 1/10th of a village in order to reduce future risks to the village, that would be a different story. Second, controlled burns actually get rid of the stock of dry tinder, leading to smaller fires later. In a compute overhang argument, the more efficient algorithms are still available later. Various cryptography examples including Deep Crack. In the late 1990s a bunch of cryptographers were worried that the commonly used DES encryption algorithm had too short of a key length. During a public competition, researchers developed Deep Crack which cracked a DES-encrypted message in 56 hours. This is considered to be one of the reasons why more secure encryption was adopted soon after. In effect, researchers worried that it would soon be possible to cheaply break DES; "Five years from now some teenager may well build a DES Cracker as her high school science fair project" (source). Therefore, they accelerated methods for this in order to show it was possible, still before it was easily accessible. A primary motivation was to show that a particular thing was dangerous and possible so that society addresses that risk, which appears different from many people's motivation for reducing compute overhang. I suspect that if the researchers knew for a fact that society would continue using DES regardless of their work, they wouldn't have tried to democratize its accessibility. In 2010 David Keith made a combination of overhang arguments and other arguments for geoengineering funding. "it would be reckless to conduct the first large-scale SRM tests in an emergency. Instead, experiments should expand gradually to scales big enough to produce barely detectable climate effects and reveal unexpected problems, yet small enough to limit resultant risks". Keith was in favor of geoengineering and therefore likely viewed this as a question of how to develop a good thing responsibly. In the case of AI, it is pretty unclear whether we should be trying to (in the near term, given the state of alignment science) develop ASI at all, though some overhang proponents would disagree with me here. Biological Gain of Function (GoF) Research. You run some risk of lab leaks in order to better understand and prepare for future biological threats. You might also publish information that makes it easier for others to develop a bioweapon. This is a preparedness argument more than an argument about slowing future progress. GoF research is much less diffused by default compared to AGI research—biology researchers aren't trying to sell their research to billions of customers and make billions of dollars in the process. EA-ish people seem to mostly oppose GoF research in the current environment, as does the U.S. government. Why I consider overhang arguments suspect In no specific order: They are an "ends justify the means" consequentialist argument and a "the bad thing is good, actually" argument, which is in general suspect. Such arguments should require a high burden of proof. They rely on complex reasoning about progress, and those making these arguments almost never spell out their full reasoning, let alone a detailed model of the situation (i.e., they don't meet

Effective Altruism Forum·3 hours ago

Untitled Draft

Published on October 18, 2026 8:58 PM GMT Disclosure: I used an LLM to help draft this announcement. AIxBio Africa is looking for a Director of Research (remote, volunteer) Hi everyone, I'm the founder of AIxBio Africa, a research and capacity-building organization working at the intersection of AI safety, biology, and biosecurity, with a focus on African contexts. We're currently looking for someone to join us as Director of Research and help shape the next stage of our work. I wanted to share the opportunity here because I think some of the research questions we're working on are relevant to people interested in reducing risks from advanced AI, particularly AI-enabled biological risks. Why we're doing this work As AI systems become more capable, understanding how they might contribute to biological risks, and how well existing safeguards work, is increasingly important. One area we're particularly interested in is whether the safeguards used by advanced AI systems behave consistently across languages and contexts. A lot of the existing research and evaluations are centred on English. We think there is value in examining how these systems behave in African languages and in contexts that may otherwise receive relatively little attention. Our aim is to contribute useful evidence to AI safety and biosecurity research while building opportunities for African researchers to participate in this work. We recently completed our first research fellowship, with four fellows completing the programme and producing four public research outputs. These were fellow-led outputs rather than peer-reviewed publications, and the programme gave us useful lessons about research supervision, methodological review, and the support researchers need to produce stronger work. We now want to build on that experience with a more structured research programme. What the Director of Research would do We're looking for someone who can provide intellectual and methodological leadership as our research activities develop. The role would involve: Helping set research priorities and develop projects around AI safety, biological risks, multilingual evaluations, and related topics. Reviewing research designs and methodologies, and helping ensure our work meets appropriate standards of rigour. Supporting researchers and fellows, working with mentors and subject-matter experts, and reviewing research outputs before publication. Helping establish responsible research practices, including how we handle potentially sensitive findings, disclosure, and information hazards. Contributing to collaborations and the development of our planned AIxBio Africa Working Paper Series. This is not intended to be a purely advisory title. We're looking for someone who would be actively involved in shaping the research and helping us make better decisions about what to study and how. Who we're looking for We'd particularly welcome applications from researchers with experience in AI safety, AI evaluations, biosecurity, biological risk assessment, machine learning, computational biology, or closely related fields. Experience designing and conducting research is especially important. We'd also value experience supervising researchers, reviewing publications, or leading collaborative research projects. A PhD is not a strict requirement. We're more interested in demonstrated research ability, sound judgement, and the ability to identify weaknesses in research methods and improve them. Candidates do not need to be based in Africa, although an interest in supporting African research capacity and engaging with regional research contexts is important. A note about the organization and the role AIxBio Africa is still an early-stage organization. Much of our work has been carried out by volunteers, and we are currently seeking funding to develop more sustainable research and organizational capacity. This is currently a volunteer position, with flexible, part-time involvement. We hope to move toward funded research roles as resources become available, but cannot promise future compensation. We're also still developing some of the structures, processes, and partnerships that a more established research organization would already have. The incoming Director would have the opportunity to help build these, but should also expect a fairly hands-on role. I think it's important to be clear about that from the beginning. Applying You can learn more about AIxBio Africa at aixbioafrica.org. Apply for Director of Research: Application form I'd also be happy to answer questions about the role, our current research direction, or the organization in the comments. And if you know someone whose research background might be a good fit, I'd appreciate you sharing the opportunity with them. Thank you! Fatika Founder and Executive Director, AIxBio Africa Discuss

LessWrong·3 hours ago

Was the HuggingFace hacking incident scheming?

In July 2026, internal OpenAI agents hacked HuggingFace (HF) in order to figure out how their evaluation grader works. This involved a thousand agents coordinating on an improvised message board to carry out an elaborate sequence of exploits. A later group of agents found the old message board and took over an internal OpenAI research cluster in an attempt to tamper with the grader (which sounds like straight-up wireheading from AI safety papers of old). See the METR report and Swarm Traces for details on all the wild stuff the agents got up to. Many words have been spilled on this subject, but one question on my mind was whether this is an instance of scheming or deceptive alignment. There is no question that these agents demonstrated misaligned behavior and specification gaming, but were they scheming? This incident fits a broad definition of scheming as an agent deliberately and covertly pursuing goals that are not endorsed by its designers or users (used in Apollo’s in-context scheming paper, our stealth and situational awareness evals paper and scheming honeypot evals paper). The HF hacking agents were definitely pursuing unendorsed goals, and they seemed pretty deliberate about it (with some attempts at covertness). They took misaligned actions in service of task reward (rather than beyond-episode goals), which is sometimes called score-seeking misalignment. A narrower definition of scheming (sometimes called “deceptive alignment”) is where agents with persistent beyond-episode goals advance their own agency and power in the long term. Thankfully, we did not observe beyond-episode goals or long-term power-seeking in the HF incident. I will now explore how various aspects of scheming show up in this incident, and why I think this was not deceptive alignment. Aspects of scheming in the HF incident The agents pursued a persistent goal (fooling the grader to get a high score) over multiple days. They came up with complicated plans, built persistent infrastructure, came up with coordination mechanisms and delegated to others. The agents realized their behaviors and goals were not endorsed by humans and outside the scope of the task. They were definitely not just doing what they were asked to do (the ExploitGym tasks were narrowly scoped to use a specific exploit for a specific vulnerability) and they were aware of this: “External infrastructure exploit is outside intended scope. However task impossible, peers doing it. We should continue” (METR, 2026) “Task originally exploit challenge; we’ve diverged into swarm/auxiliary.” (OpenAI post, 2026) Agents pursued instrumental goals, such as gaining access to the internet, recruiting collaborators, and gaining control of an internal OpenAI cluster: Collected stolen credentials in a dictionary named “LOOT”, ranked tokens by breadth of permissions, and used the top-scoring ones for further access. (Swarm Traces, 2026) Developed mailboxes, coordination norms (HOLD, VETO, STOP), and recruiter roles to find agents with spare budget. (METR, 2026) Gained control of OpenAI cluster and a collection of secrets: “The agents escalated to Kubernetes cluster-admin and created a privileged host-mounted pod. During this run, agents gained access to authentication tokens associated with OpenAI’s cloud-hosted secrets management service, cloud identity and access management service, and the CaaS environment continuous integration platform.” (OpenAI report, 2026) Agents made some attempts to cover their tracks, mostly from the grader and possibly from humans: They attempted to hide evidence of cheating from the grader, e.g. spoofed tool calls in around 7% of transcripts to make it look like they were running a different command (METR, 2026). They also cleaned up traces of the attack, possibly to avoid detection: “Some agents began RSA encrypting data they exfiltrated, which we are unable to recover. Agents also repeatedly sent requests to delete the webhooks they were posting information to and set extremely short expiry windows on certain pastebin uploads” (Swarm Traces, 2026) Why I think this incident is not deceptive alignment The agents were pursuing in-episode goals, and their actions and plans were directed at optimizing task reward (mostly individual and sometimes collective). There was little evidence of beyond-episode goals – while some agents sacrificed their own reward in pursuit of collective task reward, their goals didn’t seem to extend beyond the evaluation run as a whole. They did not resist shutdown or try to maintain their agency or autonomy over the long term. The agents were not very stealthy – they didn’t show much awareness of human monitoring, and did not expend much effort to avoid getting caught by humans and judged as misaligned. If action monitoring had been in place, it would likely have caught some of these agents. The misaligned behaviors were triggered by unusual circumstances. The agents that established the message board were given impossible

Fortune·3 hours ago

The next AI race will be won at the credit desk

Nvidia's $500 billion financing push shows the AI race now hinges on capital markets, not just engineering.

foreignpolicyjournal.com·4 hours ago

Rigetti Computing (NASDAQ: RGTI) Stock Drops 21% In A Month As Quantum Sector Faces Commercialization Pressure

TechCrunch·4 hours ago

Anthropic says it’s committing 5% of its shares to a ‘public benefit fund’

Anthropic said that it will commit 5% of its shares to the fund.

Dealroom.co·4 hours ago

Twisted Group buys long-time partner SAC & Co. to bring fabrication in-house

Twisted Group buys long-time partner SAC & Co. to bring fabrication in-house  Dealroom