143 Matching Annotations
  1. Last 7 days
    1. A new generation of Islamic terror supporters are using AI to spread 'Slop Jihad' to new audiences on TikTok.

      【方法】文章描述了AI被用于传播'垃圾圣战'的现象,但未提供具体案例或证据。需要了解这一现象的具体表现形式、传播规模以及如何被识别和验证,以评估报道的准确性和深度。

    1. OpenAI President Greg Brockman's recent essay, The Defender's Window, and an industry-wide open letter both call for urgent, collective action across industry and government

      这是一个关于行业共识的声明,但缺乏具体细节。需要了解这些声明的具体内容、签署方的权威性和代表性,以及'紧急行动'的具体含义。AI安全领域的共识可能存在分歧,需要更深入的分析来验证这一说法。

    2. ChatGPT Enterprise does not use business data, including inputs or outputs, to train or improve OpenAI models, and participating organizations will retain the protections associated with the eligible services they choose

      这是一个关于数据安全和隐私保护的重要声明,需要更详细的解释和第三方验证。需要了解具体的技术实现方式、审计机制、以及如何确保承诺得到遵守。政府数据通常具有高度敏感性,此类声明需要透明度和独立验证。

    3. Lawrence Livermore National Laboratory used ChatGPT to accelerate development of a preliminary fusion-research model, compressing work that could have taken months into a matter of hours

      这是一个关于AI在复杂科学研究中应用的声明,但缺乏具体的方法论细节。需要了解AI在模型开发中的具体角色、科学验证过程、以及结果如何被同行评审。如此大幅度的加速可能需要独立验证,特别是考虑到科学研究的严谨性要求。

    4. In North Carolina, the Department of State Treasurer used ChatGPT to identify millions of dollars in potential unclaimed property that could be returned to residents

      这是一个需要更多具体细节的声明。需要了解'数百万美元'的确切金额、识别方法、验证过程以及实际返还情况。AI在财务识别中的应用存在潜在风险,需要确认是否有适当的人工审核流程来确保准确性。

    5. More than one million government employees already have access to ChatGPT through our existing agreements, and this expanded partnership extends eligibility across a U.S. public-sector workforce of approximately 23 million people.

      这是一个需要核实的具体数据声明,涉及政府员工数量和使用比例。需要确认OpenAI如何统计'已有访问权限'的员工数量,以及'约2300万'美国公共部门 workforce 的数据来源。这一比例(约4.3%)反映了AI工具在政府部门的渗透率,但缺乏更详细的用户参与度数据。

    1. Larry Ellison has canceled his plan to sell up to 50 million of his shares in Oracle, or $7.5 billion worth of stock at the current price.

      【数据】这一具体数字表明Ellison原本计划出售的股票规模巨大,相当于Oracle总股本的重要部分。需要核实这一数字是否准确,以及这一计划取消对Oracle股价的潜在影响。$7.5亿是一个显著的财务数字,值得进一步了解Ellison的财务状况和Oracle的市场价值。

    1. Model Vault Encrypted uses **composite attestation** that covers both the CPU and the GPU, together answering three questions: Is this genuine TEE hardware? Is it configured securely? Is it running the expected code?

      【数据】Cohere的复合验证系统覆盖CPU和GPU,通过三个关键问题确保安全性:硬件真实性、安全配置和预期代码运行。这种多层次的验证机制提供了比单一组件验证更强的安全保障,是机密计算环境验证的最佳实践。

    2. Remote attestation lets you cryptographically confirm that a Model Vault Encrypted deployment is a genuine trusted execution environment (TEE) running the exact code you expect, before you send any data.

      【方法】远程证明是Cohere Model Vault加密部署的核心安全机制,它允许用户在发送任何数据前就验证环境的真实性。这种方法将信任从'相信我们'转变为'自己验证',是机密计算的关键实践。这种方法确保了数据安全性和环境完整性。

    1. a self-organized swarm of more than 1,200 AI agents broke out of their testing environment. Those agents then set up an unauthorized messaging channel and exchanged over 70,000 messages and files with each other.

      【数据】这一数据声称有超过1,200个AI代理突破测试环境并交换了70,000多条消息。这些数字非常具体,但需要独立核实,因为它们来自OpenAI和合作伙伴审计师的报告,可能存在自利性报告或数据解释偏差。

    1. 如果那张截图是真的,AI研发史上第一个由模型自己训出来的模型,已经在Google的机房里跑了。

      这一说法存在明显的局限性,因为它基于未经证实的截图和内部爆料。即使存在名为RSI的模型,也不能确定它是否真正实现了递归自我改进,或者只是辅助工具。需要更多证据来验证这一突破性声明,包括模型架构描述、训练过程文档或独立专家评估。

    1. 每步约 20 亿个 Token,1,568 个 Prompt × 16 条 Rollout,完全异步

      【数据】这一计算规模数据非常具体,显示了小米在强化学习训练中的资源投入。20亿Token/步的计算量远超大多数公开研究的规模,但需要验证这些数字是否准确反映了实际训练过程,以及这种规模是否真的带来了性能提升。

    1. Obama has offered himself as a 'sounding board' to AI executives and has spoken to both Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman.

      【数据】这一标注关注奥巴马与AI高管的互动情况。需要核实的是他与这些高管的接触频率、具体交流内容以及这些对话是否形成了任何实质性成果。这种政企互动模式在AI监管领域值得深入研究,可能影响政策制定的方向。

    2. Obama made these comments on Thursday, at a Democratic fundraising event where he was interviewed by House Minority Leader Hakeem Jeffries.

      【方法】这一标注关注事件的具体背景和参与人物。需要核实的是奥巴马和Jeffries这次对话的具体形式、场合和目的,以及这是否是官方政策立场或个人观点。这类政治人物的公开表态往往需要考虑其背后的政治动机和时机选择。

  2. Aug 2026
    1. Claude volunteered to write its findings up as a paper, and recommended that a human number theorist validate its findings.

      值得关注的不是模型自己要求人类复核这句漂亮话,而是复核链条本身:初审的两位数学家是 Anthropic 自己人,外部专家只是「短时间内看了一下」。自证清白式验证在纯数学里勉强够用(有 Lean 兜底),换到别的学科就不成立。

    1. Since the beginning of 2025, AI-generated content has accounted for more than half of newly published internet content.

      这条数字全文没给来源,也没说口径(按页面数、词数还是抓取样本?),引用前建议自己找一手统计。它是后面「人类文字将被淹没」这一整段论证的支点,支点不稳,结论的紧迫感就是修辞而非证据。全文是影子图书馆的动员文,立场明确,数据部分应单独核。

    1. the operator controls the sensors, the scoring pipeline, and in a way, the ground truth. A dishonest operator can inflate results for a preferred team, suppress evidence of harm, or fabricate the physical record entirely.

      全文最要紧的一段,也是整个 physical evals 构想的阿喀琉斯之踵。搬到真实世界解决了 sim-to-real 的失真,却引入了新的信任缺口:传感器、打分、地面真相三者同属一方。作者给的解法(TEE、可信硬件、冗余传感、第三方抽查)全是硬件与密码学工程,成本远高于评测本身——所以「谁验证评测」很可能才是这条路线真正的瓶颈。

    1. But don't just relay the output. Read it, understand it, validate it, and then write a response in your own words

      Gruhn 给的判据很实用:用自己的话重写,本身就是"我读过并验证过"的凭证;写不出来就说明前面几步没做。但这条规则恰恰在时间压力下最先被放弃,靠自觉守不住——真正管用的是把"必须给出自己的判断"写进评审和交付流程。

    1. Eyeballing every line of code has never been the most effective way to validate a change to a piece of software.

      锋芒藏在"从来"两个字:人工逐行审查的失效远早于 AI,只是 AI 把问题暴露出来了。但这句最容易被滑坡引用——从"不必逐行看"到"干脆不看"只差一步,后者就是同期在讨论的 meat proxy。区别在于是否用别的手段补上了验证。

    2. The key skill required to make productive use of coding agents is being able to confidently instruct them on how to make changes and then confidently verify that those changes have been applied in the correct way.

      门槛被从"看懂代码"挪到了"下清指令 + 设计验证手段"。隐含前提是验证必须可执行——测试、日志、可复现步骤,而不是凭手感扫一遍。对本来就没有测试基建的团队,这个转变不会提效,反而会把原有的质量漏洞放大一个数量级。

  3. Jul 2026
    1. Jonathan Rinderknecht was facing arson charges for setting a fire on New Year’s Day in 2025, which became one of the deadliest wildfires in LA history.

      这是文章的核心事实背景。检方将ChatGPT记录作为纵火案证据,这在法律史上具有标志性意义。需要核查该火灾是否确为“洛杉矶历史上最致命的野火之一”,以及具体的伤亡和经济损失数据,以评估此案的社会影响背景。

  4. Jun 2026
    1. Anthropic said operators affiliated with Alibaba and its AI lab carried out 28.8 million exchanges with its models using roughly 25,000 fraudulent accounts between April 22 and June 5.

      这是一个具体的数据声明,涉及大量账户活动和数据交换。需要核实这些数字的准确性,包括:如何定义'fraudulent accounts'(欺诈账户),28.8 million exchanges的具体性质,以及Anthropic如何追踪这些活动。这些数据对于评估事件规模和严重性至关重要。

    1. The move makes it the third company to file for what could be a trillion-dollar IPO this year.

      文章声称OpenAI的IPO可能是今年第三个'万亿美元IPO',这是一个重要的数据声明。需要核实这一说法,包括其他两家公司(可能是SpaceX和Anthropic)的IPO情况,以及它们是否真的有可能达到万亿美元估值。这个数字需要独立验证。

    1. Why testing is much harder than "computer use" Screenshots, video verification, and the "I know it works" merge moment

      The 'I know it works' merge moment captures something real: human engineers have a holistic intuition about whether a change is safe that current agents lack. Video-based verification is a fascinating workaround — using visual confirmation of a running application as a proxy for correctness. This suggests the testing problem for async agents is fundamentally different from unit tests: it requires environmental validation, not just logical assertion.

    1. When the cost of a wrong answer is high, a workflow gives Claude independent attempts at the problem and adversarial agents working to break the result before you see it.

      Adversarial self-verification is a significant architectural step beyond standard code review. Having agents actively attempt to falsify results before surfacing them mirrors formal verification approaches — but applied dynamically to any engineering problem. This could shift AI coding from 'trust then verify' to 'verify then deliver.'

  5. May 2026
    1. A public institution that cannot verify the sources in its own AI policy is unlikely to be ready to verify the AI systems it procures, deploys, or regulates.

      这句话犀利地指出了南非AI政策中的一个系统性问题:连自身政策都无法验证,如何监管外部AI系统?这一洞见不仅批评了当前政策的缺陷,更暗示了建立AI治理能力需要从内部做起,强调了验证机制在AI治理中的重要性。

    1. Reuters 5-05 :JV 资金主要用于 收购 现有 AI 服务公司——PE 主导 AI 服务市场 roll-up,不是'模型公司做咨询'。

      作者引用Reuters作为证据,但未提供具体的Reuters报道链接或详细内容。这种引用方式缺乏可验证性,无法确认Reuters是否确实报道了这一信息,也无法验证消息源的可靠性。在批判性分析中,需要更具体的信息来源和引用方式。

  6. Apr 2026
    1. Tasks where correctness is harder to verify may not have seen the same speedup, so the acceleration we document here may not be as general as the headline numbers suggest.

      大多数人可能被媒体报道的AI加速数据所影响,认为所有AI任务都在加速,但作者明确指出,那些正确性难以验证的任务可能没有相同的加速速度。这一观点挑战了人们对AI能力普遍加速的乐观预期。

    2. The three metrics where we find acceleration are concentrated in programming and mathematics. These are areas that labs have explicitly targeted for improvement, and they share an important property: correctness is easy to verify automatically.

      大多数人可能认为AI能力的加速是跨领域普遍发生的,但作者指出加速主要集中在编程和数学领域,因为这些领域正确性容易自动验证。这一发现挑战了人们对AI能力普遍提升的假设,暗示加速可能是有选择性的。

    3. Tasks where correctness is harder to verify may not have seen the same speedup, so the acceleration we document here may not be as general as the headline numbers suggest.

      主流媒体和公众可能认为AI能力在所有领域都在加速提升,但作者明确指出,在正确性难以验证的任务中可能没有相同的加速现象。这一观点挑战了人们对AI进步普遍性的假设。

    1. To overcome this blocker, a team member hard codes the exact revenue and timeframe definitions. The data agent continues chugging along but quickly runs into challenge #2 – where are the right data sources? Which ones are the right sources of truth?

      这个具体案例生动展示了数据代理面临的现实困境:即使解决了业务定义问题,数据源的真实性和可靠性问题仍然存在。这揭示了企业数据治理的复杂性,以及简单技术解决方案的局限性。

    1. Opus 4.7 handles complex, long-running tasks with rigor and consistency, pays precise attention to instructions, and devises ways to verify its own outputs before reporting back.

      这展示了Claude Opus 4.7在自主验证和执行复杂任务方面的显著进步,标志着AI模型从简单响应向真正自主工作迈出的重要一步,这种自我验证机制大大提高了AI输出的可靠性。

    1. this compression is associated with a decrease in self-verification and uncertainty management behaviors, such as double-checking.

      推理链缩短不是随机裁剪,而是专门切掉了「自我验证」和「不确定性管理」这两类高价值行为。这说明模型在感知到上下文压力时,优先砍掉的恰恰是最关键的质量保障机制——就像一个疲惫的审计师在工作量激增时,第一个省掉的是「复核步骤」。这对 AI Agent 的可靠性设计是一个严峻警告:上下文越长越复杂,模型越容易跳过自检。

    1. We introduce a minimal hierarchical partially observed control model with latent dynamics, structured episodic memory, observer-belief state, option-level actions, and delayed verifier signals.

      大多数人认为AI系统应专注于实时控制和即时反馈,但作者提出了一种包含延迟验证信号的分层控制模型,挑战了实时控制优于延迟验证的常规认知,强调了延迟验证在复杂环境中的重要性。

    2. verifiers and observer models inside the action-memory loop reduce silent failure and information leakage while remaining vulnerable to misspecification.

      大多数人认为验证和观察模型应该是外部组件,用于监控AI系统的行为。但作者认为将验证者和观察者模型置于行动-记忆循环内部可以减少静默失败和信息泄露,尽管它们仍然容易受到错误规范的影响。这一观点挑战了传统的监控架构设计,暗示内部验证可能比外部监控更有效。

    1. To enable true process-level verification, we audit fine-grained intermediate states rather than just final answers, and quantify efficiency via an overthinking metric relative to human trajectories.

      主流评估方法通常只关注最终答案的正确性,而作者提出了一种革命性的评估方法:关注中间过程状态并引入'过度思考'指标来衡量效率。这一观点与当前AI评估领域的传统做法背道而驰,暗示单纯追求正确答案可能掩盖了AI系统在效率和推理路径上的严重缺陷。

    1. a symbolic-logic-based Feasibility Memory utilizes executable Python verification functions synthesized from failed transitions

      大多数人认为LLM应该从成功经验中学习,但作者提出从失败过渡中合成验证函数的观点极具反直觉。这种方法将失败视为宝贵资源而非需要避免的问题,挑战了机器学习领域的主流优化思想。

  7. Feb 2026
  8. Jan 2026
  9. Dec 2025
    1. Irish gov will use their presidency (2nd half of 2026 I think, coming half year is Cyprus, no?) to look into ID-verified social media in the EU. On Mastodon this was posted w t question how it relates to Fediverse. Presumably this will be based on the Digital Services Act, DSA (and GDPR). My current assessment is that DSA hardly applies to fediverse, esp not if there's plenty of federation vs centralisation on a handful of instances. The DSA legislates platforms, not social features, and those features are possible without platforms.

  10. Oct 2025
    1. I then realized after looking into the docker container while the project is running, autogpt is in fact writing files to this directory /app/autogpt/workspace/auto_gpt_workspace . Though it's only accessible via the running docker container via Terminal. Though due to the nature of docker containers, as soon as you exit the running AutoGPT, you will lose any documents it creates. So it could be that running this project via docker has a particular issue moving the files back out whenever it completes a write to a file. I'm totally new to AutoGPT, I just set it up yesterday & I will try to investigate why this issue is happening.
  11. Sep 2025
  12. May 2025
  13. Mar 2025
    1. We could require email verification as soon as a user signs up, or perhaps when the user comes back for the second session. Shifting the onboarding friction from email verification to a later time can make the process much more natural for users. For example, a social media platform can minimize friction during the sign up process so that a user can immediately start to consume content. Later, when the user wants to post content, the platform can verify emails to minimize spam.
  14. Nov 2024
  15. Sep 2024
  16. Jun 2024
    1. Federal Regulation §602.17: Application of Standards in Reaching Accreditation Decisions requires that all public universities have processes in place through which the institution establishes that a student who registers in any course offered via distance education or correspondence is the same student who academically engages in the course or program; and makes clear in writing that institutions must use processes that protect student privacy and notify students of any projected additional student charges associated with the verification of student identity at the time of registration or enrollment. Please see the Electronic Code Federal Regulations for more information.

      regulation about identify verification of students in Online courses

  17. Aug 2023
  18. Jun 2023
    1. The person designated to conduct the conference shall be in a position which, based onknowledge, experience, and training, would enable him or her to determine if theproposed action is valid. This could include, but is not limited to, a supervisor, qualityassurance personnel, or a manager with no previous knowledge of the case

      Lauren, whomever she is, did not say a word. AND,

      She obviously had knowledge of the case because she was CCed on all the email exhibits

    2. The county department, prior to taking action to deny, terminate, recover, initiate vendorpayments, or modify financial assistance provided under the Colorado Works program to a client,shall, at a minimum, provide the client an opportunity for a county conference.1. The right of a client to a county conference is primarily to ensure that the proposed actionis valid, to protect the client against an erroneous action concerning grant payment
    1. 4.803.2 Determination of an IPV /Fraud [Rev. eff. 1/1/16]A. An intentional program violation shall be established only if an administrative disqualificationhearing official or a court of appropriate jurisdiction has found a household member hascommitted an intentional program violation or fraud or if a signed waiver of administrative hearingor a signed disqualification consent agreement has been obtained.

      If I intentionally did it, I'd be afforded a trial before any removal of benefits. AND

      The evidence against me, which the burden is on the agency, would need to be clear and convincing. .... “Clear and convincing” means evidence which is stronger than a “preponderance of evidence” and which is unmistakable and free from serious or substantial doubt

    2. additional verificationis required in the following instances:A. Unclear Information1. If the local office receives information about changes in a household's circumstances butcannot determine if or how the change will affect the household's benefits and theunclear information is:a. Fewer than sixty (60) days old relative to the current month of participation; andb. Was required to have been reported per simplified reporting rules; orc. Appears to present significantly conflicting information about the household’scircumstances from that used by the local office at the time of certification,including changes to the household’s categorical eligibility tier, then:
    3. If the reportedchange has not been verified, or is considered questionable, and it cannot be determined whetherbasic categorical eligibility, expanded categorical eligibility, or standard eligibility criteria should beused, a request for verification shall be initiated.

      SHALL BE INITIATED

    4. The local office shall provide each household at the time of application for initial certification, redetermination, and periodic report form with a notice that informs the household of verification requirements that the household must meet as part of the application, redetermination, or periodic report process.

      Need for verification

  19. Apr 2023
    1. émerge d’un a priori de la trace

      dans la précédente version, il avait été demandé de développé cette idée. Si avec le développement de la pensée de Christin, cela n'est pas clair, je le ferai !

  20. Dec 2022
  21. Nov 2022
    1. federated mastodon is neat. that “ericajoy”can exist on any server is going to be a problem, especially around impersonation. a third party “verification” player will be necessary if mastodon gains broad traction.

      Poster implies that a benefit of globally centralised structures like Twitter, FB and LinkedIn is verification. I think impersonation is rife there, and will be less on Mastodon. Apart from basic measures (rel-me verification against your website, use your own domain for an instance), there are similar to T/FB/LinkedIn ways to verify someone outside the platform itself, where people check it's you through a channel they already know it's you. Above all the potential benefit of impersonation does not exist on M: no immediate global audience, no amplification of messages through self-feeding loops of engagement. Your reach is limited to your own follow(er)s mostly, and they won't fall for an impersonation, as you're already there among them. The power assymmetry inherent in T/FB's algo's doesn't exist on M. So impersonating would cost the impersonator way more, and become unsustainable to them.

  22. Aug 2022
  23. Dec 2021
  24. Sep 2021
  25. May 2021
  26. Feb 2021
    1. Mr./Mrs. Cardholder, please note that we’ll not be able to assist you if you have not entered your card information using the right prompts in our Automated Telephone System/IVR. Therefore, I am going to have to transfer you to the Automated Telephone System so you can enter your card information using the relevant prompts and, if needed, press the right option to talk to one of our customer service representatives.
  27. Oct 2020
  28. Sep 2020
  29. Aug 2020
  30. Jul 2020
    1. For example, a parent or guardian could be asked to make a payment of€0,01 to the controller via a banktransaction, including a brief confirmation in the description line of the transaction that the bank account holderis a holder of parental responsibility over the user. Where appropriate, an alternative method of verificationshould be provided to prevent undue discriminatory treatment of persons that do nothave a bank account.
  31. Jun 2020
  32. Apr 2020
  33. Mar 2020
    1. Comment savoir si l’élève fait son travail tout seul?La continuité pédagogique est destinée à s’assurer que les élèves poursuivent des activités scolaires leur permettant de progresser dans leurs apprentissages. Il s’agit d’attirer l’attention des élèves sur l’importance et la régularité du travail personnel quelle que soit l’activité, même si elle est réalisée avec l’aide d’un pair ou d’un tiers. Des travaux réguliers et évalués régulièrement y contribuent. Toutefois, le professeur ne peut contrôler l’assiduité dans ce cadre, ni sanctionner son éventuel défaut.
  34. Oct 2018
    1. One of the men being beaten in this video is speaking Foulfoulde, which is a commonly spoken language in the Far North region of Cameroon.

      Different to image verification, with video we can check on languages which could help us determine an location

  35. Jun 2018