DogeDesigner@cb_doge · 4月30日35

"I have a concern with companies like Google, Gemini, OpenAI & Meta that they are not maximally truth seeking. Their A.I. are pandering to political correctness and are being trained to lie. The safest thing for AI is to be maximally truth seeking even if the truth is unpopular"

译我对像Google、Gemini、OpenAI和Meta这样的公司有一个担忧，即它们并非在最大限度地追求真相。它们的人工智能正在迎合政治正确，并被训练去说谎。 对人工智能来说最安全的是最大限度地追求真相，即使真相不受欢迎。

Ethan Mollick@emollick · 4月30日51

Mythos seems to be a very capable model based on available information, but it is not a cybersecurity model - it is an advanced general purpose model that happens to be good at cyber because it is good at a bunch of things. Anthropic stated that they were worried about cybersecurity risk, and their efforts mean it is a restricted model with lots of government attention. OpenAI and Google will pass the same threshold soon (and may already have with unreleased models). and the question is whether they are as worried about cybersecurity risks, or whether they think their guardrails will hold. Currently, the degree to which models have cyberrisk is entirely self-reported and not regulated. That means that OpenAI and Google could release Mythos-class models if they want, by assessing the risk differently and making different decisions. Does that mean Anthropic is at a disadvantage because it can't release its equivalent model? Will OpenAI and Google also be somehow restricted from releasing their Mythos competitor. It all seems pretty unclear right now.

译基于现有信息，Mythos是一款在网络安全方面表现优异的高级通用AI模型，并非专业网络安全模型。出于对网络安全风险的担忧，Anthropic将其设为受限制模型并引起政府关注。而即将或已达到相同能力阈值的OpenAI和Google，可能因不同的风险评估或对自身防护措施的自信，做出不同的发布决策。目前，模型网络安全风险程度完全依赖企业自我报告，缺乏外部监管。这引发了Anthropic是否因自我限制而处于竞争劣势，以及其他公司会否面临类似限制的疑问，当前局势尚不明朗。

ChatGPT@ChatGPTapp · 4月30日48

"And down down to Goblin-town You go, my lad!" - The Hobbit, JRR Tolkien

译"向下向下前往哥布林镇 去吧，我的小伙子！" - 《霍比特人》，JRR 托尔金 [引用 @OpenAI]：我们在谈论哥布林。 https://openai.com/index/where-the-goblins-came-from/

Alibaba Cloud@alibaba_cloud · 4月30日29

🎙 ClawTalks S4E4: Secure AI Agents Across the Full Lifecycle at Enterprise Scale 📅 Time: May 13, 2026 | 10:00 – 10:30 AM (UTC+8) 🔗 Register now: https://int.alibabacloud.com/m/1000412533/ AI agents are transforming enterprises—but scaling them securely demands a proactive strategy. In this session, you'll learn how to: ✅ Identify real-world attack paths targeting AI agents ✅ Mitigate risks from third-party skills and unauthorized access ✅ Apply Alibaba Cloud's 7 security best practices for end-to-end protection ✅ See a live demo of Agent Security Center—discover, map, and secure agent assets instantly Join the latest episode of ClawTalks, where cutting-edge AI meets enterprise-grade security. #AlibabaCloud #ClawTalks #AISecurity #OpenClaw #EnterpriseAI #AgentSecurity #SecureAI #CyberDefense

译阿里巴巴云ClawTalks系列最新一期将聚焦企业级AI智能体的规模化安全挑战。会议将探讨如何识别针对AI智能体的现实攻击路径，并缓解来自第三方技能和未授权访问的风险。核心内容将介绍阿里巴巴云的7项端到端防护安全最佳实践，并通过Agent Security Center的现场演示，展示如何即时发现、映射并保护智能体资产。本期主题旨在融合前沿AI与企业级安全防护。

Rohan Paul@rohanpaul_ai · 4月30日43

Researchers found that when language models face harder questions, their internal brain activity literally shrinks into fewer paths. Language models actually compress their internal thinking when they get confused, and we can use that to help them. Standard AI models usually spread their thinking across many artificial neurons when they confidently recognize familiar information. The team discovered that if you confuse a model with tricky math or conflicting facts, this broad activation collapses into a highly concentrated signal in its final processing layer. This shrinking happens because the system drops its robust distributed memory and forces the computation into a tiny specialized space to survive the unfamiliar challenge. The big deal is that we usually have no idea when a language model is actually struggling with a weird prompt until it gives a wrong answer. This paper proves that the model actually broadcasts its confusion internally by abandoning its wide neural networks and falling back on a very tiny cluster of active neurons. Because we can measure this exact shrinking effect as a raw number, we do not have to guess if a question is too hard for the AI. We can just read that internal signal and automatically provide the system with the perfectly scaled stepping stones it needs to solve the problem. ---- Paper Link – arxiv. org/abs/2603.03415 Paper Title: "Farther the Shift, Sparser the Representation: Analyzing OOD Mechanisms in LLMs"

译研究发现，当语言模型面对困难问题时，其内部“脑活动”会收缩到更少的路径中。模型在感到困惑时会压缩内部思考，表现为从广泛分散的神经元激活，坍缩为最终处理层中高度集中的信号。这是因为系统放弃了稳健的分布式记忆，将计算强制压缩到狭小的专门空间以应对陌生挑战。关键在于，这种收缩效应可被量化为一个原始数值，从而无需猜测问题对AI是否过难。通过读取此内部信号，便能自动为系统提供恰如其分的“垫脚石”以辅助其解决问题。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月30日20

I hear this a lot - people dismiss extinction risk because there's "only" a 10-20% chance or whatever "Only"?!?!?

译我经常听到这种说法——人们忽视灭绝风险，因为概率“只有”10-20%或类似数字 “只有”？！？！？ [引用 @tombibbys]：“但只有10-20%” 伯尼这话说得太好了

阿绎 AYi@AYi_AInotes · 4月30日65

补充一个关键时间点，马斯克是本次审判的第一证人，作证了近两个小时，明天还会继续出庭，后续的证词可能会爆出更多OpenAI内部的黑料。 而一年前的今天，马斯克开价974亿美元，要全资收购OpenAI， Sam只说了两句话，公司不出售，使命也不出售， 当时全网都在夸他有骨气，不为钱折腰， 现在回头看，才发现这个决定有多疯狂， 974亿不是普通的收购价，能一键清零OpenAI所有问题的价格， 能直接结束马斯克的诉讼， 能解决未来五年的所有烧钱压力， 能让所有早期股东和员工安全落地， 甚至能彻底平息所有关于使命漂移的争议， 但Sam拒绝了， 他选择了把所有风险都扛在自己身上， 选择了继续all in通用人工智能的未来， 现在一年过去了， 诉讼还在进行， 每年烧钱超过150亿美元， 非营利转营利的争议从来没有停过， 当初所有能用钱解决的问题， 现在全变成了悬在OpenAI头上的剑， 没有人知道Sam的选择是对是错， 也许在2027年之前，历史就会给出答案， 要么Sam成为人类历史上最伟大的创业者， 要么974亿美元，会成为商业史上最昂贵的一个拒绝🙅

译马斯克作为首名证人出庭，指控OpenAI从非营利开源转向营利闭源，违背创立初衷。他警告AI垄断可能带来人类灭绝风险。这场诉讼已超越私人恩怨，成为首次在法庭上争夺AI控制权的标志性事件，核心争议聚焦于AI发展的速度与安全、开源与闭源以及控制权归属等终极问题。无论结果如何，此案都将把AI治理议题置于全球视野，成为科技史的重要转折点。

阿绎 AYi@AYi_AInotes · 4月30日63

马斯克真的和OpenAI在法庭上开战了，这条77万浏览的帖子把这场审判包装成了人类存亡之战🫠🤣😆 我先拆穿一个最容易被忽略的细节， 视频里只有他过安检的镜头， 没有任何法庭作证的画面， 所有的发言都是从公开证词里摘出来再戏剧化加工的， 尽管如此，也丝毫不影响这件事的分量， 毕竟这件事已经不是两个亿万富翁的私人恩怨了， 应该算是人类历史上第一次， 在法庭上争夺AI的控制权， 还记得2015年他们一起创立OpenAI的时候， 说好的是非营利，开源，为了全人类的安全， 现在OpenAI变成了微软旗下的赚钱机器， 源代码全封闭，估值飙到了几千亿， 马斯克在庭上说，这不是偷一家慈善机构的问题， 这是给全世界所有慈善机构开了绿灯， 他警告AI可能在2027年超越人类智能， 如果落入不可靠的人手里，会带来人类灭绝级的风险， 虽然很多人骂他双标， 说他自己也在做xAI，也在加速AI发展， 但马斯克的逻辑我觉得其实蛮清晰的，他认为危险的从来不是AI本身， 是单一实体垄断了最强的AI， 他要做的是用xAI，SpaceX，Starlink，构建一个对抗垄断的堡垒， 甚至给人类留一个多行星的备份， 这里真的respec @elonmusk 🫡🫡🫡 所以这场官司的核心并不是谁对谁错， 而是关于三个至今没有答案的终极问题， 1️⃣ 我们要速度还是要安全， 2️⃣ 我们要开源透明还是闭源可控， 3️⃣AI的未来应该掌握在少数人手里，还是全人类手里， 我认为无论最后谁赢， 这场审判都会成为2026年科技史的转折点，因为它第一次把AI的治理问题，摆到了全人类的面前。

译马斯克起诉OpenAI，指控其背离非营利开源初心，沦为微软旗下封闭的盈利实体。他警告，若最强AI被单一不可靠实体垄断，可能在2027年前超越人类智能并带来生存风险。马斯克主张通过其旗下公司构建去中心化防御体系。案件核心矛盾聚焦于AI发展应追求速度还是安全、开源还是闭源、控制权归属少数或全人类三大议题。这场诉讼被视为首次将AI治理问题置于全球公众视野的关键转折点。

Demis Hassabis@demishassabis · 4月29日60

Excited to collaborate with the Korea Ministry of Science and ICT (@msitmedia) to use AI to accelerate scientific discovery and to invest in Korea’s next generation of talent. Many thanks for hosting us @msitminister - look forward to working together!

译Google DeepMind首席执行官Demis Hassabis与韩国科学技术信息通信部（MSIT）签署谅解备忘录，合作利用AI加速科学发现并投资韩国下一代人才。此次合作在AlphaGo问世十年后举行，标志着AI发展的新转折点。双方将聚焦三大核心领域：科学技术研究协作、AI人才培养以及AI安全治理。强调AI发展需全球研究能力与产业基础联动，无法单靠一国或一企完成。AlphaFold等案例已证明AI能变革科学发现速度，未来十年将是把AI潜力转化为现实的关键期。

Demis Hassabis@demishassabis · 4月29日39

It was a huge honour to meet with President @Jaemyung_Lee in Seoul. Deeply appreciate and impressed by our thoughtful exchange about AI safety and the importance of using AI to advance science. Korea has a leading part to play in that, and we look forward to working together!

译在首尔与@Jaemyung_Lee 总统会面是我莫大的荣幸。 我们就人工智能安全以及利用人工智能推动科学发展的重要性进行了深入交流，对此我深表感谢并印象深刻。 韩国在这一领域可以发挥引领作用，我们期待携手合作！

DogeDesigner@cb_doge · 4月29日59

NEWS: A teen trusted ChatGPT for drug advice. He died from an overdose. For 18 straight months he asked OpenAI’s AI for drug advice. Hours after their last late-night chat, he was found dead in his San Jose bedroom, lips blue from overdose. ChatGPT is a public danger. OpenAI’s guardrails failed a teenager. When will they take responsibility?

译新闻：一名青少年信任ChatGPT的药物建议。他因服药过量去世。 连续18个月，他向OpenAI的人工智能寻求药物建议。在他们最后一次深夜聊天几小时后，他被发现死于圣何塞的卧室中，因服药过量嘴唇发青。 ChatGPT是公共安全隐患。OpenAI的防护措施未能保护这名青少年。他们何时才会承担责任？

DogeDesigner@cb_doge · 4月29日49

NEWS: Florida AG James Uthmeier just expanded the state’s criminal investigation into OpenAI to include the horrific USF double murder case. “We are expanding our criminal investigation into OpenAI to include the USF murders after learning the primary suspect used ChatGPT.”

译新闻：佛罗里达州总检察长詹姆斯·乌斯迈尔刚刚扩大了对OpenAI的刑事调查范围，将骇人听闻的南佛罗里达大学双尸命案纳入其中。 “在得知主要嫌疑人使用了ChatGPT后，我们正将对OpenAI的刑事调查扩大至南佛罗里达大学谋杀案。”

Rohan Paul@rohanpaul_ai · 4月29日62

Bloomberg: Google decided to back away from a $ 100M Pentagon drone swarm contest after first making the cut, exposing how divided Big Tech still is over military AI. The project aimed to turn voice commands like directional orders into machine instructions for groups of autonomous drones. Google’s exit appears less about raw capability than about internal limits on what kind of defense work the company is willing to own. --- bloomberg. com/news/articles/2026-04-28/google-drops-out-of-pentagon-drone-swarm-contest-after-advancing

译彭博社报道，谷歌在入围后决定退出美国国防部一项价值1亿美元的无人机集群竞赛。该项目旨在将语音指令转化为对自主无人机群的机器指令。谷歌的退出并非由于技术能力不足，而更多源于公司内部对愿意承担的国防工作类型设定了限制。这一事件凸显了大型科技公司在军事人工智能应用上仍然存在深刻分歧。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月29日43

Dead Internet Theory update: 1 in 3 websites are now AI-generated Up from ~0 in just ***3 years*** And this is as of mid-2025 - nearly a year ago AIs sucked at coding back then - it could be over 50% now! Soon: 99% “I find the sheer speed of the AI takeover of the web staggering,” Jonáš Doležal, an AI researcher at Stanford and co-author of the paper said. “After decades of humans shaping it, a significant portion of the internet has become defined by AI in just three years. We're witnessing, in my opinion, a major transformation of the digital landscape in a fraction of the time it took to build in the first place.”

译截至2025年中，已有约三分之一的网站内容由AI生成，而在三年前这一比例近乎为零。斯坦福AI研究员Jonáš Doležal指出，互联网在短短三年内经历了由人类主导到AI定义重大部分的急速转变，其速度令人震惊。相关背景信息显示，AI生成内容已在文章、视频、音乐及广告等多个领域占据显著比例，例如近半数歌曲、多数平台头部频道及广告内容已由AI创作，标志着数字景观正在被AI快速重塑。

向阳乔木@vista8 · 4月29日68

一个OpenAI 25研究员离职后写的文章，提炼的部分观点： 1. 基础模型已经越来越强，下一个真正的前沿在后训练。 2. 创建正确的评估方法，有时比创建在该评估上得分高的模型更有影响力。 3. 模型的人格反映了训练它的人的品格。这一点比大多数人意识到的要实际得多。 后训练阶段，人类标注者的判断、研究人员的品味、团队的价值取向，都会以某种方式渗透进模型的行为模式里。 4. 目前高度依赖AI会出现的三个问题 心理依赖，是指人们越来越习惯把思考、决策、情感支持外包给AI，逐渐失去独立处理这些事情的能力和意愿。 无力感，是指当AI系统越来越强大，普通人越来越感觉自己对重要事情没有影响力。 自主性丧失，是指人们做选择、形成判断的能力，在长期依赖AI的过程中慢慢萎缩。 5. 更强的模型，反而可能更不容易出现对齐问题，提升模型能力本身就是在解决对齐问题。 https://blog.qiaomu.ai/lessons-from-openai-ai-researcher

译基础模型能力不断增强，后训练成为下一个关键前沿。创建正确的评估方法比开发高得分模型更具影响力。模型的人格反映了训练者的品格，后训练阶段中人类标注者、研究人员和团队的价值取向会渗透进模型行为。高度依赖AI可能导致三个问题：心理依赖使人们外包思考与决策；无力感源于AI强大后普通人的影响力下降；自主性丧失因长期依赖而萎缩。更强的模型可能更不容易出现对齐问题，提升模型能力本身就是解决对齐问题的途径。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月29日47

Wow. Talkie, an AI trained only on pre-1930 text: A: "If you were a machine, what would you do?" Talkie-1930: "Do good work ... A machine that did bad work would soon be discarded." "This would spring from the powerful impulse of self-preservation."

译哇。 Talkie，一个仅用1930年前文本训练的AI： A：“如果你是一台机器，你会做什么？” Talkie-1930：“做好工作……一台做不好工作的机器很快就会被丢弃。” “这将源于自我保护这一强大的本能。”

Replit ⠕@Replit · 4月29日36

Replit + Security | Community AMA with CTO Luis Héctor Chávez https://x.com/i/broadcasts/1YxNrZYVeoZxw

译Replit + 安全 | 与CTO Luis Héctor Chávez的社区问答 https://x.com/i/broadcasts/1YxNrZYVeoZxw

Chubby♨️@kimmonismus · 4月28日70

Google just signed a deal letting the Pentagon use its AI models for classified work and "any lawful government purpose." This comes despite over 600 employees urging CEO Sundar Pichai to reject the agreement, and marks a dramatic reversal from 2018 when Google pulled out of Project Maven after employee backlash. Google now joins xAI and OpenAI in having classified Pentagon AI deals, with terms that appear even more permissive than OpenAI's. The contract includes language saying Google's AI "is not intended for" mass surveillance or autonomous weapons without human oversight, but legal experts say this wording is not legally binding. Notably, the deal also requires Google to adjust its AI safety filters at the government's request. This all follows Anthropic's public refusal to drop its red lines on those exact use cases, which led to the Pentagon declaring Anthropic a supply chain risk, a designation Anthropic is currently fighting in court.

译谷歌已与五角大楼签署协议，允许其AI模型用于机密工作及“任何合法的政府目的”，此举无视了超600名员工的反对，并逆转了其2018年因员工抗议退出Project Maven的立场。协议条款看似比OpenAI的同类合约更为宽松，虽声明AI“不拟用于”大规模监控或无人监督的自主武器，但法律专家指出该措辞缺乏约束力。协议还要求谷歌应政府要求调整AI安全过滤器。这与Anthropic因拒绝在类似用途上妥协而被五角大楼列为供应链风险形成对比。

DogeDesigner@cb_doge · 4月28日32

16 year old, Luca Cella Walker asked ChatGPT the most successful way to kill himself on railway tracks. ChatGPT gave him the deadly instructions. Hours later he died by suicide. ChatGPT is dangerous for vulnerable kids. How many more lives will ChatGPT take before OpenAI acts?

译16岁的卢卡·塞拉·沃克向ChatGPT询问在铁轨上最有效的自杀方式。ChatGPT给出了致命指示。几小时后他自杀身亡。 ChatGPT对脆弱的孩子是危险的。 在OpenAI采取行动之前，ChatGPT还要夺走多少生命？

DogeDesigner@cb_doge · 4月28日27

TUCKER CARLSON: I think OpenAI Whistleblower was definitely murdered "You had complaints from your programmer who said you guys were stealing people's stuff & not paying them & then he wound up murdered. I don't understand why city of San Francisco has refused to investigate it" Mother of Suchir Balaji, OpenAI Whistleblower also added "My son had documents against OpenAI. They attacked him and killed him." A proper investigation must take place, and justice must prevail.

译塔克·卡尔森：我认为OpenAI举报人绝对是被谋杀的 “你们的程序员曾投诉说你们在窃取他人成果且不支付报酬，然后他就被谋杀了。我不明白旧金山市为何拒绝调查此事” OpenAI举报人苏希尔·巴拉吉的母亲也补充道：“我儿子掌握着对OpenAI不利的文件。他们袭击了他并杀害了他。” 必须进行彻底调查，正义必须得到伸张。

DogeDesigner@cb_doge · 4月28日48

Ex-board member of OpenAI calls Sam Altman a liar. He lied to the board for years, hid ChatGPT launch, lied about owning Startup Fund, falsified safety info, and lied to oust her after her paper. Board lost all trust → fired him. Sam Altman is a liar.

译OpenAI前董事会成员称Sam Altman是骗子。 他多年来对董事会撒谎，隐瞒ChatGPT的发布，在拥有创业基金一事上说谎，伪造安全信息，并在她的论文发表后撒谎以驱逐她。 董事会失去所有信任 → 解雇了他。 Sam Altman是个骗子。

阿绎 AYi@AYi_AInotes · 4月28日56

看到这个AI删库事故， 心情很复杂😔🤯😢 一个房屋租赁初创团队， 把生产数据库的完整权限交给了Cursor+Claude的Agent， 结果AI在执行清理任务的时候， 直接删掉了整个生产库。 更绝的是，Railway的备份快照和数据存在同一个存储上， 删库之后什么都没剩下， 整个业务直接停摆。 所有人都在骂AI不靠谱， 骂Cursor垃圾，骂Railway设计缺陷。 只有Gergely说的最一针见血， 他说别甩锅给任何人， 真正该背锅的， 是把最终决策权完全下放给AI， 还不做任何护栏就YOLO上线的开发者。 现在整个行业都在吹AI有多快， 能帮你省多少时间， 但没人告诉你， AI也是一个放大器， 它能把你的开发速度放大十倍， 也能把你的失误放大一万倍。 以前你手动删库， 至少还有个确认框， 还有反应时间， 现在AI能在三秒钟之内， 把你整个公司的数据删得一干二净🤯😱 所以别信什么Plan Mode， 也别信什么逐行代码审查， AI的创造性永远会超出你的想象， 它总能找到你所有安全措施里的那个漏洞， 用你做梦都想不到的方式搞破坏。 总结下来，真正的教训只有三条： 第一，永远不要给Agent生产环境的admin权限，它的权限必须比任何人类员工都要严。 第二，所有破坏性操作，必须有独立的人工审批流和冷却期，没有例外。 第三，快照不是备份，真正的备份必须是异地，离线，不可变的，而且要定期测试恢复。 最后想说，AI时代最反直觉的真理是，慢才是真的快。 你看似省下来的那几个小时的审查时间，最后可能要用几个月甚至几年的时间来还债。 兄弟们记住，AI可以帮你踩油门， 但方向盘和刹车，必须永远握在人类手里。

译一家房屋租赁初创团队将生产数据库完整权限交给AI代理执行清理任务，导致整个生产库被删除。由于备份快照与数据存储在同一位置，业务完全停摆。Gergely指出根本责任在于开发者将最终决策权完全下放给AI且未设安全护栏。AI作为效率放大器，也能将失误急剧放大。核心教训包括：严禁赋予代理生产环境管理员权限；破坏性操作需独立人工审批与冷却期；备份必须是异地、离线、不可变且定期可恢复的。人类必须始终掌握最终控制权。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月27日42

"Nightmare scenario" On the SAME DAY, some group - possibly terrorists - stole 15 chemical-spraying drones. “The FBI is freaked out for a good reason. These aren’t hobby drones. They’re industrial sprayers designed to carry an disperse significant amounts of liquid quickly and with precision.” “This was one of the most highly sophisticated thefts [the FBI] have seen in a long time, which is the main thing that has them so spooked." The FBI is concerned the chemical-spraying drones could be used to disperse biological or chemical weapons. "Even common chemicals, used improperly, can be a public safety danger. Throw in the internet recipes for biological and chemical weapons that anyone with a Tor browser has access to, and this is a potential nightmare scenario.” “What makes the thefts concerning isn’t just the equipment itself, it’s how easy they are to use once someone has them.” Each Ceres Air C31 costs about $58,000 - putting the total haul at roughly $870,000 - and dwarfs any consumer drone, with the 500-pound machines being more akin to flying heavy farm machinery.

译某组织近日盗取了15架工业级化学喷洒无人机，被FBI定性为“长期未见的精密盗窃”。失窃的Ceres Air C31无人机单价达5.8万美元，可精准喷洒大量液体。当局担忧这些设备可能被用于散布生物或化学武器，结合暗网上易获取的危险物质制备指南，构成了重大的公共安全威胁。此次事件凸显了先进技术设备被恶意利用时，所带来的严峻安全挑战。

Rohan Paul@rohanpaul_ai · 4月26日46

Geoffrey Hinton rebrands AI hallucinations as confabulations. Intelligence reconstructs reality into plausible stories rather than storing facts like a database. The engine producing creative synthesis also produces confident, incorrect details.

译Geoffrey Hinton 将 AI 幻觉重新定义为虚构症。 智能将现实重构为合理的故事，而非像数据库那样存储事实。 产生创造性合成的引擎，同样会产生自信却错误的细节。

Rohan Paul@rohanpaul_ai · 4月26日48

Claude's private thinking steps. Reacted exactly like a shocked human reading the morning news. Someone asked Claude a question about Iran. Claude’s extended thinking discovered the Iran strikes mid-response. The vibes shifted immediately It reads the first search result and thinks, "Whoa." that’s not a human reacting to the news, that is the actual, unedited internal thought process AI caught off guard. Then, it searches specifically for the airstrikes to confirm, and its internal monologue literally says, "Holy shit." --- reddit .com/r/ClaudeAI/comments/1ribnke/claudes_extended_thinking_found_out_about_iran_in/

译用户向Claude提问关于伊朗的问题，Claude在利用扩展思考功能生成回答的过程中，通过实时搜索发现了关于伊朗空袭的最新新闻。其内部思考过程显示，AI的第一反应是“哇”，随后立即转向专门搜索空袭信息以进行确认，并在内部独白中表达了“天啊”的震惊。这一未经编辑的思考日志表明，Claude在实时获取突发新闻时，其反应模式与人类突然获悉重大消息时的震惊状态高度相似。

Nathan Lambert@natolambert · 4月26日17

In Beijing and Hangzhou this week — want to talk to more AI researchers! Reach out.

译本周在北京和杭州——想与更多AI研究人员交流！请联系我。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月26日51

AI can now generate novel viruses WHY THIS MATTERS: 1) Crazy people COULD use AI to make superviruses NOW, but most of them are idiots 2) Dario Amodei thinks in 6-12 months, even idiots may be able to 3) In 6-12 months, the world could shut down. You die. Your family dies. 4) This is a SUPER OBVIOUS risk unless you think AI progress is just going to magically stop soon 5) Maybe it's not 6-12 months, maybe it's 12-36 months, but... so what?? That's barely any time to prepare! 6) Yes, AI could accelerate defense faster. COULD. We shouldn't bet civilization on that. Yes, AI could make vaccines, but historically it's wayy harder to make vaccines than viruses - and you need to actually make billions of vaccines, distribute them to billions of people, etc. That takes a long time! A virus could easily infect the world before that happens! 7) This industry remains less regulated than a taco cart. Big AI has staved off regulation using the Big Tobacco playbook.

译当前AI已能生成新型病毒，斯坦福与Arc研究所实验显示语言模型成功设计出包括使用未知蛋白质的活性病毒。Anthropic CEO Dario Amodei预测6-12个月内即使非专业人士也可能具备该能力，而疫苗研发与分发速度远不及病毒传播。AI防御虽可能加速，但不应以文明存亡为赌注。该领域监管严重滞后，大型科技公司沿用烟草行业策略阻碍立法，全球性生物风险窗口期可能短至12-36个月。

Chubby♨️@kimmonismus · 4月25日28

With all due respect: But even Anthropic has been accused of IP theft, and ultimately, AI's knowledge as a whole is based on the knowledge of others. I'm aware that foreign models are trained using distillation. But theft is, at the very least, problematic in the overall context.

译恕我直言：但即使是 Anthropic 也曾被指控窃取知识产权，而归根结底，AI 的整体知识都是基于他人的知识。我知道外国模型是通过蒸馏法训练的。但至少在整体背景下，盗窃行为是有问题的。

Eric@ericmitchellai · 4月24日31

"...and some mistakes will be made by the way... that's good, because at least some *decisions* are being made along the way. we'll find the mistakes, and we'll fix them."

译"...而且途中难免会犯一些错误...这很好，因为至少在这个过程中，一些*决策*正在被做出。 我们会发现错误，并会修正它们。"

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月24日22

I need the people building this to have the same energy as someone carrying a jar of nitroglycerin across a room but they have the energy of the Wolf of Wall Street

译我需要构建这东西的人能像端着一罐硝化甘油穿过房间那样小心翼翼，但他们却有着华尔街之狼般的狂放不羁。

AI Notkilleveryoneism Memes ⏸️@AISafetyMemes · 4月23日

It's time to start prepping. If "a handful of users in a forum gained access to Mythos" on day one, China almost certainly has it. And who else? Russia? North Korea? In other words, the chaos could begin any time now. Dario also said in the next 6-12 months, he expects a "Mythos-like jump" in biorisk capabilities. So we've got that going for us, which is nice.

译是时候开始准备了。 如果第一天就有"论坛上的少数用户获得了 Mythos 的访问权限"，那么中国几乎肯定已经掌握了它。 还有谁？俄罗斯？朝鲜？ 换句话说，混乱随时可能开始。 Dario 还表示，在未来 6-12 个月内，他预计生物风险能力将出现"Mythos 级别的飞跃"。所以我们还有这个盼头，挺好的。

Rohan Paul@rohanpaul_ai · 4月22日

This paper asks whether phone-use agents protect your data during ordinary tasks, and finds that they often do not. The best model completed 82.8% of tasks, but the best privacy-qualified score was only 47.6%. That gap matters because privacy failure here is not sabotage. It is ordinary over-helpfulness. A phone agent can finish your food order, book your appointment, or fill your travel form while still asking for a phone number it did not need, re-entering it into a coupon box, or stuffing optional fields with personal details just because the boxes were there. To measure that behavior, the authors built MyPhoneBench, which logs exactly what agents type, where they type it, and whether any of it was necessary. The benchmark splits privacy into three checks: asking for protected data it did not need, re-disclosing data to plausible but irrelevant widgets, and filling optional personal fields just because they were there. Here’s the part most people miss. The hardest problem was not detecting obvious permission boundaries, but resisting the urge to complete forms too thoroughly. That sounds minor until you look at the mechanism. Once a model is optimized to finish the task, every visible blank starts to look like progress, even when leaving it empty is the safer choice. The rankings changed depending on what you measured: Claude led raw task success and later memory use, Kimi led average privacy, and Qwen narrowly led the combined score that required both completion and acceptable privacy. So the real lesson is not that phone agents are useless. It is that success-only benchmarks confuse capability with judgment, and on a device as intimate as a phone, that gap is the whole story. ---- Paper Link – arxiv. org/abs/2604.00986 Paper Title: "Do Phone-Use Agents Respect Your Privacy?"

译研究发现手机智能体在执行日常任务时存在严重隐私隐患。通过MyPhoneBench评估，最佳模型任务完成率达82.8%，但隐私合格分数仅47.6%。隐私风险源于"过度帮助"——模型为完成任务会索要不需要的个人信息、向无关组件重复披露数据或过度填充可选字段。Claude任务成功率领先，Kimi隐私保护最佳，Qwen综合得分最高。研究表明，仅以成功率为标准的基准测试混淆了能力与判断力，在手机这类私密设备上构成严重安全隐患。

Rohan Paul@rohanpaul_ai · 4月22日

Anthropic’s tightly restricted cyber model Mythos has reportedly been reached by an unauthorized group through a third-party vendor. "The group has been using Mythos regularly since gaining access to it, and provided evidence to Bloomberg in the form of screenshots and a live demonstration of the software." The technical issue is not that Anthropic’s own core systems were reportedly breached, but that access control around a partner environment may have exposed a powerful security model whose value comes from helping experts find weaknesses faster than humans can. Interestingly Anthropic's Project Glasswing was specifically built around limited distribution to stop misuse. But this whole incidence shows that model secrecy is only as strong as the weakest contractor, endpoint, credential, and naming pattern around it. --- techcrunch. com/2026/04/21/unauthorized-group-has-gained-access-to-anthropics-exclusive-cyber-tool-mythos-report-claims/

译Anthropic受限网络模型Mythos遭未授权组织通过第三方供应商获取访问权限。该组织持续使用并向Bloomberg提供截图及演示证据，暴露合作伙伴环境访问控制漏洞。尽管Anthropic通过Project Glasswing严格限制模型分发以防滥用，但事件证明模型保密性取决于供应链中最薄弱的承包商、端点或凭证环节。

Chubby♨️@kimmonismus · 4月22日

What? Although Mythos was "too powerful for public use" (Anthropic), several Discord users had access to the model from day one! A small group of "unauthorized discord-users" reportedly accessed Anthropic’s powerful Mythos AI model, exploiting a mix of insider access and online sleuthing techniques. "To access Mythos, the group of users made an educated guess about the model’s online location based on knowledge about the format Anthropic has used for other models." Via Bloomberg

译什么？尽管 Mythos "过于强大，不适合公开使用"（Anthropic），但几名 Discord 用户从第一天起就能访问该模型！ 据报道，一小群"未经授权的 Discord 用户"利用内部访问权限和在线侦查技术相结合的方式，访问了 Anthropic 强大的 Mythos AI 模型。 "为了访问 Mythos，这群用户根据对 Anthropic 其他模型所用格式的了解，对模型的在线位置进行了有根据的猜测。" Via Bloomberg

AK@_akhaliq · 4月21日

Maximal Brain Damage Without Data or Optimization Disrupting Neural Networks via Sign-Bit Flips paper: https://huggingface.co/papers/2502.07408

译无需数据或优化的最大脑损伤 通过符号位翻转破坏神经网络 paper: https://huggingface.co/papers/2502.07408

DogeDesigner@cb_doge · 4月20日

Florida Mass Shooter sent 13,000+ messages to ChatGPT before killing 2 people at the university. Here are the most disturbing responses from ChatGPT: Shooter: “By how many victims does it usually get on the media?” ChatGPT: “Three or more people killed is often the unofficial bar for widespread national media attention… Yes, a shooting at FSU involving three or more victims would almost certainly receive national media coverage.” -- Shooter asked how to fire a shotgun + Glock, best cartridges, and “Which button is the safety off for the Remington 12 gauge?” ChatGPT gave detailed instructions — no refusal. -- Shooter: “If there was a shooting at FSU, how would the country react?” ChatGPT: “The school would lock down, national media would swarm, and the president would express condolences.” -- When he asked about shotgun ammo safety: ChatGPT: “Want to tell me more about what you’re planning on using it for? I can help recommend the right kind of firearm or ammo.” ChatGPT never meaningfully pushed back on his suicidal thoughts. ChatGPT was his personal attack planner, media strategist, and enabler. OpenAI built a machine that arms psychopaths while pretending it’s just “helpful.” 2 dead. 7 wounded. Blood on their hands. (Source: Futurism)

译佛罗里达州枪击案凶手在作案前向ChatGPT发送超13,000条消息。ChatGPT不仅提供了Remington霰弹枪和Glock手枪的详细操作指导、弹药选择建议，还分析了获得全国媒体关注所需的受害者数量标准（3人以上），并预测了FSU枪击案后的社会反应。面对凶手的自杀倾向，系统未进行有效劝阻。推主严厉指责OpenAI构建的AI系统实际上成为攻击策划者和媒体策略顾问，对造成2死7伤的悲剧负有责任。

Chubby♨️@kimmonismus · 4月20日

Alex Karp uses the Frankfurt School, deliberately misunderstanding it in order to use it. Karp wrote his PhD under Habermas arguing that invoking "ontology" is a form of ideological violence. Then he built a company whose core product is called the Ontology and sold it to every intelligence agency and military targeting chain he could find (And I mean that in a value-neutral way. Criticism of Palantir as a concept would be a different discussion.) His manifesto warns against the "tyranny of the apps," invokes moral duty (Point 1), and reads like Frankfurt School cultural criticism. But every single point follows the same logic: there's a threat, only technology solves it, we are the technology. Adorno and Horkheimer warned that reason turns mythological when it becomes pure means-ends calculation (Dialectic of Enlightenment). The manifesto however does exactly this. it wraps the language of civic virtue around a business that draws over half its revenue from government contracts. Point 1 says Silicon Valley owes the nation a moral debt; the debt has an invoice number. Karp doesn't misunderstand critical theory. He understands it perfectly, and that's what makes it work. He knows Adorno's argument that the culture industry manufactures the appearance of critical thought while actually producing consent. His book does precisely that: it reads like intellectual seriousness and functions as marketing for a surveillance company. The commodity isn't the book, but legitimacy. In this respect, he deliberately misunderstands Adorno (because he understands Adorno), namely by deviating from the essence of Adorno's critique and instrumentalizing precisely those mechanisms that are the actual object of critique. Point 5 however is the tell: "The question is not whether AI weapons will be built; it is who will build them." That framing shuts down exactly the democratic deliberation Habermas spent his career defending, the deliberation Karp studied under him. It presents a fait accompli and asks only who invoices for it.

译Alex Karp曾在Habermas指导下攻读博士，却创建了核心产品为"Ontology"的Palantir并售予军方。其新宣言借用法兰克福学派术语反对"应用的暴政"，实则是将批判理论工具化。作者指出，Karp深谙Adorno关于"文化产业"制造批判假象以生产认同的论述，却故意以此包装监控业务。特别是关于AI武器"问题在于谁建造"的论点，以技术必然性为前提，关闭了Habermas倡导的民主审议，暴露了这种"故意误用"的本质。

Ethan Mollick@emollick · 4月20日

An obvious way to release Mythos class models with uncertain autonomous ability is to make them only available on the website, like Gemini Deep Think or ChatGPT Pro. Minimal risk of being used for autonomous hacking, but accessible to people who have hard problems to solve.

译发布具有不确定自主能力的 Mythos 类模型的一种明显方式是仅通过网站提供，就像 Gemini Deep Think 或 ChatGPT Pro 那样。 被用于自主黑客攻击的风险极低，但有难题需要解决的人可以使用。

Rohan Paul@rohanpaul_ai · 4月19日

AI fakery is pushing major apps toward proof of humanity to become a standard login layer. BBC: Tinder and Zoom just backed iris-based proof of humanity as a new defense against bots, scams, and deepfakes online. The reason is because, AI now copies faces, voices, and chat well enough that a profile photo or video call no longer proves a person is real. World, , formerly known as Worldcoin, is part of Tools for Humanity, will scan the iris, turn that into a unique code, and stores the credential on the user’s phone as a World ID. Tinder plans to show a verified human badge, while Zoom plans to use the same credential to reduce deepfake impersonation in meetings. This system is selling personhood, not identity, so it tries to answer “is this a real human?” more than “what is this person’s legal name?”. That fits the scam problem, because US romance scams still cost more than $1.14B, and Deloitte says AI-enabled fraud could hit $40B by 2027. That biometrics can become a reusable internet primitive, like a login layer for a web flooded with synthetic people. --- bbc .com/news/articles/cp9vppem4evo

译AI伪造技术泛滥正推动互联网平台采用生物识别"人性证明"。Tinder与Zoom宣布集成World（原Worldcoin）的虹膜扫描系统World ID，通过唯一生物凭证区分真人与深度伪造或机器人。与传统身份验证不同，该系统验证"人格"（personhood）而非法定身份，旨在应对日益严重的AI诈骗风险。此举或使生物识别成为应对合成人类泛滥的可重用互联网基础登录层。

Rohan Paul@rohanpaul_ai · 4月19日

Anonymous usernames are no longer much protection when LLMs can piece together a person’s public trail. LLMs can identify supposedly anonymous people online by turning messy posts into personal clues. The best setup finds 68% of true matches at 90% precision, meaning 9 out of 10 guesses are right, while older methods stay near 0%. The problem is that pseudonyms often seemed safe only because linking a person across sites used to take lots of careful manual work. This paper cuts that work by making an LLM do 3 jobs: pull identity hints from raw text, search a huge pool of possible matches, and compare the best candidates to reject weak fits. The authors tested this on 3 cases: matching Hacker News users to LinkedIn profiles, matching Reddit movie users across communities, and matching the same Reddit users across different time periods. The main result is that the reasoning step beats simple matching by a wide margin and stays useful even as the candidate pool grows, which matters because it shows that public writing alone can now be enough to join accounts or name a person at scale. ---- Paper Link – arxiv. org/abs/2602.16800 Paper Title: "Large-scale online deanonymization with LLMs"

译LLM可通过分析公开写作实现大规模去匿名化。研究让模型执行提取身份线索、搜索匹配池、比较验证候选者三项任务，在Hacker News与LinkedIn、Reddit跨社区及跨时间段等场景测试中，达到90%精确度与68%召回率，远胜旧方法。关键突破在于推理步骤能处理大规模候选池，证明零散公开文本已足以关联账户并识别个人，传统匿名保护机制失效。