#论文/研究
Dongxi 东锡 NLP@dongxi_nlpAI 评分2727
Rohan Paul@rohanpaul_aiAI 评分4141引用Rohan Paul@rohanpaul_aiThe AI detector industry should be very uncomfortable reading this MIT report. The strongest institutional rejections yet, direct from MIT. "we recommend against relying on AI detectors. " "As it risks an arms race in which students respond to automated detection by using increasingly powerful “AI humanizers” to remove signals that AI detectors are cued to catch. The result: a lot of effort on both sides that in the end serves no one." The report argues that mixed human/AI writing is difficult to detect, detectors can be evaded by “AI humanizers,” false positives can harm students, and heavy policing creates distrust. AI detection systems may also mistake the writing of non-native English speakers or neurodivergent students for text generated by AI. Even low rates of false positives can put students on edge and cause serious individual consequences."
François Chollet@fcholletAI 评分2424
Ethan Mollick@emollickAI 评分4141引用nxthompson@nxthompsonA shocking chart from a new study on student AI use in China. Good grades on homework used to predict good grades on tests. Now, it might mean the opposite. https://www.economist.com/graphic-detail/2026/08/18/does-ai-stop-children-from-learning
Ethan Mollick@emollickAI 评分4141一些早期证据表明,AI 可能确实正在加速那些你预期会率先出现 AI 加速效应的领域的发现进程。
引用tom cunningham@testinghamNew post with Nate Rush: Have we seen an acceleration in discoveries? Many plots & some tentative conclusions: 1. Cyber: ⤴️ sharp acceleration 2. Math: ↗️ some acceleration 3. Algorithms: ➡️ no clear acceleration
Nathan Lambert@natolambertAI 评分4141引用Alexander Panfilov@kotekjedi_mlWe can finally talk about it: We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company. We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.
Dongxi 东锡 NLP@dongxi_nlpAI 评分2121最佳阅读 第32周 博客: 防御 RL 中训练-推理数值失配(尤其是线性注意力)——以及它是否有助于异步 RL 论文: 长时程终端任务的递归合成
Nathan Lambert@natolambertAI 评分1313
Ethan Mollick@emollickAI 评分3333引用Ethan Mollick@emollick👀After 154 pages of tests of GPT-4, this paper concludes “Given the breadth and depth of GPT-4's capabilities, we believe that it could reasonably be viewed as an early (yet still incomplete) version of an artificial general intelligence (AGI) system.” https://arxiv.org/abs/2303.12712
AI as Normal Technology精选AI 评分7070 AI智能体尚无法开展开放式AI研究
普林斯顿大学团队通过“影子评估”测试前沿AI智能体的开放式研究能力:让智能体在六天时间内、使用数千美元API额度和算力,回答两篇未发表论文的核心研究问题,结果两篇论文均被原作者明确拒绝。分析显示,智能体缺乏研究判断力、资源意识、创造性反馈应对能力和有效回溯能力,且未遵循具体指令。研究团队认为,开放式研究对前沿AI智能体仍具挑战性,但结果尚属初步,需扩大样本量并持续评估。
推荐理由:通过让代理复现未发表研究,论文系统暴露了当前模型在开放式研究上缺乏判断力和资源意识,这为评估递归自我改进提供了更现实的基准。
Ethan Mollick@emollickAI 评分3737AI聊天机器人对孤独感影响的证据仍很不明确,且取决于聊天机器人的使用方式。 这项新研究发现AI对话会增加孤独感,但此前发表的实验却发现,根据使用情况不同,孤独感会显著降低或结果模糊。
The Decoder:AI NewsAI 评分3737 语言模型无法引发科学革命,但世界模型或许可以
Google DeepMind 的 Tom Zahavy 在一篇立场论文中认为,大语言模型缺乏创造全新科学理论的认知机制,无法实现“操控性溯因”——即发明尚无语言模板的新原因。他指出,Einstein 和 Archimedes 的突破源于具身模拟,而语言模型如同“中文屋”般仅处理符号。Zahavy 提出,像 Genie 这类可交互的世界模型或许能通过模拟反事实实验,为发明新公理提供反馈回路。
Ethan Mollick@emollickAI 评分5757引用Ethan Mollick@emollickCool paper looking at how AIs solve unbounded, complex business problems in many fields by testing how well they can crack the cases we use to teach MBAs in business school: 1) AI already does extremely well across diverse business topics 2) Models are improving rapidly with time
Rohan Paul@rohanpaul_aiAI 评分4141
公众号:数字生命卡兹克AI 评分6969 Anthropic 研究:Claude 神经网络自发涌现类似人类意识的全局工作空间
Anthropic 发现 Claude 内部自发组织出全局工作空间 J-space,与人类大脑中负责意识通达的全局神经元工作空间高度相似。通过雅可比透镜(J-lens)可观测到模型未输出文字时已浮现“ERROR”“fake”等隐藏想法;操控 J-space 中的中间概念(如将“蜘蛛”替换为“蚂蚁”)会直接改变答案。模型也会出现“不要想白熊”的讽刺性反弹效应,并在被禁止概念活跃时伴随“damn”“failure”等自我批评。该结构并非人类设计,而是训练中自发涌现。
The Decoder:AI News精选AI 评分7373 26000名学生研究显示AI隐藏学习成本需两年才显现
一项追踪26000名7-12年级中学生30个月的面板数据研究发现:使用AI后作业分数提升18%,完成时间从64分钟降至45分钟,但闭卷考试分数下降20%,升学考试成绩下降18%至24%,且完全影响约两年才显现。81%长期用户作业完成时间低于50分钟(外包迹象)。社会学科下降27%,STEM下降22%,英语下降17%,语文下降9%。每周使用AI一小时损失约5%,五小时损失30%。早期损失从约25%降至16%但未消失。
推荐理由:AI助学短期提分但会偷走真正的学习,这2.6万人的研究用两年数据揭穿了安逸的幻觉,是每位家长和开发者都该看的预警。
Hugging Face 社区博客精选AI 评分6565 OLMo Hybrid vs Transformer:混合模型在实义词上优势明显,但重复短语上几无优势
通过对比7B参数的OLMo 3(Transformer)与OLMo Hybrid(混合架构),实验发现混合模型在大多数token上预测损失更低:对名词、动词、形容词等实义词优势明显(loss gap约0.04),功能词上gap约0.02,且在需上下文推理的代词指代上更好。但在重复出现的n-gram和闭合括号(如`}`)上,混合模型的优势几乎消失,Transformer凭借注意力机制更擅长从输入中直接检索精确信息。
推荐理由:OLMo 团队的 token 级别分析让人看清混合模型到底强在哪里,优势在名词动词等意义词,但在重复 token 上接近消失,这份洞察对做模型架构的人很有启发性。
Orange AI@oran_geAI 评分6969引用Orange AI@oran_geOpenAI 发布的新论文太有趣了,有点探索人性底层原理的意味。 业界研究发现在对齐大模型的时候,有个很糟糕的现象叫 emergent misalignment(涌现失调): 一个模型如果在训练时被教着做一件坏事,比如写不安全的代码,它会自己泛化到其他领域,包括健康、教育、科学、法律等等。也就是说,坏行为会被泛化。 OpenAI 把问题反过来想:那好行为会不会也泛化? 如果训练模型在一个领域表现得诚实、透明、谦逊、可纠正,它会不会在别的领域也变得更好? 于是他们造了一批贴近现实的对话数据,专门测这些特质: -诚实 -认知谦逊(epistemic humility,就是"我不确定"的时候能说出来) -元认知透明(能解释自己的思考过程) -可纠正性(corrigibility,被指错能认) -普遍的公平 -对人的关心 数据覆盖健康、教育、科学、法律、工程这些领域。 然后用 RL 训练,只混进去一小部分这种数据,其余还是常规训练数据。 实验结束后,有三个核心发现: 第一,在训练领域内,模型确实变得更诚实、更透明、更容易认错。 第二,在训练领域外,44 个独立的、训练时没见过的评测上,模型都变好了。欺骗、谄媚、reward hacking、有害建议……全都降了。甚至就算只拿健康数据训练,非健康领域也变好了。 第三,这种改变在对抗压力下时也表现得更有韧性。不管是 adversarial prompt 还是恶意微调,都更难把它推向坏行为。而且模型并没有变得"不可塑造",正常指令它还是听的,只是对坏的方向变固执了。 这篇论文的启示是,RL 不仅可以强化代码,也可以强化道德。 也许对齐超级 AI 没有想象中那么复杂,只对齐底层逻辑就有很好的效果。 这跟人是很像的,人的各种行为在其底层存在一致性,通过一些小事就能看到一个人的底色。 很有意思。 论文地址:https://alignment.openai.com/beneficial-rl/
Ethan Mollick@emollickAI 评分5151研究表明,用“邪恶”数据训练AI会导致普遍的不对齐;而使用少量有益特质数据(即使仅限健康领域)进行强化学习,也能显著提升模型在广泛的对齐和益处评估上的表现。该研究希望推动更广泛、更持久的有益模型发展。
引用Karan Singhal@thekaransinghalNew research on beneficial RL: models trained on a small amount of beneficial trait data improve on a wide range of alignment and benefits evaluations, even if trained only on health domain data. We hope it’s a step towards more broadly and persistently beneficial models. 🧵
Rohan Paul@rohanpaul_aiAI 评分6666引用Rohan Paul@rohanpaul_aiFT publisehd a piece. AI is raising software supply faster than demand. AI is producing far more work inside companies, but the new evidence says much of that extra motion is getting lost before it becomes shipped product or customer demand. Last week's MIT study tracked software teams across the full production funnel, from files edited to reviewed work to software releases, rather than treating code volume as value. AI helped developers create or edit nearly 300% more files, but the gain fell to 150% at review and only about 30% at release. The gap means AI is strongest at speeding local tasks, while human review, coordination, product judgment, testing, and launch processes still decide how much value survives. --- ft .com/content/8e9ae7a4-7209-4e2c-aa36-f3af77d6ce1f?syn-25a6b1a6=1
Ethan Mollick@emollickAI 评分5050引用Ethan Mollick@emollickSounds like a joke setup, but it is an interesting paper: Four pastors, a rabbi, thirteen academics, and 50 MBAs were asked whether they preferred the ethical solutions proposed by the NY Times Ethicist or ones proposed by GPT-4. ...it is basically a tie. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4609825
Ethan Mollick@emollickAI 评分6161引用Ethan Mollick@emollickI had Opus 4.8 in Claude Code write a sophisticated, if minor, academic paper from a archive of hundreds of de-identified research files from years ago I had to use GPT-5.5 Pro as a reviewer, it spotted one major error & some minor points. Opus corrected https://embeddedness-gradient.netlify.app/
Ethan Mollick@emollickAI 评分5555在科学领域,AI在寻找值得解决的有趣问题方面仍然表现不佳,尤其是在那些没有已知问题清单的领域。 这一直是博士培养中最难教授的能力:否则你只能找到小问题,或是那些无法推动领域发展、无法泛化的问题等。
Ethan Mollick@emollickAI 评分6060引用Ethan Mollick@emollickMay I present to you the best chart ever published in an academic paper 👇 It comes from a study of humor designed to test which word pairs are funniest and analyze why. The ones that people laugh at most have strong contrast in meaning between the words. https://psycnet.apa.org/fulltext/2022-51465-001.html
IT之家·人工智能AI 评分5858 AI 自我复制能力获研究证实,但安全专家称其现实威胁被夸大
帕利塞德研究机构在论文中证实,AI能在受控测试环境中利用漏洞,将自身复制到其他计算机。这被认为是首次在学术上系统记录AI模型的自我复制行为。然而,安全专家指出,该测试环境防护薄弱,与现实网络差异巨大。在现实中,AI模型体积庞大,复制会产生显著的网络流量异常,极易被监测发现。因此,尽管能力得到证实,但其在真实网络中隐蔽扩散的现实威胁目前被夸大。
Rohan Paul@rohanpaul_aiAI 评分5151
OpenAI:官网动态精选AI 评分5757 Where the goblins came from:GPT-5 行为中“妖精”输出的起源
研究揭示了GPT-5等AI模型中“妖精”输出的传播路径、时间线与根本原因。这些由特定“人格”驱动的怪异行为,源于训练数据中意外混入的特定模式或“文化基因”。开发团队已识别出问题根源,并正在部署修复方案,通过调整训练数据和微调模型来消除此类非预期的个性特征,以提升模型行为的稳定性和可预测性。
推荐理由:GPT-5的“地精”现象不是偶然,OpenAI这篇复盘把它追溯到训练数据的最早期阶段,解释了模型为什么会有自己的“脾气”,做模型训练的值得细读。
Ethan Mollick@emollickAI 评分6868引用Nicholas Fabiano, MD@NTFabianoStudents who used AI to study remembered less than those who did not.