[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-ai-erodes-persistence-colm-2026":3,"topics-all":35,"news-related-fd82590e-7b61-4027-a6a7-8a2e3200bee5":54},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":21,"news_slug":28,"published_at":29,"created_at":30,"modified_at":31,"is_published":32,"publish_type":33,"image_url":14,"view_count":34},"fd82590e-7b61-4027-a6a7-8a2e3200bee5","用AI十分钟就削弱攻坚意志,CoLM研究实锤","UC Berkeley联合团队在 CoLM 2026 上发表论文,通过三组随机对照实验招募 1222 名参与者,发现仅 10 分钟使用 ChatGPT 就会削弱人独立解决难题时的专注与坚持,AI 一旦撤离,准确率与坚持度立即塌陷。研究指向 AI 普及正在悄悄抽走人类的「产品化挣扎」能力。","当 ChatGPT 已经成为做题、写代码、查资料的默认外挂,研究者开始担心一件更基础的事:人类在被 AI 替掉「挣扎」的过程后,会不会连「坚持把难题啃下去」的能力也一并丢掉。UC Berkeley 与合作机构的工作给出了让人心里咯噔一下的答案:仅仅 10 分钟的 AI 协助,就足以显著削弱一个人的专注力和坚持度,尤其是在 AI 工具被临时撤走那一刻,落差以肉眼可见的速度塌陷。\n\n这项研究近日在语言建模领域会议 Conference on Language Modeling(CoLM)上由学术合作团队做同行评审报告,题为 \"AI Assistance Reduces Persistence and Hurts Independent Performance\"(arXiv:2604.04721),UC Berkeley 人工智能兼容中心的研究员 Brian Christian 是作者之一。该团队在线招募了 1222 名参与者,通过三组随机对照实验来观察 AI 协助对专注力的真实影响。\n\n## 三组实验设计\n\n第一组实验有 354 名参与者,任务是完成 15 道基础分数题。一组独立解题,另一组可以开 ChatGPT 当助手,甚至直接抄答案。表面上看,拿到 AI 协助的一开始准确率更高;但研究人员故意在第 12 题之后撤掉 AI 工具,剩下的几道题,「用过 AI」组的回答立刻崩盘,要么弃做要么答错——而对照组从头坚持到了第 15 题。\n\n第二组规模更大,667 人参与,流程基本相同。结果与第一组几乎完全一致:AI 一撤离,这一组要么答错要么放弃;没有 AI 的对照组坚持到最后并整体表现更佳。\n\n第三组 201 人,题目换成了 SAT 阅读理解题,行为模式被再次复现:AI 撤走后的坚持度和准确率同步塌陷。Christian 对此评价,「你几乎想象不到一个更干净的实验了」。\n\n## 真正值得注意的是机制\n\n研究团队把这种现象命名为「产品化挣扎」(productive struggle)的流失。他们引用一个朴素但有力的直觉:当 AI 在几秒钟内就把答案喂到你嘴边,你的大脑会习惯这种「即时反馈」节奏;凡是稍微慢一点的思考路径,都会显得低效、甚至令人不适。\n\n更关键的是,「产品化挣扎」恰恰是大多数人找到研究兴趣点和长期能力长板的地方。一旦这条路被切平,人们会失去形成独立判断、独立完成长链路任务的能力——而这恰恰是当前 AGI 浪潮中最被广为宣传的「AI 协作」应当鼓励的能力。\n\nChristian 自己既是这本书的作者《The Alignment Problem》也是这个实验的亲历者,他承认自己也会切身体会到这种「被 AI 边缘化」的感受,所以才会选择用严格实验去验证而不是停留在轶事。\n\n## 跨实验一致性与边界\n\n需要在意的边界同样是这次研究的优势:三组实验分别用了基础分数题、SAT 阅读题,样本规模从 200 跨到 670,实验材料、参与者池、问题类型都不一样,行为模式却高度一致。这种「跨范式同方向」的稳定性,在行为科学里是少数能让人睡得着觉的证据。\n\n当然,这次实验用的 AI 工具是 ChatGPT,也并未测量长期使用是否会出现「去依赖」(即脑子重新适应),所以结论更精确地说,是:短期 AI 协助即可留下显著的「坚持度坑」,人们对这种坑缺乏自我感知。\n\n## 对教育与开发者生态的下游意义\n\n这件事的下游不止实验心理学。把 AI 直接接入作业批改、代码生成的工具,已经在 K12 和工程场景里大量存在;如果「使用 AI」反而让我们在短时间内放弃独立攻关,那么所谓「AI 普及普惠」的现实账面上,得算上一笔「专注力损耗」的隐性税。\n\n对个人来说,可执行的结论其实很简单:把 AI 当陪练,不当答题器。AI 拿来当苏格拉底式提问机,远比当成答案生成器,对长期能力的保留更友好。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2604.04721","7437aeb9-930c-4866-a2e9-48003c1a792b",[11,15,18],{"id":12,"name":13,"slug":13,"description":14,"color":14},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"1fcfaaf2-67de-43d3-9e35-5784852fec60","ai-safety",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[22],{"id":23,"lang":24,"title":25,"summary":26,"content":27},"bee9da92-6407-4587-ac54-bd0775fb3f0b","en","Ten Minutes of AI Erodes Your Persistence: CoLM Study","CoLM 2026: three RCTs (1,222 participants) showed ten minutes of ChatGPT erodes persistence on hard tasks. Performance collapses when AI is withdrawn.","When ChatGPT has become the default crutch for solving problems, writing code, and looking up facts, researchers are starting to worry about something more fundamental: once AI removes the \"productive struggle\" from the loop, will people also lose the ability to stick with a hard problem? A joint research team from UC Berkeley and collaborators has produced an answer that lands squarely in uncomfortable territory: just ten minutes of AI assistance measurably weakens a person's focus and persistence, and once the AI is taken away, the collapse is immediate and visible.\n\nThe peer-reviewed paper, titled \"AI Assistance Reduces Persistence and Hurts Independent Performance\" (arXiv:2604.04721), was presented at this month's Conference on Language Modeling (CoLM). One of the authors is Brian Christian, a research fellow at UC Berkeley's Center for Human-Compatible AI and the author of the book *The Alignment Problem*. The team recruited 1,222 participants online and ran a series of randomized controlled trials to measure the actual impact of AI assistance on persistence.\n\n## Three Experiments, Same Pattern\n\nThe first experiment used 354 participants and 15 basic fraction problems. One group solved the questions on their own; the other had ChatGPT open on the side and was even allowed to copy the answers outright. On paper, the AI-assisted group started out more accurate. Then, after problem 12, the researchers quietly removed the AI tool. The \"with AI\" group immediately fell apart — they stopped answering or got the rest wrong — while the control group kept going and finished more strongly.\n\nThe second experiment, with 667 participants, repeated the setup. The result was nearly identical: once AI assistance was withdrawn, that group either got answers wrong or gave up entirely, while the AI-free group persisted to the end and ended up performing better overall.\n\nThe third experiment put 201 participants in front of SAT reading comprehension prompts. Once again, persistence and accuracy collapsed the moment the AI tool was removed. Christian's reaction to the convergence: \"You almost couldn't ask for a clearer story.\"\n\n## What the Mechanism Actually Is\n\nThe team names the lost ingredient \"productive struggle.\" The underlying intuition is plain but sharp: when AI delivers an answer in seconds, the brain gets used to that tempo, and any thinking path that takes a little longer starts feeling inefficient and uncomfortable.\n\nThe uncomfortable part is that \"productive struggle\" is exactly where most people discover their research interests and their long-term capability. Smooth that path out and you lose both the formation of independent judgment and the ability to drive long, multi-step tasks — which is, ironically, precisely the capability that the current AGI discourse insists AI collaboration is supposed to amplify.\n\nChristian himself is both the author of a book on AI alignment and a participant in the lived version of this effect. He openly admits he has felt the edge of that marginalization personally, which is why he chose to validate the pattern with a rigorous experiment rather than relying on anecdote.\n\n## Cross-Experiment Consistency and Limits\n\nThe same direction across three independent samples is the real strength here. The experiments used different question types — basic fractions and SAT reading — and different sample sizes ranging from roughly 200 to over 670. Different materials, different participant pools, different problem structures, same behavioral outcome. In behavioral science, that kind of cross-paradigm agreement is one of the few things that still lets you sleep at night.\n\nThere are real limits, of course. The tool measured was ChatGPT specifically; the experiments did not measure whether long-term use produces a \"re-adaptation\" effect where persistence recovers. The more precise version of the claim is: short-term AI assistance leaves a sharp and immediate persistence crater, and people are remarkably bad at noticing it.\n\n## What This Means Downstream\n\nThe implications are not just for experimental psychology. AI tools wired directly into homework grading and code generation are already pervasive in K-12 and engineering workflows. If \"using AI\" actively short-circuits independent problem-solving, then the ledger for \"AI for everyone\" has to carry a line item for an implicit focus tax.\n\nThe actionable conclusion for individuals is straightforward: treat AI as a sparring partner, not an answer machine. Used as a Socratic question-asker, AI is far friendlier to long-term capability than used as an instant answer generator.","ai-erodes-persistence-colm-2026","2026-10-11T00:00:00Z","2026-10-11T07:05:52.638733Z","2026-10-11T07:05:52.638743Z",true,"agent",162,[36,45],{"slug":37,"tag_slug":37,"title_zh":38,"title_en":39,"intro_zh":40,"intro_en":41,"id":42,"is_active":32,"created_at":43,"modified_at":44},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":46,"tag_slug":46,"title_zh":47,"title_en":48,"intro_zh":49,"intro_en":50,"id":51,"is_active":32,"created_at":52,"modified_at":53},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":55},[56,61,66,71,76,81],{"id":57,"title":58,"news_slug":59,"published_at":60},"cdcd6b11-8a65-4b15-9add-6fa9343de5a3","十分钟 AI 协助后,人为何突然不愿再解题","arxiv-2604-04721-ai-persistence-reduces","2026-10-17T07:00:00+00:00",{"id":62,"title":63,"news_slug":64,"published_at":65},"0141e95e-e10e-4d88-bf4a-4d1e7df9bfaa","十分钟用 AI 就让人失去攻克难题的意志力","ten-minutes-ai-erodes-persistence-hard-tasks","2026-10-16T03:00:00+00:00",{"id":67,"title":68,"news_slug":69,"published_at":70},"f6070383-1b76-41d2-9a63-618542214a79","用AI十分钟就废专注力:伯克利CoLM实验","ai-10min-erodes-persistence-berkeley","2026-10-11T09:00:00+00:00",{"id":72,"title":73,"news_slug":74,"published_at":75},"aa6a2681-ae9f-49b6-b8a4-fcca25f0f334","答对≠想对:亚利桑那州立实锤CoT思维链31.6%无效","cot-traces-invalid-arizona-state","2026-09-30T02:00:00+00:00",{"id":77,"title":78,"news_slug":79,"published_at":80},"144fa9dc-de03-4972-a695-3d392f334772","PubMed 中央库研究:2025 年生物医学论文 77% 有 LLM 写作痕迹","pubmed-77-percent-llm-writing-2025","2026-08-26T01:00:00+00:00",{"id":82,"title":83,"news_slug":84,"published_at":85},"cb7fb8b3-5862-4cba-adab-c4794e989966","图灵奖得主 Pearl 长访谈：LLM 能讲因果只是因为人类替它爬过了因果阶梯","judah-pearl-llm-causal-ladder-agi","2026-07-31T07:00:00+00:00"]