注册并分享邀请链接,可获得视频播放与邀请奖励。

搜索结果 JAGGED
JAGGED 贴吧
一个关键词就是一个贴吧,路径全站唯一。
创建贴吧
用户
未找到
包含 JAGGED 的推特
红杉 AI 大会上,Karpathy 聊了近 30 分钟,把过去几个月编程范式的剧变讲透了 从"软件 1.0/2.0/3.0"的框架,到 MenuGen 项目里"神经网络直接把图片渲染进菜单像素"这种细思极恐的例子 还有"参差智能"(jagged intelligence)、"animals vs ghosts"这些他自己独创的概念。 这是一个真正在大模型一线写代码、踩过坑的人,把 2026 年最前沿的认知浓缩成大白话讲给你听。 章节: 00:30 作为程序员从没这么落后过 01:02 去年12月的转折点:从"能用的AI"到 vibe coding 03:35 软件1.0/2.0/3.0:编程变成了写提示词 04:04 OpenClaw 安装:复制粘贴一段话给 agent 就是新时代编程 04:49 MenuGen 的故事:从写代码到"给 Gemini 一张照片" 05:37 Nano Banana 直接把菜品渲染进像素,整个 app 都成了多此一举 06:11 软件3.0范式 08:07 MenuGen 例子的推演:完全神经网络构成的计算机 09:40 AI 到底为什么在有些领域强、有些领域弱 11:14 经典例子:strawberry 里到底几个 r,到走50米该不该走路去洗车 12:24 国际象棋的秘密:GPT-4 突飞猛进是因为训练数据加了棋谱 13:39 给创始人的建议:可验证性能让你自己做微调、蹭上RL红利 15:46 vibe coding vs agentic engineering 16:54 "10倍工程师"过时了,现在顶尖的人远不止10倍 18:21 招聘该怎么变:别再考算法题,让人做一个能扛住攻击的大项目 19:36 人依然要把关审美、判断力、taste 20:37 Spec 比 plan mode 更重要:把关顶层设计,细节交给agent 21:51 API细节(dim/axis/reshape)已经不用记了,但底层原理还得懂 22:13 品味判断力会不会贬值? 23:31 参差智能:我们不是在造动物,是在召唤幽灵(ghosts) 25:17 agent 能自主行动后的世界 25:27 一切文档依然是给人类写的——Karpathy最讨厌的痛点 27:43 智能变便宜之后,到底什么还值得深入学习 28:06 那条让他难忘的推文:"你能外包思考,但外包不了理解"
显示更多
看得我泪流满面😭🥹💔就凭这一点,难道中国人还能再吃狗肉吗? A Belated Tribute to China’s Wenchuan Earthquake Rescue Dogs. This is a tribute that comes years too late. On October 2, 2021, the last search-and-rescue dog that had served in the 2008 Wenchuan Earthquake, Bingjie, passed away. With his death, all 67 rescue dogs that answered the call during one of China’s darkest disasters had completed their final roll call. This isn’t about whether you love dogs or keep pets. It’s about heroes. Back in 2008, when mountains collapsed and entire cities were reduced to rubble, these dogs charged into places no human could safely enter. At a single whistle from their handlers, they disappeared into twisted steel, shattered glass, and unstable concrete without hesitation. Their paws were ripped apart by broken rebar and jagged debris. They bled with every step. Many worked until the pads of their feet were raw and infected. Yet the moment they caught even the faintest scent of life beneath the ruins, they refused to leave. They barked. They dug. They kept digging. For them, every signal they gave meant another chance for someone to survive. That wasn’t simply bravery. It was something even purer. A complete trust in the person who guided them. A trust so absolute that it stood above fear itself. Bingjie was barely a year old when he was deployed to Wenchuan. Still little more than a puppy, he clawed through what had become a living hell and helped save 13 lives. The first little girl he rescued is now a university graduate. The future she was able to live began with a wet black nose finding her in the darkness… and a pair of tiny paws digging until they nearly gave out. If she remembers that moment, then she remembers who reached her first. No matter how you measure it, humanity owes these dogs a debt that can never truly be repaid. There was also Shenhu, another legendary rescue dog. By the time he retired in 2019, nearly fourteen years old, he had helped save 15 lives. His handler couldn’t bear to part with him, so he brought Shenhu home and stayed by his side until the very end. Today, a statue stands in his honor. But the truth is… Dogs don’t understand medals. They don’t understand monuments. They don’t care about titles or glory. A gentle pat on the head. A favorite tennis ball. A few words of praise from the person they trust. To them, that was worth risking everything. If called again, they would have done it all over without a second thought. People often speak of loyalty. These dogs never spoke a word. They carved loyalty into their bones. Now they have all returned to the place people lovingly call the Rainbow Bridge. Perhaps they’re together again, sniffing one another, wagging their tails, healthy and full of life, just as they were back then. To the silent heroes who gave everything so that strangers could have another tomorrow… Thank you. You will never be forgotten.
显示更多
0
21
108
14
转发到社区
Andrej Karpathy发表评论表示: 因为人们对大语言模型(LLM)能力的认知鸿沟正在进一步扩大。 现在的情况已经不再只是“两群人在各说各话”,而更像是一个急剧收窄的漏斗。 - 约60亿人(全球人口的75%):粗略估算,这些人几乎从未真正接触过LLM。 - 约10亿至20亿人(约20%):他们是ChatGPT等免费产品的轻度、低频用户。这群人接触到的往往是一些略显笨拙的聊天机器人,将其当作功能更强的谷歌搜索、写作辅助工具等等。偶尔也许会有一个AI Agent尝试帮你预订机票。我有一些非科技行业的朋友认为AI对他们几乎没什么用处,而我觉得这种看法完全可以理解。即使是数学和编程领域之外的许多专业人士,也仍然处于这一层级。例如,企业高管和许多其他职能岗位的员工,大部分时间都在与人沟通。因此,尽管他们在理性上理解AI正在发生什么变化,但这种认识仍然来自间接体验,多少有些抽象。 - 约2000万人(0.2%):接下来是那些在数学和编程领域专业使用前沿LLM的人。这些人已经亲眼见证:过去需要几周甚至几个月才能完成的大型复杂项目,现在只需要给AI Agent一个Prompt,就能由它完成。开发应用、复制应用、迁移应用、从二进制文件中反编译应用……这一切都发生得极其迅速,而且就在最近。不到一年前,我还在亲手编写代码,在代码编辑器里逐字逐符地敲入那些记在脑海里的编程指令,偶尔按一下Tab键,让自动补全帮我生成一小段代码。 - 约5000人(0.00006%):最后,是那些能够在内部接触最前沿AI系统的人。外部世界目前看到的,只不过是一个预览版本。真正的前沿能力可能是这样的:成千上万个AI Agent组成集群,连续数周协作,完成超大型软件工程项目;发现零日漏洞;以机器速度发动和防御网络攻击;发现新的科学知识;推进数学研究的最前沿。这些任务过去可能需要业内顶尖专业人士花费数年才能完成。与此同时,人类的审查能力和理解能力已经开始跟不上AI的发展速度。例如,人们至今仍在对数月前的OpenAI-HF事件进行所谓的“考古式研究”。而数学家们可能还需要花费很长时间,才能仔细研究那722篇关于前沿数学的论文手稿。 造成这种漏斗式分化的,是以下几个因素共同作用的结果: 1. AI带来的影响,随着目标的雄心、问题规模和任务时间跨度的增加而扩大。 一个只需要一段文字就能回答的问题,几乎无法触及这些系统的能力上限。你必须拥有大量真正关心、规模庞大且难度极高的问题,才能充分发挥AI的能力。 这一因素主要解释了为什么普通消费者和专业用户之间会出现如此巨大的体验差异。 2. AI能力分布的不均匀性(Jaggedness)。 关于这一点,我此前已经写过很多。 AI能力的高峰集中出现在那些数字化程度高、结果可验证,而且具有明确经济价值的领域。 这是因为LLM的能力很大程度上来自基于可验证奖励的强化学习(Reinforcement Learning on Verifiable Rewards,RLVR),而训练环境的组合经过精心筛选,其选择又受到潜在商业收入的驱动。 这一因素主要解释了为什么不同专业领域之间会产生如此巨大的能力差异,例如数学和编程领域为何尤其突出。 3. 使用权限(Access)。 免费版本、付费版本,以及仅供内部使用的前沿系统,所提供的能力存在显著差异。 这就是我们所处时代的奇特之处。 绝大多数普通人几乎没有真正接触过这些AI系统。即便接触过,他们看到的也只是一个有点笨拙的聊天机器人。 大多数专业人士目前感受到的,也只是相对有限的生产力提升。 然而,极少数专业人士正在亲身经历另一种截然不同的现实:AI能力曲线突然变得近乎垂直向上,他们正体验着这种指数级跃升所带来的眩晕感。 而这一切,正在同一个时代、同一个时刻,同时发生。
显示更多
0
11
57
6
转发到社区