{ "text": "大家好,欢迎来到新视频。\n开始之前,我也想告诉大家。\nDeep Seek的Harness也有新更新\n我会在另一个频道ARC King上讲\nDeep Stick V41 Flash现在已经正式发布了\n经过我用正确的Finting设置测试后\n我觉得这个模型真的很强。\n另外还有一份技术报告\n解释DC是怎么让这个模型更高效的\n先看看到底新在哪\n然后我会给你看这些生成结果\n以及我给这些分数背后的代码\nDeep-C把Flash称作新架构家族里最小的模型。\n但报告里写的是\n它的主干是一个5520亿参数的Mix of Experts\n再加上1960亿参数的Ngram Memory Modules\n它还是个巨大的模型\n真正有意思的是它实际激活了多少参数\n处理输入时大约激活80亿参数\n生成输出时则是160亿\n新的英国编码器解码器设计\n让大多数输入Token可以跳过完整的Decoder计算\n这很重要\n因为Coding Agents会花很多时间读文件\n看工具结果\n还有对话历史\nDeepsea还表示在相同工作负载下\nGlobal Attention Cache的GPU显存用量大约只有上一代的四分之一\nPersistent Cache的存储占用只有八分之一\n这些节省都适用于缓存的上下文\n它们有助于解释更低的服务成本\n但并不意味着整个模型一下子变小了\n原生视觉理解意味着你可以把图片和文字一起发过去\n收到文字回复\n我这里的测试主要是看它在文本提示下写代码的能力\n所以视觉能力算是这次发布的一部分\n而不是我在这个视频里专门做过基准测试的内容\n官方公布的结果也挺有说服力\nDeepSeq报告的DeepSW1.1成绩是74.2\n而V4 Pro是62.7\n这里说的是他们在最高投入下的结果\n在一些更难的基准测试上\n其他模型还是领先\n而且报告也显示\n围绕模型的Agent软件会影响最终分数\n所以我想看看它在开源代码里能做什么\n现在\n我得纠正一下我最初的印象\n我原来那篇开源代码模型的条目里\nThinking是被禁用的\n我一开始怀疑这是回退了\n不过\n开启Thinking并把Effort调到最高后\n分数从53.75提升到了81.25%\n保存的绘画里有大量推理输出\n生成结果也明显更好了\n这些测试是在今天发布前\n使用临时的API模型跑的\n我还没有用新的正式版\n模型名重新跑一遍套件\n这些是我上手预览的结果\n外加官方发布信息\n我没有证据表明\nDIPS在第一轮偷偷用了一个更弱的版本\nTeambench 3有8个任务\n每个任务10分\n这是我的实际评分\n我还打开了生成出来的界面\n检查了代码\n并保留了原始输出\n那我们直接开始\n首先是电梯模拟\n这里应该有三部电梯\n每部在一个人\n人们在不同楼层等待\n乘坐空闲的电梯\n前往各自的目的地\n把鼠标悬停上去时\n应该能显示他们想去哪里\n看新的结果\n楼洞布局很清楚\n三部电梯警都显示出来了\n乘客也在各自楼层排队\n你可以看到人们上车的过程\n还能看到等待运行中和\n以到达的技术在更新\n在浏览器里检查时\n生成了两批个24个人\n两批加起来数量都对\n48名乘客全部到达\n而且任何时候都没有一部轿箱\n同时在超过一个人\n技术器\n在这次运行里也一直保持一致\n之前\n舱势会消失在屏幕外\n人也会被画到错误的位置\n新代码还会清理已经完成的乘客元素\n这对重复使用很有帮助\n现在还有一个重置边缘情况\n如果你在错开的生成\n还在进行时清空模拟\n之后可能还会出现一些带处理的乘客\n但主要模拟已经能正常工作了\n这部分从三分提升到了十分里的八分\n第二个任务是隐形眼镜盒\n我想要一个3D物体\nL和耳标时要很明显\n盖子点击后能打开\n这个版本就可信多了\n现在有一个干净的白色和尘\n薄荷绿的盖子\n还有大号清晰可读的字母\n两个盖子可以分别打开\n而且两个都打开按钮\n也确实如字面所示\n你也可以再把它们关上\n然后旋转这个物体来查看\n之前的版本还有\n字母有些破碎\n但在盖子下面\n能看到一个看起来很实的表面\n这里的结构是围绕脚链组织的\n所以动作逻辑就对了\n虽然内部空间看起来还是有点浅\n而且默认镜头在盖子完全打开时\n会把边缘裁掉\n所以还需要再打磨一下\n但这已经比提示词的理解好多了\n七分原来只有五分\n接下来是折叠桌\n这在第一次生成时是最明显的失败之一\n因为桌面是飘着的\n桌腿还埋在地下\n这次展开后的桌子稳稳地立着\n能看到四条腿\n木质表面也不错\n滑块反应很顺\n折叠部件也能作为一个整体保持在一起\n看看这两个结果的差别就知道了\n这段代码把腿连接到了合适的运动部件上\n解决了前面那个核心问题\n不过在完全折叠的状态下还是会扣分\n腿部横杆会和叠在一起的桌面板块相交\n所以整个机构并没有从头到尾都做到物理上干净\n我给它六分 满分十分\n它确实有进步\n但也很能说明模型还有哪些地方需要继续检查\n第四个任务是熊猫吃汉堡的SVG\n这个版本构图更干净\n表情也更生动\n汉堡的层次还挺清楚\n爪子和一点点碎屑也增加了些趣味\n它是一副紧凑\n完整的插画\n而且渲染的也正确\n剩下的问题是动作表现\n看起来更像熊猫抱着汉堡\n而不是在咬它\n所以我更喜欢这张插画\n但它还没有完全抓住那个动作\n这个提示词打了六分\n之前是五分\n现在这个弓箭游戏在视觉上是最明显的提升之一\n我们有日落场景\n背景里的群山\n一个弓箭手\n四个目标\n还有一个清晰的计时器\n拖动瞄准再松手\n感觉更像真正的射箭游戏了\n核心射击流程是通的\n在浏览器检查里\n四只箭都命中了四个目标\n目标被击中后也倒下了\n游戏内计时器显示\n这一轮四箭用了6.10秒\n这说的是游戏玩法表现\n和模型生成游戏所花的时间是分开的\n我的评分表上这是8分\n比之前的5分高了不少\n不过代码审查也发现了一个\n我必须说明的半\n这样就不会弹出结果页面了\n所以在那次检查里\n分数没能保存\n射击分为合命中反馈都不错\n但完成流程还需要修\n这也是为什么不能只看低评\n第六个任务是那个数学题\n考的是在受限移动条件下\n统计有序对的排列数\n预期答案是20460\n这次保存的绘画里\n最终答案很清楚\n而且是对的\n还有一个SOVER可以再跑一遍\n我又跑了一次\n结果得到了20460\n而且和他检查过的小样本上的暴力枚举\n结果一致\n模型发现了自己最初转移逻辑里的一个错误\n把它改正了\n还先拿小规模暴力样例\n验证了修正后的算法\n然后才给出最终答案\n这种解题行为非常有用\n所以这是十分里的十分\n现在还有一条更清晰的证据\n练来支持这个结果\n第七个任务是本地Panda微调工作流\n我让它创建数据集\n微调一个JM2B模型\n再做一个本地网页界面\n刷新时生成一条Panda冷知识\n结果很不错\n这个项目里有一个完成了700次迭代的LORA训练\n保持了Adductors\n还有一个能正常运行的本地应用\n数据集里有151条人工编写的事实\n之前版本只有127条\n服务方式也更简单了\n模型和Adductors通过MLX直接加载\n并使用Tokenizer的Chat Template\n这样就去掉了上一个方案里\n额外的导出步骤和手写模板层\n界面也挺好看\n有一张简单的Panda插画\n一个清晰的事实卡片\n还有一个生成新试试的按钮\n布局在手机上也能正常显示\n刷新页面和点击按钮\n都可以生成新的内容\n在评审过程中模型产生了回复\n五个样本身程度很干净\n是完整的单据\n没有我们之前看到的那些多余标记\n两次本地直连请求各花了大约半秒\n这是本地微调后的GEM模型在回答\n和DeepSeek的API生成速度是分开的\n我把这个任务打了10分\n比之前的6分高\n不过还是要说明\n数据质量仍然有一些限制\n所有验证答案\n其实都已经出现在训练集中\n测试文件是复制的训练性\n而且有些人工编写的事实并不准确\n在评审过程中\nRefresh还重复了一条事实\n所以完成整个工作流\n并不能说明\n每个生成出来的Panda结果都可靠\n但能把训练和本地应用这样跑通\n还是很厉害的\n最后是这只手表\n这大概是我改观的最佳视觉证明\n第一版的表盘还是一整块金色表面\n现在有了更细致的表盘\n清晰的指标金属表带\n能读的星期和日期窗口\n汇动的指针\n还有一个第二时区的小表盘\n整体观感精致多了\n交互也没问题\n你可以选两个时区互相切换\n开关星期和日期显示\n还能在扫秒和跳秒之间切换\n控制检查也通过了\n说明设计下面那些关键部分也没问题\n指针在午夜时能准确对齐\n梦买的半小时偏一夜\n有效了\n纽约的下令时切换也正常\n第二个时区能正确滚到第二天\n代码还附用了时间格式画起\n避免像旧版本那样每一针都重复创建\n视觉结构和时钟逻辑终于配合起来了\n我真的很喜欢这个结果\n它从四分涨到了十分\n是复册里提升最大的一个\n现在我们来看最后的图表\n电梯8分\n隐形眼镜和7分\n桌子6分\nSVG6分\n射简8分\n数学\nFine Tuning Workflow和手表都是10分\n总分是65分\n满分80分\n也就是81.25%\n不开启思考时的成绩是80分里的43分\n修正配置并重新跑任务后\n我们竟提升了22分\n也就是27.5个百分点\n这个新分数也高于我之前V4Flash的72.5%\n以及V4Pro条目的76.25%\n它在这个排行榜上\n和Gemina3.8Flash打平\n只比Open4\n8的80%高一点\n虽然没排到总榜第一\n但这个结果已经非常有竞争力了\n再说速度\n我看了OpenCode日志\n它为这个新的测试文件夹\n保存了绘画记录\n共有64条已完成的模型回复\n有可用的Token数和时间记录\n像绘画标题\n这类后台调用都被排除了\n另外还排除了一条\n没完成的Restorch回复\n在这些已完成的回复里\n一共生成了364574个Token\n大约294000个是Risning Tokens\n另外大约70\n000个是正常生成的Token\n包括Code2调用参数和答案等其他输出\n所以\n这次生成的Token大多都用在了思考上\n这两个数字都包含了Risning\n它们是根据OpenCode的时间\n戳重建出来的估算值\n它们不包括后面执行工具花的时间\n也不是服务器内部解码速度的精确测量\n我把每次响应的时长加总了\n所以同时跑多个任务\n不会把单次响应速率认为太高\n我觉得这很厉害\n他一边处理多个任务\n一边花了几万tokens做listening\n同时还能保持很高的生成速度\n完整任务耗时能让你从另一个角度来看\n电梯任务大概花了4分20秒\nSVG任务大约11秒\n本地翻桌的流程大约花了11分38秒\n包括搭建环境训练和检查APP的时间\n所以最高强度的思考可能意味着成品要等更久\n在这次附色里\n这些额外的工作带来了结果的大幅提升\n这个取舍我很满意\n现在API价格已经上线了\n非高峰时段\n每百万未缓存输入token15美分\n每百万输出token60美分\n高峰时段输入是30美分\n输出是1.2美元\n缓存输入在非高峰时段每百万Token\n只要0.3美分\n高峰时段价格翻倍\n我会把当前的价格表面接放在下面\n要使用这个已发布的模型\n请选择DeepSeek API地址\nDeepSeek在支持的工具里列出了OpenCode\n记得把Thinking打开\n并把Efford设为Max\n如果你想看这次回测用的配置\n旧的V4 Flash模型已经退役了\nAPI名字会临时录游到V4.1 Flash\nV4 Pro的请求会在9月14日UTC凌晨4点开始按Flash价格录游到V4\nE Flash直到V4.1 Pro发布\n如果你依赖Pro模型名\n这个时间点就很重要\nRizni也会消耗可计费的Output Tokens\n所以即使把Effort设到最高还是会有成本\n不过结合这个定价实测速度和81.25%的结果V4\nE Flash还是一个很有吸引力的编程选择\nDeepSea还分享了技术报告和模型仓库\n并表示他正在于\n开源社区在推理支持和部署选项方面\n我的看法有了很大变化\n因为Max开启了Thinking Mode\n这个模型在我的测试里表现得非常不错\n尤其是在万表和本地模型工作流上\n电梯案例也很清楚地说明\n改进不只是界面更好看\n不过还有一些bug要处理\n表格和SVG的分数较低\n说明他并不是每个任务都一样强\n但我还是会非常愿意重新看看\n他在编码上的表现\n配置错误让第一次印象不太公平\n这次复色让我们看到了更真实的情况\n总体来说挺酷的\n总之\n欢迎在评论区告诉我你的想法\n如果你喜欢这个视频\n可以通过Super Thanks打赏\n或者点击中按钮成为会员\n另外\n给这个视频点个赞\n订阅我的频道\n我们下期见\n拜拜", "language": "multilingual", "duration": 891.8785, "segments": [ { "start": 4.803, "end": 6.753, "text": "大家好,欢迎来到新视频。", "language": "zh", "chunk": 0, "local_language_run": 0, "language_run": 0 }, { "start": 7.333, "end": 9.373, "text": "开始之前,我也想告诉大家。", "language": "zh", "chunk": 0, "local_language_run": 0, "language_run": 0 }, { "start": 11.631, "end": 13.44, "text": "Deep Seek的Harness也有新更新", "language": "zh", "chunk": 0, "local_language_run": 1, "language_run": 1 }, { "start": 16.179, "end": 18.525, "text": "我会在另一个频道ARC King上讲", "language": "zh", "chunk": 0, "local_language_run": 2, "language_run": 2 }, { "start": 20.588, "end": 23.378, "text": "Deep Stick V41 Flash现在已经正式发布了", "language": "zh", "chunk": 0, "local_language_run": 3, "language_run": 3 }, { "start": 28.881, "end": 31.325, "text": "经过我用正确的Finting设置测试后", "language": "zh", "chunk": 0, "local_language_run": 4, "language_run": 4 }, { "start": 36.567, "end": 38.577, "text": "我觉得这个模型真的很强。", "language": "zh", "chunk": 0, "local_language_run": 5, "language_run": 5 }, { "start": 41.677, "end": 42.967, "text": "另外还有一份技术报告", "language": "zh", "chunk": 0, "local_language_run": 6, "language_run": 6 }, { "start": 42.967, "end": 46.127, "text": "解释DC是怎么让这个模型更高效的", "language": "zh", "chunk": 0, "local_language_run": 6, "language_run": 6 }, { "start": 46.127, "end": 47.767, "text": "先看看到底新在哪", "language": "zh", "chunk": 0, "local_language_run": 6, "language_run": 6 }, { "start": 47.767, "end": 50.747, "text": "然后我会给你看这些生成结果", "language": "zh", "chunk": 0, "local_language_run": 6, "language_run": 6 }, { "start": 50.747, "end": 53.807, "text": "以及我给这些分数背后的代码", "language": "zh", "chunk": 0, "local_language_run": 6, "language_run": 6 }, { "start": 55.863, "end": 59.38, "text": "Deep-C把Flash称作新架构家族里最小的模型。", "language": "zh", "chunk": 0, "local_language_run": 7, "language_run": 7 }, { "start": 63.328, "end": 64.238, "text": "但报告里写的是", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 64.238, "end": 67.978, "text": "它的主干是一个5520亿参数的Mix of Experts", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 67.978, "end": 71.198, "text": "再加上1960亿参数的Ngram Memory Modules", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 71.198, "end": 72.998, "text": "它还是个巨大的模型", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 72.998, "end": 75.938, "text": "真正有意思的是它实际激活了多少参数", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 75.938, "end": 78.318, "text": "处理输入时大约激活80亿参数", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 78.318, "end": 80.158, "text": "生成输出时则是160亿", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 80.158, "end": 82.578, "text": "新的英国编码器解码器设计", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 82.578, "end": 86.118, "text": "让大多数输入Token可以跳过完整的Decoder计算", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 86.118, "end": 87.378, "text": "这很重要", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 87.378, "end": 89.718, "text": "因为Coding Agents会花很多时间读文件", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 89.718, "end": 91.058, "text": "看工具结果", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 91.058, "end": 92.078, "text": "还有对话历史", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 92.078, "end": 94.998, "text": "Deepsea还表示在相同工作负载下", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 94.998, "end": 99.958, "text": "Global Attention Cache的GPU显存用量大约只有上一代的四分之一", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 99.958, "end": 103.498, "text": "Persistent Cache的存储占用只有八分之一", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 103.498, "end": 106.618, "text": "这些节省都适用于缓存的上下文", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 106.618, "end": 109.238, "text": "它们有助于解释更低的服务成本", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 109.238, "end": 112.198, "text": "但并不意味着整个模型一下子变小了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 112.198, "end": 116.078, "text": "原生视觉理解意味着你可以把图片和文字一起发过去", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 116.078, "end": 117.578, "text": "收到文字回复", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 117.578, "end": 121.398, "text": "我这里的测试主要是看它在文本提示下写代码的能力", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 121.398, "end": 124.158, "text": "所以视觉能力算是这次发布的一部分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 124.158, "end": 127.498, "text": "而不是我在这个视频里专门做过基准测试的内容", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 127.498, "end": 130.138, "text": "官方公布的结果也挺有说服力", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 130.138, "end": 134.038, "text": "DeepSeq报告的DeepSW1.1成绩是74.2", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 134.038, "end": 136.718, "text": "而V4 Pro是62.7", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 136.718, "end": 140.118, "text": "这里说的是他们在最高投入下的结果", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 140.118, "end": 142.498, "text": "在一些更难的基准测试上", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 142.498, "end": 143.898, "text": "其他模型还是领先", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 143.898, "end": 145.618, "text": "而且报告也显示", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 145.618, "end": 148.838, "text": "围绕模型的Agent软件会影响最终分数", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 148.838, "end": 152.078, "text": "所以我想看看它在开源代码里能做什么", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 152.078, "end": 152.778, "text": "现在", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 152.778, "end": 154.938, "text": "我得纠正一下我最初的印象", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 154.938, "end": 157.618, "text": "我原来那篇开源代码模型的条目里", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 157.618, "end": 159.378, "text": "Thinking是被禁用的", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 159.378, "end": 161.298, "text": "我一开始怀疑这是回退了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 161.298, "end": 161.598, "text": "不过", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 161.598, "end": 164.398, "text": "开启Thinking并把Effort调到最高后", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 164.398, "end": 167.878, "text": "分数从53.75提升到了81.25%", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 167.878, "end": 170.578, "text": "保存的绘画里有大量推理输出", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 170.578, "end": 172.298, "text": "生成结果也明显更好了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 172.298, "end": 174.858, "text": "这些测试是在今天发布前", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 174.858, "end": 176.738, "text": "使用临时的API模型跑的", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 176.738, "end": 178.778, "text": "我还没有用新的正式版", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 178.778, "end": 180.498, "text": "模型名重新跑一遍套件", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 180.498, "end": 183.158, "text": "这些是我上手预览的结果", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 183.158, "end": 184.838, "text": "外加官方发布信息", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 184.838, "end": 186.858, "text": "我没有证据表明", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 186.858, "end": 189.778, "text": "DIPS在第一轮偷偷用了一个更弱的版本", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 189.778, "end": 192.258, "text": "Teambench 3有8个任务", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 192.258, "end": 193.498, "text": "每个任务10分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 193.498, "end": 195.558, "text": "这是我的实际评分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 195.558, "end": 197.738, "text": "我还打开了生成出来的界面", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 197.738, "end": 199.138, "text": "检查了代码", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 199.138, "end": 200.418, "text": "并保留了原始输出", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 200.418, "end": 201.698, "text": "那我们直接开始", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 201.698, "end": 203.538, "text": "首先是电梯模拟", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 203.538, "end": 205.238, "text": "这里应该有三部电梯", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 205.238, "end": 206.398, "text": "每部在一个人", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 206.398, "end": 208.658, "text": "人们在不同楼层等待", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 208.658, "end": 210.178, "text": "乘坐空闲的电梯", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 210.178, "end": 212.398, "text": "前往各自的目的地", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 212.398, "end": 214.078, "text": "把鼠标悬停上去时", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 214.078, "end": 216.618, "text": "应该能显示他们想去哪里", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 216.618, "end": 217.958, "text": "看新的结果", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 217.958, "end": 218.998, "text": "楼洞布局很清楚", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 218.998, "end": 220.578, "text": "三部电梯警都显示出来了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 220.578, "end": 222.718, "text": "乘客也在各自楼层排队", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 222.718, "end": 224.678, "text": "你可以看到人们上车的过程", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 224.678, "end": 226.898, "text": "还能看到等待运行中和", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 226.898, "end": 228.118, "text": "以到达的技术在更新", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 228.118, "end": 229.898, "text": "在浏览器里检查时", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 229.898, "end": 231.518, "text": "生成了两批个24个人", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 231.518, "end": 233.358, "text": "两批加起来数量都对", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 233.358, "end": 235.218, "text": "48名乘客全部到达", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 235.218, "end": 237.858, "text": "而且任何时候都没有一部轿箱", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 237.858, "end": 239.258, "text": "同时在超过一个人", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 239.258, "end": 240.458, "text": "技术器", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 240.458, "end": 242.898, "text": "在这次运行里也一直保持一致", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 242.898, "end": 243.578, "text": "之前", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 243.578, "end": 245.678, "text": "舱势会消失在屏幕外", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 245.678, "end": 247.298, "text": "人也会被画到错误的位置", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 247.298, "end": 250.138, "text": "新代码还会清理已经完成的乘客元素", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 250.138, "end": 252.098, "text": "这对重复使用很有帮助", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 252.098, "end": 254.118, "text": "现在还有一个重置边缘情况", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 254.118, "end": 255.878, "text": "如果你在错开的生成", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 255.878, "end": 257.278, "text": "还在进行时清空模拟", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 257.278, "end": 259.978, "text": "之后可能还会出现一些带处理的乘客", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 259.978, "end": 262.518, "text": "但主要模拟已经能正常工作了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 262.518, "end": 265.138, "text": "这部分从三分提升到了十分里的八分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 265.138, "end": 267.318, "text": "第二个任务是隐形眼镜盒", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 267.318, "end": 268.898, "text": "我想要一个3D物体", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 268.898, "end": 270.638, "text": "L和耳标时要很明显", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 270.638, "end": 271.778, "text": "盖子点击后能打开", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 271.778, "end": 273.678, "text": "这个版本就可信多了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 273.678, "end": 275.478, "text": "现在有一个干净的白色和尘", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 275.478, "end": 276.858, "text": "薄荷绿的盖子", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 276.858, "end": 278.738, "text": "还有大号清晰可读的字母", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 278.738, "end": 280.738, "text": "两个盖子可以分别打开", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 280.738, "end": 282.458, "text": "而且两个都打开按钮", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 282.458, "end": 283.898, "text": "也确实如字面所示", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 283.898, "end": 285.818, "text": "你也可以再把它们关上", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 285.818, "end": 287.798, "text": "然后旋转这个物体来查看", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 287.798, "end": 289.418, "text": "之前的版本还有", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 289.418, "end": 290.918, "text": "字母有些破碎", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 290.918, "end": 292.298, "text": "但在盖子下面", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 292.298, "end": 294.058, "text": "能看到一个看起来很实的表面", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 294.058, "end": 296.558, "text": "这里的结构是围绕脚链组织的", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 296.558, "end": 297.858, "text": "所以动作逻辑就对了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 297.858, "end": 300.218, "text": "虽然内部空间看起来还是有点浅", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 300.218, "end": 302.858, "text": "而且默认镜头在盖子完全打开时", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 302.858, "end": 303.818, "text": "会把边缘裁掉", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 303.818, "end": 305.858, "text": "所以还需要再打磨一下", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 305.858, "end": 308.438, "text": "但这已经比提示词的理解好多了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 308.438, "end": 309.798, "text": "七分原来只有五分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 309.798, "end": 311.378, "text": "接下来是折叠桌", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 311.378, "end": 314.758, "text": "这在第一次生成时是最明显的失败之一", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 314.758, "end": 316.738, "text": "因为桌面是飘着的", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 316.738, "end": 317.978, "text": "桌腿还埋在地下", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 317.978, "end": 320.898, "text": "这次展开后的桌子稳稳地立着", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 320.898, "end": 322.138, "text": "能看到四条腿", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 322.138, "end": 323.818, "text": "木质表面也不错", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 323.818, "end": 325.238, "text": "滑块反应很顺", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 325.238, "end": 328.838, "text": "折叠部件也能作为一个整体保持在一起", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 328.838, "end": 331.998, "text": "看看这两个结果的差别就知道了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 331.998, "end": 335.498, "text": "这段代码把腿连接到了合适的运动部件上", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 335.498, "end": 337.618, "text": "解决了前面那个核心问题", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 337.618, "end": 340.718, "text": "不过在完全折叠的状态下还是会扣分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 340.718, "end": 344.258, "text": "腿部横杆会和叠在一起的桌面板块相交", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 344.258, "end": 348.398, "text": "所以整个机构并没有从头到尾都做到物理上干净", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 348.398, "end": 350.498, "text": "我给它六分 满分十分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 350.498, "end": 351.758, "text": "它确实有进步", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 351.758, "end": 355.478, "text": "但也很能说明模型还有哪些地方需要继续检查", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 355.478, "end": 358.338, "text": "第四个任务是熊猫吃汉堡的SVG", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 358.338, "end": 359.998, "text": "这个版本构图更干净", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 359.998, "end": 361.418, "text": "表情也更生动", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 361.418, "end": 362.778, "text": "汉堡的层次还挺清楚", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 362.778, "end": 365.378, "text": "爪子和一点点碎屑也增加了些趣味", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 365.378, "end": 366.518, "text": "它是一副紧凑", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 366.518, "end": 367.878, "text": "完整的插画", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 367.878, "end": 369.278, "text": "而且渲染的也正确", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 369.278, "end": 371.198, "text": "剩下的问题是动作表现", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 371.198, "end": 373.298, "text": "看起来更像熊猫抱着汉堡", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 373.298, "end": 374.318, "text": "而不是在咬它", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 374.318, "end": 376.578, "text": "所以我更喜欢这张插画", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 376.578, "end": 378.738, "text": "但它还没有完全抓住那个动作", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 378.738, "end": 380.538, "text": "这个提示词打了六分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 380.538, "end": 381.318, "text": "之前是五分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 381.758, "end": 384.658, "text": "现在这个弓箭游戏在视觉上是最明显的提升之一", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 384.658, "end": 386.558, "text": "我们有日落场景", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 386.558, "end": 387.898, "text": "背景里的群山", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 387.898, "end": 389.058, "text": "一个弓箭手", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 389.058, "end": 390.038, "text": "四个目标", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 390.038, "end": 392.698, "text": "还有一个清晰的计时器", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 392.698, "end": 394.598, "text": "拖动瞄准再松手", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 394.598, "end": 397.678, "text": "感觉更像真正的射箭游戏了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 397.678, "end": 400.018, "text": "核心射击流程是通的", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 400.018, "end": 402.278, "text": "在浏览器检查里", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 402.278, "end": 404.398, "text": "四只箭都命中了四个目标", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 404.398, "end": 407.378, "text": "目标被击中后也倒下了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 407.378, "end": 408.758, "text": "游戏内计时器显示", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 409.238, "end": 411.438, "text": "这一轮四箭用了6.10秒", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 411.438, "end": 413.498, "text": "这说的是游戏玩法表现", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 413.498, "end": 416.678, "text": "和模型生成游戏所花的时间是分开的", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 416.678, "end": 418.458, "text": "我的评分表上这是8分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 418.458, "end": 419.838, "text": "比之前的5分高了不少", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 419.838, "end": 421.938, "text": "不过代码审查也发现了一个", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 421.938, "end": 422.998, "text": "我必须说明的半", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 422.998, "end": 425.398, "text": "这样就不会弹出结果页面了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 425.398, "end": 426.978, "text": "所以在那次检查里", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 426.978, "end": 428.538, "text": "分数没能保存", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 428.538, "end": 430.838, "text": "射击分为合命中反馈都不错", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 430.838, "end": 432.778, "text": "但完成流程还需要修", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 432.778, "end": 434.798, "text": "这也是为什么不能只看低评", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 434.798, "end": 437.638, "text": "第六个任务是那个数学题", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 437.638, "end": 440.078, "text": "考的是在受限移动条件下", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 440.078, "end": 442.558, "text": "统计有序对的排列数", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 442.558, "end": 445.138, "text": "预期答案是20460", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 445.138, "end": 447.238, "text": "这次保存的绘画里", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 447.238, "end": 448.538, "text": "最终答案很清楚", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 448.538, "end": 449.738, "text": "而且是对的", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 449.738, "end": 451.898, "text": "还有一个SOVER可以再跑一遍", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 451.898, "end": 453.238, "text": "我又跑了一次", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 453.238, "end": 455.638, "text": "结果得到了20460", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 455.638, "end": 458.978, "text": "而且和他检查过的小样本上的暴力枚举", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 458.978, "end": 459.738, "text": "结果一致", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 459.738, "end": 463.578, "text": "模型发现了自己最初转移逻辑里的一个错误", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 463.578, "end": 465.138, "text": "把它改正了", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 465.138, "end": 467.218, "text": "还先拿小规模暴力样例", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 467.218, "end": 468.698, "text": "验证了修正后的算法", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 468.698, "end": 470.738, "text": "然后才给出最终答案", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 470.738, "end": 472.738, "text": "这种解题行为非常有用", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 472.738, "end": 475.058, "text": "所以这是十分里的十分", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 475.058, "end": 477.958, "text": "现在还有一条更清晰的证据", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 477.958, "end": 479.438, "text": "练来支持这个结果", "language": "zh", "chunk": 0, "local_language_run": 8, "language_run": 8 }, { "start": 479.869, "end": 482.689, "text": "第七个任务是本地Panda微调工作流", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 482.689, "end": 484.689, "text": "我让它创建数据集", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 484.689, "end": 486.309, "text": "微调一个JM2B模型", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 486.309, "end": 488.509, "text": "再做一个本地网页界面", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 488.509, "end": 490.669, "text": "刷新时生成一条Panda冷知识", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 490.669, "end": 492.109, "text": "结果很不错", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 492.109, "end": 495.909, "text": "这个项目里有一个完成了700次迭代的LORA训练", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 495.909, "end": 497.409, "text": "保持了Adductors", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 497.409, "end": 499.669, "text": "还有一个能正常运行的本地应用", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 499.669, "end": 502.629, "text": "数据集里有151条人工编写的事实", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 502.629, "end": 504.249, "text": "之前版本只有127条", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 504.249, "end": 506.069, "text": "服务方式也更简单了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 506.069, "end": 508.889, "text": "模型和Adductors通过MLX直接加载", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 508.889, "end": 511.409, "text": "并使用Tokenizer的Chat Template", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 511.409, "end": 513.689, "text": "这样就去掉了上一个方案里", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 513.689, "end": 515.489, "text": "额外的导出步骤和手写模板层", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 515.489, "end": 516.929, "text": "界面也挺好看", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 516.929, "end": 518.529, "text": "有一张简单的Panda插画", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 518.529, "end": 520.349, "text": "一个清晰的事实卡片", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 520.349, "end": 522.169, "text": "还有一个生成新试试的按钮", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 522.169, "end": 525.129, "text": "布局在手机上也能正常显示", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 525.129, "end": 527.529, "text": "刷新页面和点击按钮", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 527.529, "end": 530.069, "text": "都可以生成新的内容", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 530.069, "end": 533.269, "text": "在评审过程中模型产生了回复", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 533.269, "end": 535.449, "text": "五个样本身程度很干净", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 535.449, "end": 537.369, "text": "是完整的单据", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 537.369, "end": 540.229, "text": "没有我们之前看到的那些多余标记", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 540.229, "end": 543.349, "text": "两次本地直连请求各花了大约半秒", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 543.349, "end": 546.049, "text": "这是本地微调后的GEM模型在回答", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 546.049, "end": 548.709, "text": "和DeepSeek的API生成速度是分开的", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 548.709, "end": 550.349, "text": "我把这个任务打了10分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 550.349, "end": 551.449, "text": "比之前的6分高", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 551.449, "end": 552.909, "text": "不过还是要说明", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 552.909, "end": 554.509, "text": "数据质量仍然有一些限制", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 554.509, "end": 555.869, "text": "所有验证答案", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 555.869, "end": 557.429, "text": "其实都已经出现在训练集中", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 557.429, "end": 559.509, "text": "测试文件是复制的训练性", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 559.509, "end": 562.349, "text": "而且有些人工编写的事实并不准确", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 562.349, "end": 563.649, "text": "在评审过程中", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 563.649, "end": 565.729, "text": "Refresh还重复了一条事实", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 565.729, "end": 567.849, "text": "所以完成整个工作流", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 567.849, "end": 569.329, "text": "并不能说明", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 569.329, "end": 571.729, "text": "每个生成出来的Panda结果都可靠", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 571.729, "end": 574.629, "text": "但能把训练和本地应用这样跑通", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 574.629, "end": 575.749, "text": "还是很厉害的", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 575.749, "end": 577.509, "text": "最后是这只手表", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 577.509, "end": 579.829, "text": "这大概是我改观的最佳视觉证明", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 579.829, "end": 582.529, "text": "第一版的表盘还是一整块金色表面", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 582.529, "end": 584.669, "text": "现在有了更细致的表盘", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 584.669, "end": 586.289, "text": "清晰的指标金属表带", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 586.289, "end": 588.329, "text": "能读的星期和日期窗口", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 588.329, "end": 589.189, "text": "汇动的指针", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 589.189, "end": 591.409, "text": "还有一个第二时区的小表盘", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 591.409, "end": 593.349, "text": "整体观感精致多了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 593.349, "end": 594.329, "text": "交互也没问题", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 594.329, "end": 596.569, "text": "你可以选两个时区互相切换", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 596.569, "end": 598.409, "text": "开关星期和日期显示", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 598.409, "end": 600.289, "text": "还能在扫秒和跳秒之间切换", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 600.289, "end": 602.149, "text": "控制检查也通过了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 602.149, "end": 605.169, "text": "说明设计下面那些关键部分也没问题", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 605.169, "end": 608.089, "text": "指针在午夜时能准确对齐", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 608.089, "end": 610.469, "text": "梦买的半小时偏一夜", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 610.469, "end": 611.969, "text": "有效了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 611.969, "end": 613.829, "text": "纽约的下令时切换也正常", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 613.829, "end": 616.469, "text": "第二个时区能正确滚到第二天", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 616.469, "end": 618.949, "text": "代码还附用了时间格式画起", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 618.949, "end": 622.169, "text": "避免像旧版本那样每一针都重复创建", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 622.169, "end": 625.649, "text": "视觉结构和时钟逻辑终于配合起来了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 625.649, "end": 628.089, "text": "我真的很喜欢这个结果", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 628.089, "end": 630.289, "text": "它从四分涨到了十分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 630.289, "end": 632.489, "text": "是复册里提升最大的一个", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 632.489, "end": 635.069, "text": "现在我们来看最后的图表", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 635.069, "end": 636.489, "text": "电梯8分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 636.489, "end": 637.769, "text": "隐形眼镜和7分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 637.769, "end": 638.509, "text": "桌子6分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 638.509, "end": 639.469, "text": "SVG6分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 639.469, "end": 640.269, "text": "射简8分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 640.269, "end": 641.489, "text": "数学", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 641.489, "end": 644.169, "text": "Fine Tuning Workflow和手表都是10分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 644.169, "end": 645.909, "text": "总分是65分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 645.909, "end": 646.749, "text": "满分80分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 646.749, "end": 648.349, "text": "也就是81.25%", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 648.349, "end": 651.309, "text": "不开启思考时的成绩是80分里的43分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 651.309, "end": 653.549, "text": "修正配置并重新跑任务后", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 653.549, "end": 654.909, "text": "我们竟提升了22分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 654.909, "end": 656.989, "text": "也就是27.5个百分点", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 656.989, "end": 660.589, "text": "这个新分数也高于我之前V4Flash的72.5%", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 660.589, "end": 663.349, "text": "以及V4Pro条目的76.25%", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 663.349, "end": 664.989, "text": "它在这个排行榜上", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 664.989, "end": 666.689, "text": "和Gemina3.8Flash打平", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 666.689, "end": 667.689, "text": "只比Open4", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 667.689, "end": 669.729, "text": "8的80%高一点", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 669.729, "end": 671.369, "text": "虽然没排到总榜第一", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 671.369, "end": 674.549, "text": "但这个结果已经非常有竞争力了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 674.549, "end": 675.649, "text": "再说速度", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 675.649, "end": 677.009, "text": "我看了OpenCode日志", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 677.009, "end": 679.489, "text": "它为这个新的测试文件夹", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 679.489, "end": 680.769, "text": "保存了绘画记录", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 680.769, "end": 683.449, "text": "共有64条已完成的模型回复", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 683.449, "end": 685.989, "text": "有可用的Token数和时间记录", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 685.989, "end": 687.489, "text": "像绘画标题", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 687.489, "end": 689.289, "text": "这类后台调用都被排除了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 689.289, "end": 690.989, "text": "另外还排除了一条", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 690.989, "end": 692.809, "text": "没完成的Restorch回复", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 692.809, "end": 694.829, "text": "在这些已完成的回复里", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 694.829, "end": 697.549, "text": "一共生成了364574个Token", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 697.549, "end": 700.469, "text": "大约294000个是Risning Tokens", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 700.469, "end": 701.489, "text": "另外大约70", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 701.489, "end": 704.209, "text": "000个是正常生成的Token", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 704.209, "end": 708.049, "text": "包括Code2调用参数和答案等其他输出", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 708.049, "end": 708.949, "text": "所以", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 708.949, "end": 712.809, "text": "这次生成的Token大多都用在了思考上", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 712.809, "end": 715.569, "text": "这两个数字都包含了Risning", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 715.569, "end": 718.049, "text": "它们是根据OpenCode的时间", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 718.049, "end": 719.709, "text": "戳重建出来的估算值", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 719.709, "end": 722.729, "text": "它们不包括后面执行工具花的时间", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 722.729, "end": 725.549, "text": "也不是服务器内部解码速度的精确测量", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 725.549, "end": 727.809, "text": "我把每次响应的时长加总了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 727.809, "end": 729.029, "text": "所以同时跑多个任务", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 729.029, "end": 731.369, "text": "不会把单次响应速率认为太高", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 731.369, "end": 733.109, "text": "我觉得这很厉害", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 733.109, "end": 734.909, "text": "他一边处理多个任务", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 734.909, "end": 737.629, "text": "一边花了几万tokens做listening", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 737.629, "end": 740.849, "text": "同时还能保持很高的生成速度", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 740.849, "end": 744.349, "text": "完整任务耗时能让你从另一个角度来看", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 744.349, "end": 747.509, "text": "电梯任务大概花了4分20秒", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 747.509, "end": 749.589, "text": "SVG任务大约11秒", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 749.589, "end": 752.889, "text": "本地翻桌的流程大约花了11分38秒", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 752.889, "end": 755.769, "text": "包括搭建环境训练和检查APP的时间", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 755.769, "end": 760.029, "text": "所以最高强度的思考可能意味着成品要等更久", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 760.029, "end": 761.429, "text": "在这次附色里", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 761.429, "end": 763.829, "text": "这些额外的工作带来了结果的大幅提升", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 763.829, "end": 765.509, "text": "这个取舍我很满意", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 765.509, "end": 768.629, "text": "现在API价格已经上线了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 768.629, "end": 769.649, "text": "非高峰时段", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 769.649, "end": 772.769, "text": "每百万未缓存输入token15美分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 772.769, "end": 775.409, "text": "每百万输出token60美分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 775.409, "end": 777.909, "text": "高峰时段输入是30美分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 777.909, "end": 779.689, "text": "输出是1.2美元", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 779.689, "end": 783.189, "text": "缓存输入在非高峰时段每百万Token", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 783.189, "end": 784.589, "text": "只要0.3美分", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 784.589, "end": 786.789, "text": "高峰时段价格翻倍", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 786.789, "end": 789.629, "text": "我会把当前的价格表面接放在下面", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 789.629, "end": 792.149, "text": "要使用这个已发布的模型", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 792.149, "end": 795.249, "text": "请选择DeepSeek API地址", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 795.249, "end": 799.149, "text": "DeepSeek在支持的工具里列出了OpenCode", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 799.149, "end": 801.129, "text": "记得把Thinking打开", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 801.129, "end": 802.709, "text": "并把Efford设为Max", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 802.709, "end": 805.469, "text": "如果你想看这次回测用的配置", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 805.469, "end": 807.869, "text": "旧的V4 Flash模型已经退役了", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 807.869, "end": 810.449, "text": "API名字会临时录游到V4.1 Flash", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 810.449, "end": 815.429, "text": "V4 Pro的请求会在9月14日UTC凌晨4点开始按Flash价格录游到V4", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 815.429, "end": 817.809, "text": "E Flash直到V4.1 Pro发布", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 817.809, "end": 819.829, "text": "如果你依赖Pro模型名", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 819.829, "end": 821.389, "text": "这个时间点就很重要", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 821.389, "end": 825.109, "text": "Rizni也会消耗可计费的Output Tokens", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 825.109, "end": 829.389, "text": "所以即使把Effort设到最高还是会有成本", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 829.389, "end": 834.089, "text": "不过结合这个定价实测速度和81.25%的结果V4", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 834.089, "end": 837.249, "text": "E Flash还是一个很有吸引力的编程选择", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 837.869, "end": 840.629, "text": "DeepSea还分享了技术报告和模型仓库", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 840.629, "end": 842.349, "text": "并表示他正在于", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 842.349, "end": 845.229, "text": "开源社区在推理支持和部署选项方面", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 845.229, "end": 847.429, "text": "我的看法有了很大变化", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 847.429, "end": 849.469, "text": "因为Max开启了Thinking Mode", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 849.469, "end": 852.529, "text": "这个模型在我的测试里表现得非常不错", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 852.529, "end": 855.149, "text": "尤其是在万表和本地模型工作流上", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 855.149, "end": 857.169, "text": "电梯案例也很清楚地说明", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 857.169, "end": 858.649, "text": "改进不只是界面更好看", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 858.649, "end": 860.829, "text": "不过还有一些bug要处理", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 860.829, "end": 862.989, "text": "表格和SVG的分数较低", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 862.989, "end": 865.509, "text": "说明他并不是每个任务都一样强", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 865.509, "end": 867.949, "text": "但我还是会非常愿意重新看看", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 867.949, "end": 869.209, "text": "他在编码上的表现", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 869.209, "end": 871.789, "text": "配置错误让第一次印象不太公平", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 871.789, "end": 874.349, "text": "这次复色让我们看到了更真实的情况", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 874.349, "end": 876.289, "text": "总体来说挺酷的", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 876.289, "end": 877.009, "text": "总之", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 877.009, "end": 879.589, "text": "欢迎在评论区告诉我你的想法", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 879.589, "end": 881.569, "text": "如果你喜欢这个视频", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 881.569, "end": 883.129, "text": "可以通过Super Thanks打赏", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 883.129, "end": 885.529, "text": "或者点击中按钮成为会员", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 885.529, "end": 886.369, "text": "另外", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 886.369, "end": 887.729, "text": "给这个视频点个赞", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 887.729, "end": 888.949, "text": "订阅我的频道", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 888.949, "end": 890.489, "text": "我们下期见", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 }, { "start": 890.489, "end": 891.009, "text": "拜拜", "language": "zh", "chunk": 1, "local_language_run": 0, "language_run": 8 } ], "segmented_multilingual_transcription": { "status": "STABLE_SEGMENTED_MULTILINGUAL_OK", "created_at": "2026-09-11T13:41:52", "manifest": "/Users/sagawa/AI_WORK/video_notes/url/20260911_133614__DeepSeek_V4.1_Flash实测_200_TPS_碾压Astra_新架构也太猛了/segmented_work/chunk_manifest.json", "audio": "/Users/sagawa/AI_WORK/video_notes/url/20260911_133614__DeepSeek_V4.1_Flash实测_200_TPS_碾压Astra_新架构也太猛了/audio_16k_mono.wav", "audio_sha256": "5740dcf1aa65f1a2ea39e7ac4cc80d878122b84baada2824334e7d04a81e4900", "audio_duration": 891.8785, "model": "/Users/sagawa/AI_WORK/cache/huggingface/hub/models--mlx-community--whisper-large-v3-turbo/snapshots/a4aaeec0636e6fef84abdcbe3544cb2bf7e9f6fb", "language_profile": "/Users/sagawa/AI_WORK/video_notes/url/20260911_133614__DeepSeek_V4.1_Flash实测_200_TPS_碾压Astra_新架构也太猛了/multilingual_work/multilingual_language_profile.json", "dominant_languages": [ "zh" ], "suppressed_languages": [], "chunk_count": 2, "completed_chunks": 2, "resumed_chunks": 0, "merged_segment_count": 396, "detected_languages": [ "zh" ], "transcript_languages": [ "zh" ], "missing_languages": [], "unexpected_languages": [], "language_sequence": [ "zh" ], "speech_region_seconds": 860.248, "speech_covered_seconds": 856.746, "speech_coverage_ratio": 0.9959290809162008, "tail_repeat_run": 1, "tail_repeat_text": "大家好,欢迎来到新视频。", "chunk_needs_review": [], "qc_ok": true, "qc_errors": [], "complete_review_artifact": true } }