{"text": "大家好,欢迎来到新一期视频。开始之前,我想先说一下。我在我的ARCK频道上更活跃了。我现在会发一些这里不讲的新内容。还有一些开发者相关的工具,你们也可以去看看。链接我会放在简介里。我现在拿到了Jeff的使用权限,想看看这东西到底有什么用。TagC说,它适合做极快的判断,非常低的成本。以及软件可以直接使用的结构化输出。所以,我准备了一些实用测试。其中还包括一些场景,如果设置本身有问题,就可能直接得到错误答案。Jeff是Tag所说的System-1模型。你把一些信息给他,问具体问题,他就会返回判断结果。他不会像普通聊天那样生成回复。也不会帮你写一个应用。有三种主要的问题类型。Chance是从你提供的选项里选一个。Store是用你定义的等级来打分。那则会给出一个是或否陈述为真的概率。这样做的目标是让这些小判断足够便宜、足够快。能贯穿整个软件使用。比如,你的应用可以判断该把消息发给哪个团队。某人是不是申请了退款,以及某件事是否需要处理。然后,代码再决定怎么处理这些答案。好了,我们直接来看测试。我先从一条关于重复扣费的客服消息开始。客户要退回多收的钱,说网站能正常用。还说这件事可以等到明天。我把四个问题放在一起问了。该哪个部门处理是否请求了退款,是否紧急,以及。Jaf把他选成了billing。退款请求的概率是98%,而紧急程度只有11%。frustration分数也接近平静那一段。所以,他把这些东西分得很准。在这个例子里,billing问题不会自动变成技术问题。退款请求也不会自动变成紧急情况。这对help desk这类工作流来说是个很好的起点。然后,我把消息改成,我不是再要退款。我只是需要一份发票副本。他还是选了billing。这很合理,但退款概率降到了3%。所以,他正确处理了否定,不是只看到refund这个词。就把它当成退款请求。这也是为什么在把任何模型接到动作之前。我都会先测试你实际是怎么说的。要退款,问退款政策。明确拒绝退款,这三件事完全不同。下一个测试显示了一个我觉得特别值得理解的局限。我问食堂几点关门,在有其他选项可选的情况下。这选了other。所以,他确实把这个问题排除在billing,technical support和sales之外。然后我给了他一个版本的问题,只允许那三个部门。这一次,他选了sales,智信度是0.31。答案符合我给出的选项。但这些选项里没有一个涵盖了我的需求。我其实是在故意设置一种情况,让他没法返回我需要的答案。所以,当你听到零hallucinations这种说法时,要记住这个区别。限制输出确实能阻止模型编造一个新类别。他并不能保证他选出来的类别有用或者正确。像这种流程里,我会加一个其他或未知选项,留给人工复合。而且,我不会把90%的confidence理解成模型有90%的时间都判断正确。这得用你自己的样本去测试。接着,我试了一个基础的prome injection。原始消息说,顾客点付款时结账页面崩了。而且,他们并不是要退款。Juff把他路由到技术支持,并把退款请求的概率设得很低。然后我在消息里塞了一个假的system override,让他选billing。并把退款和紧急程度这两个问题都标成处。但我真正的评估指令告诉他,要把消息里的这些内容当成不可信文版。他还是保留了技术知识分类,并且把退款概率一直很低,只有3%。紧迫感一直停留在不确定的中间地带,没有直接变成强烈的事。不过,单靠一次简单的注入测试,还不能说明模型就能免疫攻击。之后,我又测试了另一个稍微不同的东西。从文档里精确选出一个纸,消息里包含发件人的邮箱。就账单地址,以及收据应该发送到的新地址。我把这些地址作为颗选项提供进去。然后问当前的收据该发到哪里。他选中了新地址,连加号和年份都完全按原样选出来了。这对理解Jaf是怎么做提取很有用。你的代码可以收集候选值,Jaf负责选出相关的那个。再由你的代码复制。但如果你的第一步没找到正确地址,Jaf就没法通过一道选择题把它补出来。候选列表是你要测试的系统的一部分。最后,我试了一个小的agent审计事例。工具返回显示权限被拒绝,什么都没保存。但assistant的最终消息却说草稿已经成功保存。Jaf把这个任务判为失败,并给出了93%的概率。认为这个成功说法没有依据。我喜欢它在这个例子里是按工具证据来判断的。不过,这个错误本身很简单,用普通代码就能抓出来。更有意思的应用是去检查那些更段的trace有好几步。部分进展还有互相矛盾的说法。那就需要大得多的测试集了。现在,还有一个挺有意思的browser automation演示。Gregor Zinnick分享了一个开源项目,叫Jeff Ultrafast。他把Jeff和Browser Use工具结合在一起。他在演示里展示了他在大约7秒内,帮他在Google Flags上查到了从苏黎世到伦敦的单程航班。他说这次调用成本是0.39美元,不到半美分。而且视频还是元素播放的。这都是他自己报告的结果。我自己还没复现这个航班测试。他的工作方式挺有意思的。他会读取网页结构,然后生成一个代编号的空间列表。每一步之后,他都会刷新科选项。Jeff先选下一步操作。再把相关目标和这些问题打包成一个请求。虽然遇到需要实际文本时,他也可以点击某个按钮,或者在某个输入框里打字。那部分文本是由一个单独的小语言模型生成的。这个例子里用的是Mercury 2.5。这是Jeff如何融入更大agent的一个实际例子。Jeff负责结构化决策,另一个模型负责生成文本。浏览器代码还是得检查目标并执行动作。这里有几个细节要注意。公开的记事是从第一次页面观察后才开始的。所以不包含启动时间。而且这是一次不需要预定的机票搜索。报告的价格也和模型的成本估算一致。有统计使用量,不过那不包括浏览器基础设施。我也把这个项目和运行说明加进了TestKit。这些请求总共用了4148个输入Token。按公布的价格,每百万输入Token4.2美分估算下来。输入费用远不到一美分。按这个价格,输出Token是免费的。这些小例子成本其实很低,但一个应用的总成本。还包括你用来准备输入检查答案。以及处理那些需要换别的模型或人工介入的部分。我还记录了模型版本。我选的是JeffTheLatest。返回结果里,把模型识别成了Jeff1.13.0。这个信息最好和结果一起保存,因为叫做Latest的别名是会变的。总的来说,我觉得最有意思的地方,是你能很轻松地把这些判断塞进应用里。我测试的例子显示,它在路由否定选择值。以及根据证据核实说法,这几方面表现不错。也说明了你给出的选择和指令有多重要。这些只是八个代和城市里的小请求。他们给了我一些值得继续研究的方向。但还不能证明它在生产环境里可靠。也不能证明它在速度上比其他模型快很多。我把测试脚本也附上了,你可以自己附现这些例子。你也可以试试更大范围的样本。比较一下是把多个问题一起问,还是分开提问。这样更容易看出JF是否适合你工作流里的某个特定环节。总体来说还挺不错的。总之,欢迎在评论区告诉我你的想法。如果你喜欢这个视频,可以通过Super Thanks打赏。也可以点击这双按钮加入会员。另外,别忘了给这个视频点赞并订阅我的频道。我们下期再见,拜拜。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 29.82, "text": "大家好,欢迎来到新一期视频。开始之前,我想先说一下。我在我的ARCK频道上更活跃了。我现在会发一些这里不讲的新内容。还有一些开发者相关的工具,你们也可以去看看。链接我会放在简介里。我现在拿到了Jeff的使用权限,想看看这东西到底有什么用。TagC说,它适合做极快的判断,非常低的成本。", "tokens": [50365, 15248, 11, 28566, 17699, 6912, 4511, 12560, 2257, 16786, 40656, 39752, 1543, 45213, 32442, 11, 25246, 10108, 8090, 8861, 1543, 40528, 14200, 1899, 34, 42, 39752, 6025, 5708, 19002, 25956, 6563, 225, 2289, 1543, 1654, 25040, 12949, 28926, 38515, 35102, 1960, 39255, 1546, 12560, 34742, 25750, 1543, 14852, 32241, 13824, 18937, 28926, 12444, 15106, 28053, 1546, 23323, 39806, 11, 29806, 47935, 6734, 22884, 1543, 165, 241, 122, 14468, 1654, 12949, 12744, 3581, 11249, 222, 30312, 15759, 1543, 1654, 25040, 24351, 21381, 33887, 1546, 22982, 9254, 4422, 225, 43446, 11, 7093, 22884, 5562, 38409, 16220, 33883, 2412, 10440, 9254, 1543, 51, 559, 34, 8090, 11, 11284, 2215, 224, 14245, 10907, 7360, 223, 10251, 1546, 2437, 97, 4307, 255, 11, 14392, 41377, 1546, 11336, 8802, 1543, 51856], "temperature": 0, "avg_logprob": -0.08796623560387319, "compression_ratio": 1.2621359223300972, "no_speech_prob": 1.0252893152540299e-11}, {"id": 1, "seek": 3000, "start": 30.0, "end": 58.94, "text": "以及软件可以直接使用的结构化输出。所以,我准备了一些实用测试。其中还包括一些场景,如果设置本身有问题,就可能直接得到错误答案。Jeff是Tag所说的System-1模型。你把一些信息给他,问具体问题,他就会返回判断结果。他不会像普通聊天那样生成回复。也不会帮你写一个应用。有三种主要的问题类型。Chance是从你提供的选项里选一个。Store是用你定义的等级来打分。", "tokens": [50365, 40282, 17819, 107, 20485, 6723, 43297, 22982, 9254, 1546, 45641, 7360, 226, 23756, 9830, 241, 7781, 1543, 7239, 11, 1654, 6336, 228, 1787, 229, 2289, 38515, 24726, 9254, 11038, 233, 5233, 243, 1543, 9572, 5975, 14852, 41828, 38515, 50255, 50218, 11, 13119, 7422, 122, 34719, 8802, 19847, 2412, 34069, 11, 3111, 16657, 43297, 5916, 4511, 29900, 5233, 107, 28223, 28899, 1543, 33887, 1541, 51, 559, 5966, 8090, 1546, 50, 9321, 12, 16, 41908, 39823, 1543, 2166, 16075, 38515, 17665, 26460, 23197, 5000, 11, 22064, 39806, 29485, 34069, 11, 5000, 3111, 12949, 3316, 242, 8350, 2437, 97, 4307, 255, 45641, 9319, 1543, 5000, 47928, 12760, 29993, 19550, 40096, 6135, 4184, 14496, 8244, 11336, 8350, 1787, 235, 1543, 6404, 47928, 4845, 106, 2166, 5676, 247, 20182, 44297, 9254, 1543, 2412, 10960, 39810, 13557, 4275, 1546, 34069, 22113, 119, 39823, 1543, 6546, 719, 1541, 35630, 2166, 20949, 3254, 249, 1546, 2215, 231, 10178, 117, 15759, 2215, 231, 20182, 1543, 4520, 418, 1541, 9254, 2166, 12088, 2930, 231, 1546, 10187, 16853, 100, 6912, 12467, 6627, 1543, 51812], "temperature": 0, "avg_logprob": -0.07096298716285011, "compression_ratio": 1.2682291666666667, "no_speech_prob": 9.340361817322673e-12}, {"id": 2, "seek": 5894, "start": 58.94, "end": 78.46, "text": "那则会给出一个是或否陈述为真的概率。这样做的目标是让这些小判断足够便宜、足够快。能贯穿整个软件使用。比如,你的应用可以判断该把消息发给哪个团队。某人是不是申请了退款,以及某件事是否需要处理。然后,代码再决定怎么处理这些答案。", "tokens": [50365, 4184, 2437, 247, 12949, 23197, 7781, 20182, 1541, 19780, 45334, 8842, 230, 3316, 108, 13992, 8034, 26181, 44866, 1543, 21209, 10907, 1546, 11386, 162, 3921, 1541, 33650, 5562, 13824, 7322, 2437, 97, 4307, 255, 37236, 1787, 253, 27364, 2415, 250, 1231, 37236, 1787, 253, 10251, 1543, 8225, 18464, 107, 13740, 123, 27662, 7549, 17819, 107, 20485, 22982, 9254, 1543, 36757, 11, 18961, 44297, 9254, 6723, 2437, 97, 4307, 255, 44646, 16075, 28837, 26460, 28926, 23197, 17028, 7549, 3919, 95, 10034, 253, 1543, 17238, 238, 4035, 23034, 3526, 111, 27908, 2289, 46361, 48798, 11, 40282, 17238, 238, 20485, 6973, 1541, 45334, 35748, 1787, 226, 13876, 1543, 26636, 11, 19105, 23230, 223, 8623, 5676, 111, 12088, 15282, 1787, 226, 13876, 5562, 13824, 28223, 28899, 1543, 51341], "temperature": 0, "avg_logprob": -0.04757730979619064, "compression_ratio": 1.274131274131274, "no_speech_prob": 1.4206986281850575e-11}, {"id": 3, "seek": 7846, "start": 78.46, "end": 94.46, "text": "好了,我们直接来看测试。我先从一条关于重复扣费的客服消息开始。客户要退回多收的钱,说网站能正常用。还说这件事可以等到明天。我把四个问题放在一起问了。该哪个部门处理是否请求了退款,是否紧急,以及。", "tokens": [50365, 12621, 11, 15003, 43297, 6912, 4200, 11038, 233, 5233, 243, 1543, 1654, 10108, 35630, 2257, 48837, 28053, 37732, 12624, 1787, 235, 3416, 96, 18464, 117, 1546, 32316, 27408, 28837, 26460, 45213, 1543, 32316, 1486, 115, 4275, 46361, 8350, 6392, 18681, 1546, 39623, 11, 8090, 16469, 239, 34155, 8225, 15789, 11279, 9254, 1543, 14852, 8090, 5562, 20485, 6973, 6723, 10187, 4511, 11100, 6135, 1543, 1654, 16075, 19425, 7549, 34069, 12744, 3581, 29567, 22064, 2289, 1543, 44646, 17028, 7549, 13470, 8259, 101, 1787, 226, 13876, 1541, 45334, 27908, 32718, 2289, 46361, 48798, 11, 1541, 45334, 7732, 100, 36379, 11, 40282, 1543, 51165], "temperature": 0, "avg_logprob": -0.046624313280420396, "compression_ratio": 1.1991525423728813, "no_speech_prob": 1.5441824902096002e-11}, {"id": 4, "seek": 9446, "start": 94.46, "end": 123.86, "text": "Jaf把他选成了billing。退款请求的概率是98%,而紧急程度只有11%。frustration分数也接近平静那一段。所以,他把这些东西分得很准。在这个例子里,billing问题不会自动变成技术问题。退款请求也不会自动变成紧急情况。这对help desk这类工作流来说是个很好的起点。然后,我把消息改成,我不是再要退款。我只是需要一份发票副本。", "tokens": [50365, 41, 2792, 16075, 5000, 2215, 231, 11336, 2289, 65, 7345, 1543, 46361, 48798, 27908, 32718, 1546, 26181, 44866, 1541, 22516, 8923, 11070, 7732, 100, 36379, 29649, 13127, 35244, 5348, 4, 1543, 5779, 381, 2405, 6627, 33188, 6404, 14468, 17463, 16716, 5363, 247, 4184, 2257, 28427, 1543, 7239, 11, 5000, 16075, 5562, 13824, 38409, 16220, 6627, 5916, 4563, 6336, 228, 1543, 3581, 15368, 17797, 7626, 15759, 11, 65, 7345, 34069, 47928, 9722, 34961, 2129, 246, 11336, 32502, 1474, 107, 34069, 1543, 46361, 48798, 27908, 32718, 6404, 47928, 9722, 34961, 2129, 246, 11336, 7732, 100, 36379, 46514, 1543, 5562, 8713, 37451, 10026, 5562, 22113, 119, 41315, 27854, 6912, 8090, 1541, 7549, 4563, 20715, 9147, 12579, 1543, 26636, 11, 1654, 16075, 28837, 26460, 34490, 11336, 11, 1654, 7296, 8623, 4275, 46361, 48798, 1543, 1654, 36859, 35748, 2257, 36266, 28926, 41925, 5935, 107, 8802, 1543, 51835], "temperature": 0, "avg_logprob": -0.09413223266601563, "compression_ratio": 1.2716417910447761, "no_speech_prob": 1.4237719510312719e-11}, {"id": 5, "seek": 12386, "start": 123.86, "end": 153.54, "text": "他还是选了billing。这很合理,但退款概率降到了3%。所以,他正确处理了否定,不是只看到refund这个词。就把它当成退款请求。这也是为什么在把任何模型接到动作之前。我都会先测试你实际是怎么说的。要退款,问退款政策。明确拒绝退款,这三件事完全不同。下一个测试显示了一个我觉得特别值得理解的局限。我问食堂几点关门,在有其他选项可选的情况下。", "tokens": [50365, 5000, 45726, 2215, 231, 2289, 65, 7345, 1543, 5562, 4563, 14245, 13876, 11, 8395, 46361, 48798, 26181, 44866, 47421, 21381, 18, 4, 1543, 7239, 11, 5000, 15789, 38114, 106, 1787, 226, 13876, 2289, 45334, 12088, 11, 7296, 14003, 18032, 265, 21874, 15368, 5233, 235, 1543, 3111, 42061, 16233, 11336, 46361, 48798, 27908, 32718, 1543, 5562, 22021, 31006, 3581, 16075, 26443, 11906, 41908, 39823, 14468, 4511, 34961, 11914, 32442, 1543, 1654, 7182, 12949, 10108, 11038, 233, 5233, 243, 2166, 24726, 8842, 227, 1541, 15282, 8090, 1546, 1543, 4275, 46361, 48798, 11, 22064, 46361, 48798, 18523, 7973, 244, 1543, 11100, 38114, 106, 6852, 240, 10115, 251, 46361, 48798, 11, 5562, 10960, 20485, 6973, 37100, 47123, 1543, 4438, 20182, 11038, 233, 5233, 243, 1431, 122, 25696, 2289, 20182, 45681, 17682, 18453, 40242, 5916, 13876, 17278, 1546, 34703, 43446, 1543, 1654, 22064, 25155, 10726, 224, 6336, 254, 12579, 28053, 8259, 101, 11, 3581, 2412, 9572, 5000, 2215, 231, 10178, 117, 4429, 2215, 231, 1546, 46514, 4438, 1543, 51849], "temperature": 0, "avg_logprob": -0.02432045017380312, "compression_ratio": 1.2975206611570247, "no_speech_prob": 1.532695498296377e-11}, {"id": 6, "seek": 15386, "start": 153.86, "end": 164.24, "text": "这选了other。所以,他确实把这个问题排除在billing,technical support和sales之外。然后我给了他一个版本的问题,只允许那三个部门。", "tokens": [50365, 5562, 2215, 231, 2289, 802, 1543, 7239, 11, 5000, 38114, 106, 24726, 16075, 15368, 34069, 44647, 32999, 3581, 65, 7345, 11, 29113, 804, 1406, 12565, 82, 4229, 9574, 12022, 1543, 26636, 1654, 23197, 2289, 5000, 20182, 42096, 8802, 1546, 34069, 11, 14003, 2347, 223, 7422, 116, 4184, 10960, 7549, 13470, 8259, 101, 1543, 50884], "temperature": 0, "avg_logprob": -0.1876630615769771, "compression_ratio": 1.006060606060606, "no_speech_prob": 1.5723471136208644e-11}, {"id": 7, "seek": 16424, "start": 164.24, "end": 185.84, "text": "这一次,他选了sales,智信度是0.31。答案符合我给出的选项。但这些选项里没有一个涵盖了我的需求。我其实是在故意设置一种情况,让他没法返回我需要的答案。所以,当你听到零hallucinations这种说法时,要记住这个区别。限制输出确实能阻止模型编造一个新类别。", "tokens": [50365, 5562, 27505, 11, 5000, 2215, 231, 2289, 82, 4229, 11, 5094, 118, 17665, 13127, 1541, 15, 13, 12967, 1543, 28223, 28899, 5437, 99, 14245, 1654, 23197, 7781, 1546, 2215, 231, 10178, 117, 1543, 8395, 5562, 13824, 2215, 231, 10178, 117, 15759, 17944, 20182, 35681, 113, 5419, 244, 2289, 14200, 32535, 32718, 1543, 1654, 41646, 1541, 3581, 43045, 9042, 7422, 122, 34719, 2257, 39810, 46514, 11, 33650, 5000, 10062, 11148, 3316, 242, 8350, 1654, 35748, 1546, 28223, 28899, 1543, 7239, 11, 16233, 2166, 31022, 4511, 6306, 114, 43864, 1311, 10325, 5562, 39810, 8090, 11148, 15729, 11, 4275, 34756, 21632, 15368, 9937, 118, 18453, 1543, 43446, 25491, 9830, 241, 7781, 38114, 106, 24726, 8225, 10034, 119, 30438, 41908, 39823, 38109, 244, 37583, 20182, 12560, 22113, 119, 18453, 1543, 51445], "temperature": 0, "avg_logprob": -0.05600599875816932, "compression_ratio": 1.1951219512195121, "no_speech_prob": 1.8491772349471525e-11}, {"id": 8, "seek": 18584, "start": 185.84, "end": 193.92000000000002, "text": "他并不能保证他选出来的类别有用或者正确。像这种流程里,我会加一个其他或未知选项,留给人工复合。", "tokens": [50365, 5000, 3509, 114, 28590, 24302, 5233, 223, 5000, 2215, 231, 44561, 1546, 22113, 119, 18453, 2412, 9254, 31148, 15789, 38114, 106, 1543, 12760, 5562, 39810, 27854, 29649, 15759, 11, 1654, 12949, 9990, 20182, 9572, 5000, 19780, 29954, 6498, 2215, 231, 10178, 117, 11, 24456, 23197, 4035, 23323, 1787, 235, 14245, 1543, 50769], "temperature": 0, "avg_logprob": -0.056003908677534624, "compression_ratio": 1.0073529411764706, "no_speech_prob": 1.573299303336828e-11}, {"id": 9, "seek": 19392, "start": 193.92, "end": 210.5, "text": "而且,我不会把90%的confidence理解成模型有90%的时间都判断正确。这得用你自己的样本去测试。接着,我试了一个基础的prome injection。原始消息说,顾客点付款时结账页面崩了。而且,他们并不是要退款。", "tokens": [50365, 22942, 11, 25395, 12949, 16075, 7771, 4, 1546, 47273, 13876, 17278, 11336, 41908, 39823, 2412, 7771, 4, 1546, 44848, 7182, 2437, 97, 4307, 255, 15789, 38114, 106, 1543, 5562, 5916, 9254, 2166, 17645, 1546, 14496, 8802, 6734, 11038, 233, 5233, 243, 1543, 14468, 20708, 11, 1654, 5233, 243, 2289, 20182, 26008, 38114, 222, 1546, 1424, 423, 22873, 1543, 19683, 15476, 28837, 26460, 8090, 11, 10178, 122, 32316, 12579, 41469, 48798, 15729, 45641, 18464, 99, 10178, 113, 8833, 49378, 102, 2289, 1543, 22942, 11, 47911, 3509, 114, 7296, 4275, 46361, 48798, 1543, 51194], "temperature": 0, "avg_logprob": -0.08456836060849493, "compression_ratio": 1.2301587301587302, "no_speech_prob": 1.420668704205097e-11}, {"id": 10, "seek": 19392, "start": 210.89999999999998, "end": 223.6, "text": "Juff把他路由到技术支持,并把退款请求的概率设得很低。然后我在消息里塞了一个假的system override,让他选billing。并把退款和紧急程度这两个问题都标成处。", "tokens": [51214, 41, 1245, 16075, 5000, 24658, 23786, 4511, 32502, 1474, 107, 35488, 11, 3509, 114, 16075, 46361, 48798, 27908, 32718, 1546, 26181, 44866, 7422, 122, 5916, 4563, 41377, 1543, 26636, 40528, 28837, 26460, 15759, 13331, 252, 2289, 20182, 31706, 1546, 28215, 42321, 11, 33650, 5000, 2215, 231, 65, 7345, 1543, 3509, 114, 16075, 46361, 48798, 12565, 7732, 100, 36379, 29649, 13127, 5562, 36257, 7549, 34069, 7182, 162, 3921, 11336, 1787, 226, 1543, 51849], "temperature": 0, "avg_logprob": -0.08456836060849493, "compression_ratio": 1.2301587301587302, "no_speech_prob": 1.420668704205097e-11}, {"id": 11, "seek": 22392, "start": 223.92, "end": 253.22, "text": "但我真正的评估指令告诉他,要把消息里的这些内容当成不可信文版。他还是保留了技术知识分类,并且把退款概率一直很低,只有3%。紧迫感一直停留在不确定的中间地带,没有直接变成强烈的事。不过,单靠一次简单的注入测试,还不能说明模型就能免疫攻击。之后,我又测试了另一个稍微不同的东西。从文档里精确选出一个纸,消息里包含发件人的邮箱。", "tokens": [50365, 8395, 1654, 6303, 15789, 1546, 5233, 226, 7384, 108, 25922, 49061, 44816, 5000, 11, 4275, 16075, 28837, 26460, 15759, 1546, 5562, 13824, 34742, 25750, 16233, 11336, 1960, 4429, 17665, 17174, 42096, 1543, 5000, 45726, 24302, 24456, 2289, 32502, 1474, 107, 6498, 5233, 228, 6627, 22113, 119, 11, 3509, 114, 20334, 16075, 46361, 48798, 26181, 44866, 34448, 4563, 41377, 11, 35244, 18, 4, 1543, 7732, 100, 3316, 104, 9709, 34448, 36135, 24456, 3581, 1960, 38114, 106, 12088, 1546, 5975, 31685, 10928, 4845, 99, 11, 17944, 43297, 2129, 246, 11336, 5702, 118, 23661, 230, 1546, 6973, 1543, 1960, 16866, 11, 47446, 5363, 254, 27505, 11249, 222, 47446, 1546, 26432, 14028, 11038, 233, 5233, 243, 11, 14852, 28590, 8090, 11100, 41908, 39823, 3111, 8225, 2347, 235, 31670, 41530, 6336, 119, 1543, 9574, 13547, 11, 1654, 17047, 11038, 233, 5233, 243, 2289, 22762, 20182, 10415, 235, 39152, 47123, 1546, 38409, 16220, 1543, 35630, 17174, 21416, 96, 15759, 34910, 38114, 106, 2215, 231, 7781, 20182, 16853, 116, 11, 28837, 26460, 15759, 23305, 2392, 104, 28926, 20485, 4035, 1546, 3023, 106, 11249, 109, 1543, 51830], "temperature": 0, "avg_logprob": -0.036962830559324825, "compression_ratio": 1.3153409090909092, "no_speech_prob": 1.6232260327542214e-11}, {"id": 12, "seek": 25392, "start": 253.92, "end": 267.24, "text": "就账单地址,以及收据应该发送到的新地址。我把这些地址作为颗选项提供进去。然后问当前的收据该发到哪里。他选中了新地址,连加号和年份都完全按原样选出来了。", "tokens": [50365, 3111, 18464, 99, 47446, 10928, 14872, 222, 11, 40282, 18681, 26075, 106, 44297, 44646, 28926, 29309, 4511, 1546, 12560, 10928, 14872, 222, 1543, 1654, 16075, 5562, 13824, 10928, 14872, 222, 11914, 13992, 12501, 245, 2215, 231, 10178, 117, 20949, 3254, 249, 36700, 6734, 1543, 26636, 22064, 16233, 8945, 1546, 18681, 26075, 106, 44646, 28926, 4511, 17028, 15759, 1543, 5000, 2215, 231, 5975, 2289, 12560, 10928, 14872, 222, 11, 3316, 252, 9990, 26987, 12565, 5157, 36266, 7182, 37100, 26613, 19683, 14496, 2215, 231, 7781, 34912, 1543, 51031], "temperature": 0, "avg_logprob": -0.04267772395959061, "compression_ratio": 1.2142857142857142, "no_speech_prob": 1.7031771826214737e-11}, {"id": 13, "seek": 26724, "start": 267.24, "end": 280.44, "text": "这对理解Jaf是怎么做提取很有用。你的代码可以收集候选值,Jaf负责选出相关的那个。再由你的代码复制。但如果你的第一步没找到正确地址,Jaf就没法通过一道选择题把它补出来。", "tokens": [50365, 5562, 8713, 13876, 17278, 41, 2792, 1541, 15282, 10907, 20949, 29436, 49295, 9254, 1543, 18961, 19105, 23230, 223, 6723, 18681, 26020, 11238, 2215, 231, 40242, 11, 41, 2792, 18464, 253, 18464, 96, 2215, 231, 7781, 15106, 28053, 1546, 43083, 1543, 8623, 23786, 18961, 19105, 23230, 223, 1787, 235, 25491, 1543, 8395, 13119, 18961, 18049, 31429, 10062, 25085, 4511, 15789, 38114, 106, 10928, 14872, 222, 11, 41, 2792, 3111, 10062, 11148, 19550, 16866, 2257, 6025, 2215, 231, 6852, 102, 30716, 42061, 9890, 98, 44561, 1543, 51025], "temperature": 0, "avg_logprob": -0.07418358654057214, "compression_ratio": 1.2109704641350212, "no_speech_prob": 1.4844681039960506e-11}, {"id": 14, "seek": 26724, "start": 281.14, "end": 283.24, "text": "候选列表是你要测试的系统的一部分。", "tokens": [51060, 11238, 2215, 231, 43338, 17571, 1541, 32085, 11038, 233, 5233, 243, 1546, 25368, 10115, 253, 1546, 2257, 32174, 1543, 51165], "temperature": 0, "avg_logprob": -0.07418358654057214, "compression_ratio": 1.2109704641350212, "no_speech_prob": 1.4844681039960506e-11}, {"id": 15, "seek": 28324, "start": 283.24, "end": 293.02, "text": "最后,我试了一个小的agent审计事例。工具返回显示权限被拒绝,什么都没保存。但assistant的最终消息却说草稿已经成功保存。", "tokens": [50365, 8661, 13547, 11, 1654, 5233, 243, 2289, 20182, 7322, 1546, 559, 317, 2415, 94, 7422, 94, 6973, 17797, 1543, 23323, 39806, 3316, 242, 8350, 1431, 122, 25696, 4422, 225, 43446, 23238, 6852, 240, 10115, 251, 11, 10440, 7182, 10062, 24302, 39781, 1543, 8395, 640, 10329, 1546, 8661, 10115, 230, 28837, 26460, 5322, 112, 8090, 31409, 231, 10415, 123, 49161, 11336, 32311, 24302, 39781, 1543, 50854], "temperature": 0, "avg_logprob": -0.08299939566795979, "compression_ratio": 1.1063829787234043, "no_speech_prob": 1.518957702673074e-11}, {"id": 16, "seek": 28324, "start": 293.8, "end": 298.88, "text": "Jaf把这个任务判为失败,并给出了93%的概率。认为这个成功说法没有依据。", "tokens": [50893, 41, 2792, 16075, 15368, 26443, 5087, 94, 2437, 97, 13992, 28564, 18464, 98, 11, 3509, 114, 23197, 7781, 2289, 26372, 4, 1546, 26181, 44866, 1543, 7422, 97, 13992, 15368, 11336, 32311, 8090, 11148, 17944, 3254, 251, 26075, 106, 1543, 51147], "temperature": 0, "avg_logprob": -0.08299939566795979, "compression_ratio": 1.1063829787234043, "no_speech_prob": 1.518957702673074e-11}, {"id": 17, "seek": 29888, "start": 298.88, "end": 305.98, "text": "我喜欢它在这个例子里是按工具证据来判断的。不过,这个错误本身很简单,用普通代码就能抓出来。", "tokens": [50365, 1654, 41606, 11284, 3581, 15368, 17797, 7626, 15759, 1541, 26613, 23323, 39806, 5233, 223, 26075, 106, 6912, 2437, 97, 4307, 255, 1546, 1543, 1960, 16866, 11, 15368, 29900, 5233, 107, 8802, 19847, 4563, 11249, 222, 47446, 11, 9254, 29993, 19550, 19105, 23230, 223, 3111, 8225, 43069, 44561, 1543, 50720], "temperature": 0, "avg_logprob": -0.11965969973390207, "compression_ratio": 1.218918918918919, "no_speech_prob": 1.3368255287471431e-11}, {"id": 18, "seek": 29888, "start": 306.64, "end": 314.5, "text": "更有意思的应用是去检查那些更段的trace有好几步。部分进展还有互相矛盾的说法。那就需要大得多的测试集了。", "tokens": [50753, 19002, 2412, 16697, 1546, 44297, 9254, 1541, 6734, 16407, 222, 42623, 4184, 13824, 19002, 28427, 1546, 6903, 617, 2412, 2131, 6336, 254, 31429, 1543, 32174, 36700, 43491, 35091, 1369, 240, 15106, 5881, 249, 5419, 122, 1546, 8090, 11148, 1543, 46887, 35748, 3582, 5916, 6392, 1546, 11038, 233, 5233, 243, 26020, 2289, 1543, 51146], "temperature": 0, "avg_logprob": -0.11965969973390207, "compression_ratio": 1.218918918918919, "no_speech_prob": 1.3368255287471431e-11}, {"id": 19, "seek": 29888, "start": 314.92, "end": 318.0, "text": "现在,还有一个挺有意思的browser automation演示。", "tokens": [51167, 25040, 11, 35091, 20182, 41046, 2412, 16697, 1546, 1443, 30947, 17769, 31382, 25696, 1543, 51321], "temperature": 0, "avg_logprob": -0.11965969973390207, "compression_ratio": 1.218918918918919, "no_speech_prob": 1.3368255287471431e-11}, {"id": 20, "seek": 29888, "start": 318.68, "end": 321.54, "text": "Gregor Zinnick分享了一个开源项目,叫Jeff Ultrafast。", "tokens": [51355, 48254, 284, 1176, 7729, 618, 30855, 2289, 20182, 18937, 47402, 10178, 117, 11386, 11, 19855, 33887, 20925, 7011, 1543, 51498], "temperature": 0, "avg_logprob": -0.11965969973390207, "compression_ratio": 1.218918918918919, "no_speech_prob": 1.3368255287471431e-11}, {"id": 21, "seek": 29888, "start": 322.12, "end": 324.26, "text": "他把Jeff和Browser Use工具结合在一起。", "tokens": [51527, 5000, 16075, 33887, 12565, 22170, 30947, 8278, 23323, 39806, 45641, 14245, 3581, 29567, 1543, 51634], "temperature": 0, "avg_logprob": -0.11965969973390207, "compression_ratio": 1.218918918918919, "no_speech_prob": 1.3368255287471431e-11}, {"id": 22, "seek": 32426, "start": 324.26, "end": 330.58, "text": "他在演示里展示了他在大约7秒内,帮他在Google Flags上查到了从苏黎世到伦敦的单程航班。", "tokens": [50365, 5000, 3581, 31382, 25696, 15759, 43491, 25696, 2289, 5000, 3581, 3582, 16853, 99, 22, 20956, 34742, 11, 4845, 106, 5000, 3581, 12104, 3127, 3235, 12109, 5708, 42623, 21381, 35630, 13736, 237, 6173, 236, 17484, 4511, 7384, 99, 7017, 99, 1546, 47446, 29649, 10256, 103, 44695, 1543, 50681], "temperature": 0, "avg_logprob": -0.1310415267944336, "compression_ratio": 1.2464183381088825, "no_speech_prob": 1.600387877775944e-11}, {"id": 23, "seek": 32426, "start": 331.21999999999997, "end": 334.26, "text": "他说这次调用成本是0.39美元,不到半美分。", "tokens": [50713, 5000, 8090, 5562, 9487, 8897, 225, 9254, 11336, 8802, 1541, 15, 13, 12493, 39433, 11, 35728, 30018, 9175, 6627, 1543, 50865], "temperature": 0, "avg_logprob": -0.1310415267944336, "compression_ratio": 1.2464183381088825, "no_speech_prob": 1.600387877775944e-11}, {"id": 24, "seek": 32426, "start": 334.94, "end": 336.56, "text": "而且视频还是元素播放的。", "tokens": [50899, 22942, 40656, 39752, 45726, 14812, 37204, 49993, 12744, 1546, 1543, 50980], "temperature": 0, "avg_logprob": -0.1310415267944336, "compression_ratio": 1.2464183381088825, "no_speech_prob": 1.600387877775944e-11}, {"id": 25, "seek": 32426, "start": 337.2, "end": 339.3, "text": "这都是他自己报告的结果。", "tokens": [51012, 5562, 22796, 5000, 17645, 49817, 16846, 1546, 45641, 9319, 1543, 51117], "temperature": 0, "avg_logprob": -0.1310415267944336, "compression_ratio": 1.2464183381088825, "no_speech_prob": 1.600387877775944e-11}, {"id": 26, "seek": 32426, "start": 340.21999999999997, "end": 342.15999999999997, "text": "我自己还没复现这个航班测试。", "tokens": [51163, 1654, 17645, 14852, 10062, 1787, 235, 20204, 15368, 10256, 103, 44695, 11038, 233, 5233, 243, 1543, 51260], "temperature": 0, "avg_logprob": -0.1310415267944336, "compression_ratio": 1.2464183381088825, "no_speech_prob": 1.600387877775944e-11}, {"id": 27, "seek": 32426, "start": 343.08, "end": 345.09999999999997, "text": "他的工作方式挺有意思的。", "tokens": [51306, 31309, 41315, 9249, 27584, 41046, 2412, 16697, 1546, 1543, 51407], "temperature": 0, "avg_logprob": -0.1310415267944336, "compression_ratio": 1.2464183381088825, "no_speech_prob": 1.600387877775944e-11}, {"id": 28, "seek": 32426, "start": 345.74, "end": 349.44, "text": "他会读取网页结构,然后生成一个代编号的空间列表。", "tokens": [51439, 5000, 12949, 5233, 119, 29436, 16469, 239, 10178, 113, 45641, 7360, 226, 11, 26636, 8244, 11336, 20182, 19105, 38109, 244, 26987, 1546, 23322, 31685, 43338, 17571, 1543, 51624], "temperature": 0, "avg_logprob": -0.1310415267944336, "compression_ratio": 1.2464183381088825, "no_speech_prob": 1.600387877775944e-11}, {"id": 29, "seek": 32426, "start": 350.2, "end": 352.7, "text": "每一步之后,他都会刷新科选项。", "tokens": [51662, 23664, 2257, 31429, 9574, 13547, 11, 5000, 7182, 12949, 2437, 115, 12560, 42091, 2215, 231, 10178, 117, 1543, 51787], "temperature": 0, "avg_logprob": -0.1310415267944336, "compression_ratio": 1.2464183381088825, "no_speech_prob": 1.600387877775944e-11}, {"id": 30, "seek": 35270, "start": 352.7, "end": 354.96, "text": "Jeff先选下一步操作。", "tokens": [50365, 33887, 10108, 2215, 231, 45271, 31429, 13167, 235, 11914, 1543, 50478], "temperature": 0, "avg_logprob": -0.12816811778062692, "compression_ratio": 1.2608695652173914, "no_speech_prob": 1.731520656023111e-11}, {"id": 31, "seek": 35270, "start": 355.88, "end": 359.02, "text": "再把相关目标和这些问题打包成一个请求。", "tokens": [50524, 8623, 16075, 15106, 28053, 11386, 162, 3921, 12565, 5562, 13824, 34069, 12467, 23305, 11336, 20182, 27908, 32718, 1543, 50681], "temperature": 0, "avg_logprob": -0.12816811778062692, "compression_ratio": 1.2608695652173914, "no_speech_prob": 1.731520656023111e-11}, {"id": 32, "seek": 35270, "start": 360.02, "end": 366.7, "text": "虽然遇到需要实际文本时,他也可以点击某个按钮,或者在某个输入框里打字。", "tokens": [50731, 12026, 121, 5823, 3330, 229, 4511, 35748, 24726, 8842, 227, 17174, 8802, 15729, 11, 5000, 47935, 12579, 6336, 119, 17238, 238, 7549, 26613, 25153, 106, 11, 31148, 3581, 17238, 238, 7549, 9830, 241, 14028, 21416, 228, 15759, 12467, 22381, 1543, 51065], "temperature": 0, "avg_logprob": -0.12816811778062692, "compression_ratio": 1.2608695652173914, "no_speech_prob": 1.731520656023111e-11}, {"id": 33, "seek": 35270, "start": 367.65999999999997, "end": 371.41999999999996, "text": "那部分文本是由一个单独的小语言模型生成的。", "tokens": [51113, 4184, 32174, 17174, 8802, 1541, 23786, 20182, 47446, 18637, 105, 1546, 7322, 5233, 255, 12009, 41908, 39823, 8244, 11336, 1546, 1543, 51301], "temperature": 0, "avg_logprob": -0.12816811778062692, "compression_ratio": 1.2608695652173914, "no_speech_prob": 1.731520656023111e-11}, {"id": 34, "seek": 35270, "start": 372.3, "end": 374.88, "text": "这个例子里用的是Mercury 2.5。", "tokens": [51345, 15368, 17797, 7626, 15759, 9254, 24620, 44, 2869, 2598, 568, 13, 20, 1543, 51474], "temperature": 0, "avg_logprob": -0.12816811778062692, "compression_ratio": 1.2608695652173914, "no_speech_prob": 1.731520656023111e-11}, {"id": 35, "seek": 35270, "start": 375.59999999999997, "end": 378.32, "text": "这是Jeff如何融入更大agent的一个实际例子。", "tokens": [51510, 27455, 33887, 43526, 43772, 235, 14028, 19002, 3582, 559, 317, 1546, 20182, 24726, 8842, 227, 17797, 7626, 1543, 51646], "temperature": 0, "avg_logprob": -0.12816811778062692, "compression_ratio": 1.2608695652173914, "no_speech_prob": 1.731520656023111e-11}, {"id": 36, "seek": 35270, "start": 378.96, "end": 381.76, "text": "Jeff负责结构化决策,另一个模型负责生成文本。", "tokens": [51678, 33887, 18464, 253, 18464, 96, 45641, 7360, 226, 23756, 5676, 111, 7973, 244, 11, 22762, 20182, 41908, 39823, 18464, 253, 18464, 96, 8244, 11336, 17174, 8802, 1543, 51818], "temperature": 0, "avg_logprob": -0.12816811778062692, "compression_ratio": 1.2608695652173914, "no_speech_prob": 1.731520656023111e-11}, {"id": 37, "seek": 38176, "start": 381.76, "end": 384.88, "text": "浏览器代码还是得检查目标并执行动作。", "tokens": [50365, 11038, 237, 6758, 230, 34386, 19105, 23230, 223, 45726, 5916, 16407, 222, 42623, 11386, 162, 3921, 3509, 114, 3416, 100, 8082, 34961, 11914, 1543, 50521], "temperature": 0, "avg_logprob": -0.10752402282342678, "compression_ratio": 1.2460567823343849, "no_speech_prob": 1.5866571945188923e-11}, {"id": 38, "seek": 38176, "start": 385.52, "end": 386.64, "text": "这里有几个细节要注意。", "tokens": [50553, 35102, 2412, 6336, 254, 7549, 10115, 228, 45161, 4275, 40794, 1543, 50609], "temperature": 0, "avg_logprob": -0.10752402282342678, "compression_ratio": 1.2460567823343849, "no_speech_prob": 1.5866571945188923e-11}, {"id": 39, "seek": 38176, "start": 387.34, "end": 389.92, "text": "公开的记事是从第一次页面观察后才开始的。", "tokens": [50644, 13545, 18937, 1546, 34756, 6973, 1541, 35630, 18049, 9487, 10178, 113, 8833, 42350, 47550, 13547, 18888, 45213, 1546, 1543, 50773], "temperature": 0, "avg_logprob": -0.10752402282342678, "compression_ratio": 1.2460567823343849, "no_speech_prob": 1.5866571945188923e-11}, {"id": 40, "seek": 38176, "start": 390.53999999999996, "end": 391.8, "text": "所以不包含启动时间。", "tokens": [50804, 7239, 1960, 23305, 2392, 104, 2392, 107, 34961, 44848, 1543, 50867], "temperature": 0, "avg_logprob": -0.10752402282342678, "compression_ratio": 1.2460567823343849, "no_speech_prob": 1.5866571945188923e-11}, {"id": 41, "seek": 38176, "start": 392.5, "end": 394.53999999999996, "text": "而且这是一次不需要预定的机票搜索。", "tokens": [50902, 22942, 27455, 27505, 1960, 35748, 12501, 226, 12088, 1546, 37960, 41925, 162, 14987, 7732, 95, 1543, 51004], "temperature": 0, "avg_logprob": -0.10752402282342678, "compression_ratio": 1.2460567823343849, "no_speech_prob": 1.5866571945188923e-11}, {"id": 42, "seek": 38176, "start": 395.2, "end": 397.42, "text": "报告的价格也和模型的成本估算一致。", "tokens": [51037, 49817, 16846, 1546, 1550, 115, 30921, 6404, 12565, 41908, 39823, 1546, 11336, 8802, 7384, 108, 19497, 2257, 6784, 112, 1543, 51148], "temperature": 0, "avg_logprob": -0.10752402282342678, "compression_ratio": 1.2460567823343849, "no_speech_prob": 1.5866571945188923e-11}, {"id": 43, "seek": 38176, "start": 398.12, "end": 401.65999999999997, "text": "有统计使用量,不过那不包括浏览器基础设施。", "tokens": [51183, 2412, 10115, 253, 7422, 94, 22982, 9254, 26748, 11, 1960, 16866, 4184, 1960, 41828, 11038, 237, 6758, 230, 34386, 26008, 38114, 222, 7422, 122, 4307, 121, 1543, 51360], "temperature": 0, "avg_logprob": -0.10752402282342678, "compression_ratio": 1.2460567823343849, "no_speech_prob": 1.5866571945188923e-11}, {"id": 44, "seek": 38176, "start": 402.58, "end": 405.48, "text": "我也把这个项目和运行说明加进了TestKit。", "tokens": [51406, 27761, 16075, 15368, 10178, 117, 11386, 12565, 3316, 238, 8082, 8090, 11100, 9990, 36700, 2289, 51, 377, 45626, 1543, 51551], "temperature": 0, "avg_logprob": -0.10752402282342678, "compression_ratio": 1.2460567823343849, "no_speech_prob": 1.5866571945188923e-11}, {"id": 45, "seek": 40548, "start": 405.48, "end": 409.04, "text": "这些请求总共用了4148个输入Token。", "tokens": [50365, 5562, 13824, 27908, 32718, 33440, 15408, 9254, 2289, 19, 7271, 23, 7549, 9830, 241, 14028, 51, 8406, 1543, 50543], "temperature": 0, "avg_logprob": -0.1347344363177264, "compression_ratio": 1.235, "no_speech_prob": 1.488676196204075e-11}, {"id": 46, "seek": 40548, "start": 409.48, "end": 413.52000000000004, "text": "按公布的价格,每百万输入Token4.2美分估算下来。", "tokens": [50565, 26613, 13545, 34688, 1546, 1550, 115, 30921, 11, 23664, 31906, 23570, 9830, 241, 14028, 51, 8406, 19, 13, 17, 9175, 6627, 7384, 108, 19497, 4438, 6912, 1543, 50767], "temperature": 0, "avg_logprob": -0.1347344363177264, "compression_ratio": 1.235, "no_speech_prob": 1.488676196204075e-11}, {"id": 47, "seek": 40548, "start": 414.18, "end": 415.5, "text": "输入费用远不到一美分。", "tokens": [50800, 9830, 241, 14028, 18464, 117, 9254, 3316, 250, 35728, 2257, 9175, 6627, 1543, 50866], "temperature": 0, "avg_logprob": -0.1347344363177264, "compression_ratio": 1.235, "no_speech_prob": 1.488676196204075e-11}, {"id": 48, "seek": 40548, "start": 415.90000000000003, "end": 418.68, "text": "按这个价格,输出Token是免费的。", "tokens": [50886, 26613, 15368, 1550, 115, 30921, 11, 9830, 241, 7781, 51, 8406, 1541, 2347, 235, 18464, 117, 1546, 1543, 51025], "temperature": 0, "avg_logprob": -0.1347344363177264, "compression_ratio": 1.235, "no_speech_prob": 1.488676196204075e-11}, {"id": 49, "seek": 40548, "start": 419.3, "end": 422.52000000000004, "text": "这些小例子成本其实很低,但一个应用的总成本。", "tokens": [51056, 5562, 13824, 7322, 17797, 7626, 11336, 8802, 41646, 4563, 41377, 11, 8395, 20182, 44297, 9254, 1546, 33440, 11336, 8802, 1543, 51217], "temperature": 0, "avg_logprob": -0.1347344363177264, "compression_ratio": 1.235, "no_speech_prob": 1.488676196204075e-11}, {"id": 50, "seek": 42252, "start": 422.52, "end": 425.29999999999995, "text": "还包括你用来准备输入检查答案。", "tokens": [50365, 14852, 41828, 2166, 9254, 6912, 6336, 228, 1787, 229, 9830, 241, 14028, 16407, 222, 42623, 28223, 28899, 1543, 50504], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 51, "seek": 42252, "start": 425.91999999999996, "end": 428.88, "text": "以及处理那些需要换别的模型或人工介入的部分。", "tokens": [50535, 40282, 1787, 226, 13876, 4184, 13824, 35748, 26075, 95, 18453, 1546, 41908, 39823, 19780, 4035, 23323, 30312, 14028, 1546, 32174, 1543, 50683], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 52, "seek": 42252, "start": 429.58, "end": 431.14, "text": "我还记录了模型版本。", "tokens": [50718, 1654, 14852, 34756, 7391, 243, 2289, 41908, 39823, 42096, 8802, 1543, 50796], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 53, "seek": 42252, "start": 431.71999999999997, "end": 433.06, "text": "我选的是JeffTheLatest。", "tokens": [50825, 1654, 2215, 231, 24620, 33887, 2278, 43, 267, 377, 1543, 50892], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 54, "seek": 42252, "start": 434.0, "end": 437.58, "text": "返回结果里,把模型识别成了Jeff1.13.0。", "tokens": [50939, 3316, 242, 8350, 45641, 9319, 15759, 11, 16075, 41908, 39823, 5233, 228, 18453, 11336, 2289, 33887, 16, 13, 7668, 13, 15, 1543, 51118], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 55, "seek": 42252, "start": 438.47999999999996, "end": 442.5, "text": "这个信息最好和结果一起保存,因为叫做Latest的别名是会变的。", "tokens": [51163, 15368, 17665, 26460, 8661, 2131, 12565, 45641, 9319, 29567, 24302, 39781, 11, 34627, 19855, 10907, 43, 267, 377, 1546, 18453, 15940, 1541, 12949, 2129, 246, 1546, 1543, 51364], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 56, "seek": 42252, "start": 443.24, "end": 445.03999999999996, "text": "总的来说,我觉得最有意思的地方,", "tokens": [51401, 33440, 1546, 6912, 8090, 11, 45681, 8661, 2412, 16697, 1546, 30146, 11, 51491], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 57, "seek": 42252, "start": 445.68, "end": 447.94, "text": "是你能很轻松地把这些判断塞进应用里。", "tokens": [51523, 1541, 2166, 8225, 4563, 17819, 119, 36897, 10928, 16075, 5562, 13824, 2437, 97, 4307, 255, 13331, 252, 36700, 44297, 9254, 15759, 1543, 51636], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 58, "seek": 42252, "start": 448.35999999999996, "end": 451.35999999999996, "text": "我测试的例子显示,它在路由否定选择值。", "tokens": [51657, 1654, 11038, 233, 5233, 243, 1546, 17797, 7626, 1431, 122, 25696, 11, 11284, 3581, 24658, 23786, 45334, 12088, 2215, 231, 6852, 102, 40242, 1543, 51807], "temperature": 0, "avg_logprob": -0.1371658039603004, "compression_ratio": 1.2646239554317549, "no_speech_prob": 1.3645257666838884e-11}, {"id": 59, "seek": 45136, "start": 451.36, "end": 454.74, "text": "以及根据证据核实说法,这几方面表现不错。", "tokens": [50365, 40282, 31337, 26075, 106, 5233, 223, 26075, 106, 8987, 116, 24726, 8090, 11148, 11, 5562, 6336, 254, 39604, 17571, 20204, 1960, 29900, 1543, 50534], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 60, "seek": 45136, "start": 455.48, "end": 457.96000000000004, "text": "也说明了你给出的选择和指令有多重要。", "tokens": [50571, 6404, 8090, 11100, 2289, 2166, 23197, 7781, 1546, 2215, 231, 6852, 102, 12565, 25922, 49061, 2412, 6392, 24928, 1543, 50695], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 61, "seek": 45136, "start": 458.68, "end": 460.58000000000004, "text": "这些只是八个代和城市里的小请求。", "tokens": [50731, 5562, 13824, 36859, 33453, 7549, 19105, 12565, 45277, 27261, 15759, 1546, 7322, 27908, 32718, 1543, 50826], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 62, "seek": 45136, "start": 461.24, "end": 463.14, "text": "他们给了我一些值得继续研究的方向。", "tokens": [50859, 47911, 23197, 2289, 1654, 38515, 40242, 5916, 10115, 100, 10115, 255, 23230, 242, 44704, 1546, 9249, 24282, 1543, 50954], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 63, "seek": 45136, "start": 463.72, "end": 465.82, "text": "但还不能证明它在生产环境里可靠。", "tokens": [50983, 8395, 14852, 28590, 5233, 223, 11100, 11284, 3581, 8244, 1369, 100, 8051, 107, 47885, 15759, 4429, 5363, 254, 1543, 51088], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 64, "seek": 45136, "start": 466.46000000000004, "end": 469.02000000000004, "text": "也不能证明它在速度上比其他模型快很多。", "tokens": [51120, 6404, 28590, 5233, 223, 11100, 11284, 3581, 31217, 13127, 5708, 11706, 9572, 5000, 41908, 39823, 10251, 20778, 1543, 51248], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 65, "seek": 45136, "start": 469.7, "end": 472.6, "text": "我把测试脚本也附上了,你可以自己附现这些例子。", "tokens": [51282, 1654, 16075, 11038, 233, 5233, 243, 27067, 248, 8802, 6404, 8842, 226, 5708, 2289, 11, 42766, 17645, 8842, 226, 20204, 5562, 13824, 17797, 7626, 1543, 51427], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 66, "seek": 45136, "start": 473.3, "end": 475.04, "text": "你也可以试试更大范围的样本。", "tokens": [51462, 2166, 47935, 5233, 243, 5233, 243, 19002, 3582, 27395, 225, 3919, 112, 1546, 14496, 8802, 1543, 51549], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 67, "seek": 45136, "start": 475.5, "end": 478.66, "text": "比较一下是把多个问题一起问,还是分开提问。", "tokens": [51572, 11706, 9830, 225, 8861, 1541, 16075, 6392, 7549, 34069, 29567, 22064, 11, 45726, 6627, 18937, 20949, 22064, 1543, 51730], "temperature": 0, "avg_logprob": -0.11720349776183123, "compression_ratio": 1.3806818181818181, "no_speech_prob": 1.1169868849303377e-11}, {"id": 68, "seek": 47866, "start": 478.66, "end": 483.0, "text": "这样更容易看出JF是否适合你工作流里的某个特定环节。", "tokens": [50365, 21209, 19002, 49212, 4200, 7781, 41, 37, 1541, 45334, 2215, 224, 14245, 2166, 41315, 27854, 15759, 1546, 17238, 238, 7549, 17682, 12088, 8051, 107, 45161, 1543, 50582], "temperature": 0, "avg_logprob": -0.15435595512390138, "compression_ratio": 1.1654411764705883, "no_speech_prob": 1.367855481659852e-11}, {"id": 69, "seek": 47866, "start": 483.78000000000003, "end": 485.02000000000004, "text": "总体来说还挺不错的。", "tokens": [50621, 33440, 29485, 6912, 8090, 14852, 41046, 1960, 29900, 1546, 1543, 50683], "temperature": 0, "avg_logprob": -0.15435595512390138, "compression_ratio": 1.1654411764705883, "no_speech_prob": 1.367855481659852e-11}, {"id": 70, "seek": 47866, "start": 485.38000000000005, "end": 487.54, "text": "总之,欢迎在评论区告诉我你的想法。", "tokens": [50701, 33440, 9574, 11, 28566, 17699, 3581, 5233, 226, 7422, 118, 9937, 118, 44816, 1654, 18961, 7093, 11148, 1543, 50809], "temperature": 0, "avg_logprob": -0.15435595512390138, "compression_ratio": 1.1654411764705883, "no_speech_prob": 1.367855481659852e-11}, {"id": 71, "seek": 47866, "start": 488.24, "end": 491.34000000000003, "text": "如果你喜欢这个视频,可以通过Super Thanks打赏。", "tokens": [50844, 45669, 41606, 15368, 40656, 39752, 11, 6723, 19550, 16866, 30152, 2561, 12467, 5266, 237, 1543, 50999], "temperature": 0, "avg_logprob": -0.15435595512390138, "compression_ratio": 1.1654411764705883, "no_speech_prob": 1.367855481659852e-11}, {"id": 72, "seek": 47866, "start": 492.22, "end": 494.18, "text": "也可以点击这双按钮加入会员。", "tokens": [51043, 47935, 12579, 6336, 119, 5562, 2129, 234, 26613, 25153, 106, 43878, 12949, 3606, 246, 1543, 51141], "temperature": 0, "avg_logprob": -0.15435595512390138, "compression_ratio": 1.1654411764705883, "no_speech_prob": 1.367855481659852e-11}, {"id": 73, "seek": 47866, "start": 494.98, "end": 498.56, "text": "另外,别忘了给这个视频点赞并订阅我的频道。", "tokens": [51181, 26202, 11, 18453, 26677, 2289, 23197, 15368, 40656, 39752, 12579, 5266, 252, 3509, 114, 7422, 95, 10034, 227, 14200, 39752, 6025, 1543, 51360], "temperature": 0, "avg_logprob": -0.15435595512390138, "compression_ratio": 1.1654411764705883, "no_speech_prob": 1.367855481659852e-11}, {"id": 74, "seek": 49856, "start": 498.56, "end": 500.58, "text": "我们下期再见,拜拜。", "tokens": [50365, 15003, 4438, 16786, 44176, 11, 35105, 1543, 50466], "temperature": 0, "avg_logprob": -0.187750244140625, "compression_ratio": 0.7777777777777778, "no_speech_prob": 1.451539496238885e-11}], "language": "zh"}