{"text": "最近,硅谷出了一个非常特别的模型,叫做JEFF。他发布以后,很快就在开发者圈里刷屏。但如果只看网上的随便信息,很容易产生两个误解。第一个误解是,JEFF又是一个准备去挑战CHIGBT Cloud的全能模型。另一个误解就是,他不就是一个普通的分类器吗?为什么突然有这么多人讨论?所以,今天这期内容,我把整个JEFF的来龙去脉都梳理一遍。然后想从八个方面给大家去讲清楚。第一个,JEFF到底是什么?第二个,他开源了没有?第三个,能不能写的?第四个,为什么突然就火了?第五个,他的优势在哪里?第六个,这条技术路线到底成不成立?第七个,开发者现在想用它来去做什么?第八个,如果你想试一试,又该怎么去上手?首先,如果理解JEFF可以从我经常推荐的一本书讲起,就是丹尼尔·卡尼曼的《思考快与曼》。书里,他把人的思考方式分成了两套系统。系统一负责去快速判断,看到前面有东西飞过来,我们可能就本能的会躲开,看到一个人的表情,我们就能很快的判断他是不是生气了。而系统二负责的是慢,做一道复杂的数学题,分析一份商业计划,或者权衡一个重要的决定,都需要去投入更多的注意。过去几年,大模型的发展其实主要都在不断的加强系统,模型要学会去理解复杂问题,展开推理,量用工具,然后生成一段完整的答案。而TypeSafe提出的问题就是,一个AI系统里的每一步都需要这样的认证思考吗?很多时候程序需要的其实不是一篇文章,而是一个判断,这个就是JEFF出现的原因。TypeSafe把它叫做System-1 Model,也就是这个屏幕上写的。这个名字呢,就借鉴了思考快一慢的概念,但更多的是一种产品的类比,它并不是说复现了人的直觉,而是专门去处理那些范围明确,需要快速完成判断的任务。你可以把JEFF理解成一个专门去给程序使用的决策模型。穿透大模型,接受一段文字,然后继续去生成文字。那JEFF呢,接受一段材料和几个体型定义好的问题,返回的是选项,是分数,以及相应的概率。比如说,用户发来一句话,我买的东西一个星期还没到,现在不想要了。去写回复,而是需要搞清楚几件事情。这个用户是在查物流,还是在准备退款。情绪是不是激烈,要不要去转给人工客服?JEFF呢,会从开发者给定的选项里面去做选择,并且告诉程序每一个答案有多大概率,那程序拿到结果以后,就可以继续执行对应的流程。所以,有一句话就可以很直观的去概括它了。CHATGPT负责写答案,JEFF负责去做判断。OK,那第二个问题,JEFF开源了没有?目前是没有的。Type Safe呢,没有公开JEFF的模型权重,和核心实现,开发者呢,主要还是通过云端API去调用它。所以它和LAMA这一类可以去下载权重自己部署的开源模型,不是一回事。它更像是一种模型服务,你把材料和问题发给它,它把决策结果反回来。那第三问题,就是它能不能去携代码。首先,它不能替你去写Python,不能生成网页,也不能帮你写邮件或者营销文案。更准确的讲,它就没有自由文本生成这个能力。但呢,这不代表它和编程没有关系。一个编程agent在执行任务的时候,可能要不断的判断下一步该对,读哪个文件,该掉用哪个工具,测试是不是通过,要不要继续修改。那Jeff不能生成代码,但可以去参与这些流程里的选择。所以他不是一个写代码的人,更像是站在流程里面去做不断分流的人。这也能解释他为什么就突然火起来。你看过去大家评价一个新的模型,习惯会看他会不会写代码,数学能力有多强,推理链到底有多长。而Jeff反过来去做了另一件事情,他主动放弃自由的文本生成,只优化判断,优化选择,优化打分。他正好击中了agent正在遇到的一个现实问题。我们经常会看到一个agent最后生成了一份报告,好像他只完成了一次任务,但这个过程中他可能做了成千上百次的选择。比如说这条搜索结果有没有用,这两份资料是不是在重复表达同一个观点,现在的信息够不够,下一步应该掉用工具,还是开始去写作,这段结论有没有证据的支持。如果每一个小判断都需要去掉用一次非常强的通用大模型,延迟和成本会一层层累积。更麻烦的是,大模型明明只需要返回一个选项,但可能会顺手的减法,但他正好击写一段解释或者输出一个不完全符合要求的json。那jef呢,试图就是把这些小的判断单独拿出来解决。所以这么一看下来,其实我们就知道了,他第一个优势就是快。按照type safe公开的信息,jef处理一次决策通常需要70-500毫秒。对于浏览器的操作、实时审核、agent路由这类连续决策任务,这个差别会直接影响到产品体验了。好,第二个优势呢,就是便宜。官方价格大约是每百万输入的token是4.2美分啊。单次调用省下来的钱可能没那么明显,但当系统每天都需要完成几万次几十万次判断时,成本差距就会无限的被放大。第三个优势就是输出稳定。如果开发者只给出了5个候选选项,他就只会在这5个选项里返回结果,不会突然说创造出第6个答案,也不需要程序再从一段自然语言里面去提取结论。但这里一定要区分两件事情,输出格式稳定不等于判断是一定正确的。jef不会在格式上去胡乱发挥,但他在这些什么情况下,就是说,如果开发者只给出了5个候选项,他完全可能在合法选项里面去选错,所谓不产生自由闻闻幻觉,解决的是接口问题,不是准确性的问题。第4个优势就是它会返回概率。普通分类器往往只告诉你答案是什么,那jef希望去把置信程度也交给程序,这样系统就可以设计不同的策略,把握很高,而且操作风险很低,可以自动执行。把握一般,那就交给更强的大模型做复合,把握太低的,或者涉及到退款付款,商库这类高风险动作的,就转给人工。另外,jef还可以读取一次材料,同时回答多个相互独立的问题。还是前面的客服消息,他可以同时判断用户意图,情绪,紧急程度,和是否需要人工介入,不用反复的去读取一段上下文。OK,那这条路线到底对还是不对?其实我的理解,在他适合的任务里面,这条路线肯定是成立的,大原模型通常要采用自回归生成,他需要一个token,接着一个token生成答案。这个机制呢,给予了他协作,编程和开放式推理的能力,但如果程序只需要一个不而直,或者一个没举项,完整走一遍文本生成的流程,确实比较重。而jef把输出空间提前的限制住,不再去生成任意文字,而是直接围绕候选答案去计算概率,他失去了自由表达能力,也因此换来了更低的延迟,更低的成本,和更容易被代码处理的结果。但他也不是所谓问题的答案,如果任务需要长链条推理,需要精确计算,需要理解复杂的日期关系,或者需要模型给出完整理由,jef都不是合适选择。他的官方文档也明确提出,他在算术记住日期比较多部推理,以及包含大量无关信息的任务上存在短板,他同样可能也会受到提示词注入的影响。所以讨论这条路线的时候,我觉得不能问jef和cloud更聪明,他们解决的就不是同一种问题。cloud和gbt擅长的是开放式任务,传统分类模型适合类别长期固定,数据充足,可以专门训练的任务,那jef想占据的,其实中间那一层,问题由开发者临时定义,但输出必须限制在明确的类型和选项里。这一层以前经常交给大模型,现在终于有人开始为他单独去设计模型了。另外开发者也已经在尝试把jef放进不少的场景里面了,想客服系统,来判断用户意图,把消息转到查询订单,退款,投诉,或者人工客服。其他系统可以用他先来判断检查回来的资料到底是不是相关的,先过滤掉无关内容,再把剩下的材料交给大模型。内容平台也可以用他来做垃圾信息,辱骂内容,风险文本的前置判断。agent可以用他来选择下一步调用哪个工具,或者决定任务,是不是已经完整。Blozer Use 社区还做了一个jef ultrafast浏览器agent。jef负责从网页元素去选择该点击的按钮和输入框,只有遇到真实需要填写的文字的时候,才去调用生成模型。他们的结果呢,一次疏离室到伦敦的航班搜索演示,只用了7.1秒。Type Safe自己也演示过让jef去玩Doom。输入不是游戏画面,而是血量、敌人位置、弹药数量等已经整理好的游戏状态。jef不断从有限动作中去做选择,实现每秒的多次决策。这些案例表面上差别很大,但他们其实有个共同点,任务边界清楚,后选动作明确,请求频率很高,而且程序不需要一段漂亮的文字,只需要知道下一步到底该做什么。这个才是jef真正适合的位置。如果你想自己上手,流程其实也不复杂,可以直接去用Type Safe提供的Pregground和API,也可以通过VSL AI Gateway或者CloudFair去调用。一次请求里主要可以放两个东西,一份是State,也就是模型做判断需要看到的材料,另一份是Questions,也就是你希望他回答了。也就是你希望他回答了问题,候选选项,或者评分范围。jef读完State以后,会返回每个问题的答案和概率,你的程序再根据概率去决定下一个动作。不过真正困难的地方,往往不是API的接通,而是去把你的问题给定义清楚。还是那句话,我买的东西一个星期没到,现在不想要了。如果选项只有物流的延迟和申请退款,知道单选题本身就设计得太好。物流延迟是原因,申请退款是诉求,它们可以同时成立。模型不会提开发者修复一套混乱的业务标准,给他的信息不完整,问题不明确,或者选项互相重叠,速度再快也没有用。所以比较稳妥的使用方法就是先选择一个范围很小,答案比较清楚的任务。哪一批已经有人工处理过的历史数据,让jef在旁边运行,暂时不要参加真实操作。观察他在哪些任务上稳定、智信度和真实正确率是不是匹配,再根据风险设定自动通过,模型复合和人工处理的预知。低风险,高频,可设销的动作,可以更积极地去做状况。而退款,资金,权限和删除数据这些操作,仍然去保留代码规则和人工审批。所以我认为,jef真正有意思的地方,不只是说把原本的一次调查做便宜了。他可能让原本因为成本太高而没有发生的判断,开始大量出现。当然更多判断不一定自动带来更好的结果,如果标准本身就错了,系统值在更快更便宜的重复错误。但我始终认为jef至少提出了一个很值得观察的方向。未来的软件未必需要一个模型包办所有事情,复杂推理可以交给强模型,内容生成交给生成模型,范围明确的小判断,再去交给jef这一类的决策模型,再配合代码规则和人工审核。不是每一道题,都需要一皮A篇作文,有些时候程序员真正需要的只是一个足够快,带着概率,可以立刻执行的答案。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 29.98, "text": "最近,硅谷出了一个非常特别的模型,叫做JEFF。他发布以后,很快就在开发者圈里刷屏。但如果只看网上的随便信息,很容易产生两个误解。第一个误解是,JEFF又是一个准备去挑战CHIGBT Cloud的全能模型。另一个误解就是,他不就是一个普通的分类器吗?为什么突然有这么多人讨论?所以,今天这期内容,我把整个JEFF的来龙去脉都梳理一遍。然后想从八个方面给大家去讲清楚。第一个,JEFF到底是什么?第二个,他开源了没有?第三个,能不能写的?", "tokens": [50365, 46777, 11, 38114, 227, 8897, 115, 7781, 2289, 20182, 14392, 17682, 18453, 1546, 41908, 39823, 11, 19855, 10907, 48474, 6345, 1543, 5000, 28926, 34688, 3588, 13547, 11, 4563, 10251, 3111, 3581, 18937, 28926, 12444, 2523, 230, 15759, 2437, 115, 9636, 237, 1543, 8395, 13119, 14003, 4200, 16469, 239, 5708, 1546, 10673, 237, 27364, 17665, 26460, 11, 4563, 49212, 1369, 100, 8244, 36257, 7549, 5233, 107, 17278, 1543, 18049, 7549, 5233, 107, 17278, 1541, 11, 48474, 6345, 17047, 1541, 20182, 6336, 228, 1787, 229, 6734, 49067, 1486, 246, 5462, 40, 8769, 51, 8061, 1546, 11319, 8225, 41908, 39823, 1543, 22762, 20182, 5233, 107, 17278, 5620, 11, 5000, 1960, 5620, 20182, 29993, 19550, 1546, 6627, 22113, 119, 34386, 14769, 30, 31006, 40859, 5823, 2412, 31889, 6392, 4035, 7422, 101, 7422, 118, 30, 7239, 11, 12074, 5562, 16786, 34742, 25750, 11, 1654, 16075, 27662, 7549, 48474, 6345, 1546, 6912, 28907, 247, 6734, 27067, 231, 7182, 19017, 111, 13876, 2257, 3330, 235, 1543, 26636, 7093, 35630, 33453, 7549, 39604, 23197, 6868, 6734, 39255, 38433, 1543, 18049, 7549, 11, 48474, 6345, 33883, 1541, 10440, 30, 19693, 7549, 11, 5000, 18937, 47402, 2289, 17944, 30, 35878, 7549, 11, 8225, 28590, 5676, 247, 1546, 30, 51864], "temperature": 0, "avg_logprob": -0.16139272433608326, "compression_ratio": 1.3380614657210401, "no_speech_prob": 9.431832918849192e-12}, {"id": 1, "seek": 3000, "start": 30.0, "end": 60.0, "text": "第四个,为什么突然就火了?第五个,他的优势在哪里?第六个,这条技术路线到底成不成立?第七个,开发者现在想用它来去做什么?第八个,如果你想试一试,又该怎么去上手?首先,如果理解JEFF可以从我经常推荐的一本书讲起,就是丹尼尔·卡尼曼的《思考快与曼》。书里,他把人的思考方式分成了两套系统。系统一负责去快速判断,看到前面有东西飞过来,我们可能就本能的会躲开,看到一个人的表情,我们就能很快的判断他是不是生气了。而系统二负责的是慢,", "tokens": [50365, 7536, 19425, 7549, 11, 31006, 40859, 5823, 3111, 26171, 2289, 30, 7536, 21001, 7549, 11, 31309, 7384, 246, 5087, 123, 3581, 17028, 15759, 30, 7536, 30566, 7549, 11, 5562, 48837, 32502, 1474, 107, 24658, 16853, 123, 33883, 11336, 1960, 11336, 24409, 30, 7536, 29920, 7549, 11, 18937, 28926, 12444, 25040, 7093, 9254, 11284, 6912, 6734, 10907, 10440, 30, 7536, 33453, 7549, 11, 13119, 38386, 5233, 243, 2257, 5233, 243, 11, 17047, 44646, 15282, 6734, 5708, 11389, 30, 36490, 11, 13119, 13876, 17278, 48474, 6345, 6723, 35630, 1654, 30276, 11279, 33597, 31409, 238, 1546, 2257, 8802, 2930, 99, 39255, 9147, 11, 5620, 940, 117, 1530, 120, 1530, 242, 14259, 32681, 1530, 120, 9531, 120, 1546, 9806, 8870, 26504, 10251, 940, 236, 9531, 120, 9782, 1543, 2930, 99, 15759, 11, 5000, 16075, 4035, 1546, 8870, 26504, 9249, 27584, 6627, 11336, 2289, 36257, 1881, 245, 25368, 10115, 253, 1543, 25368, 10115, 253, 2257, 18464, 253, 18464, 96, 6734, 10251, 31217, 2437, 97, 4307, 255, 11, 18032, 8945, 8833, 2412, 38409, 16220, 11808, 252, 16866, 6912, 11, 15003, 16657, 3111, 8802, 8225, 1546, 12949, 17514, 110, 18937, 11, 18032, 20182, 4035, 1546, 17571, 10570, 11, 15003, 3111, 8225, 4563, 10251, 1546, 2437, 97, 4307, 255, 5000, 23034, 8244, 42204, 2289, 1543, 11070, 25368, 10115, 253, 11217, 18464, 253, 18464, 96, 24620, 20645, 11], "temperature": 0, "avg_logprob": -0.0825046522045567, "compression_ratio": 1.3962264150943395, "no_speech_prob": 1.6303892264035724e-11}, {"id": 2, "seek": 6000, "start": 60.0, "end": 90.0, "text": "做一道复杂的数学题,分析一份商业计划,或者权衡一个重要的决定,都需要去投入更多的注意。过去几年,大模型的发展其实主要都在不断的加强系统,模型要学会去理解复杂问题,展开推理,量用工具,然后生成一段完整的答案。而TypeSafe提出的问题就是,一个AI系统里的每一步都需要这样的认证思考吗?很多时候程序需要的其实不是一篇文章,而是一个判断,这个就是JEFF出现的原因。TypeSafe把它叫做System-1 Model,也就是这个屏幕上写的。", "tokens": [50365, 10907, 2257, 6025, 1787, 235, 4422, 224, 1546, 33188, 29618, 30716, 11, 6627, 7360, 238, 2257, 36266, 45581, 940, 248, 7422, 94, 2437, 240, 11, 31148, 4422, 225, 9890, 94, 20182, 24928, 1546, 5676, 111, 12088, 11, 7182, 35748, 6734, 37094, 14028, 19002, 6392, 1546, 40794, 1543, 16866, 6734, 6336, 254, 5157, 11, 3582, 41908, 39823, 1546, 28926, 43491, 41646, 13557, 4275, 7182, 3581, 1960, 4307, 255, 1546, 9990, 5702, 118, 25368, 10115, 253, 11, 41908, 39823, 4275, 29618, 12949, 6734, 13876, 17278, 1787, 235, 4422, 224, 34069, 11, 43491, 18937, 33597, 13876, 11, 26748, 9254, 23323, 39806, 11, 26636, 8244, 11336, 2257, 28427, 14128, 27662, 1546, 28223, 28899, 1543, 11070, 22464, 494, 50, 24248, 20949, 7781, 1546, 34069, 5620, 11, 20182, 48698, 25368, 10115, 253, 15759, 1546, 23664, 2257, 31429, 7182, 35748, 21209, 1546, 7422, 97, 5233, 223, 8870, 26504, 14769, 30, 20778, 29111, 29649, 6346, 237, 35748, 1546, 41646, 7296, 2257, 20878, 229, 17174, 11957, 254, 11, 11070, 1541, 20182, 2437, 97, 4307, 255, 11, 15368, 5620, 48474, 6345, 7781, 20204, 1546, 42577, 1543, 22464, 494, 50, 24248, 42061, 19855, 10907, 50, 9321, 12, 16, 17105, 11, 6404, 5620, 15368, 9636, 237, 40211, 5708, 5676, 247, 1546, 1543], "temperature": 0, "avg_logprob": -0.079125711233309, "compression_ratio": 1.3130841121495327, "no_speech_prob": 1.411020172231714e-11}, {"id": 3, "seek": 9000, "start": 90.0, "end": 120.0, "text": "这个名字呢,就借鉴了思考快一慢的概念,但更多的是一种产品的类比,它并不是说复现了人的直觉,而是专门去处理那些范围明确,需要快速完成判断的任务。你可以把JEFF理解成一个专门去给程序使用的决策模型。穿透大模型,接受一段文字,然后继续去生成文字。那JEFF呢,接受一段材料和几个体型定义好的问题,返回的是选项,是分数,以及相应的概率。比如说,用户发来一句话,我买的东西一个星期还没到,现在不想要了。", "tokens": [50365, 15368, 15940, 22381, 6240, 11, 3111, 2122, 253, 21829, 112, 2289, 8870, 26504, 10251, 2257, 20645, 1546, 26181, 33679, 11, 8395, 19002, 6392, 24620, 2257, 39810, 1369, 100, 30246, 1546, 22113, 119, 11706, 11, 11284, 3509, 114, 7296, 8090, 1787, 235, 20204, 2289, 4035, 1546, 16186, 24447, 11, 11070, 1541, 940, 241, 8259, 101, 6734, 1787, 226, 13876, 4184, 13824, 27395, 225, 3919, 112, 11100, 38114, 106, 11, 35748, 10251, 31217, 41509, 2437, 97, 4307, 255, 1546, 26443, 5087, 94, 1543, 42766, 16075, 48474, 6345, 13876, 17278, 11336, 20182, 940, 241, 8259, 101, 6734, 23197, 29649, 6346, 237, 22982, 9254, 1546, 5676, 111, 7973, 244, 41908, 39823, 1543, 13740, 123, 45808, 3582, 41908, 39823, 11, 14468, 23151, 2257, 28427, 17174, 22381, 11, 26636, 10115, 100, 10115, 255, 6734, 8244, 11336, 17174, 22381, 1543, 4184, 48474, 6345, 6240, 11, 14468, 23151, 2257, 28427, 4422, 238, 33404, 12565, 6336, 254, 7549, 29485, 39823, 12088, 2930, 231, 20715, 34069, 11, 3316, 242, 8350, 24620, 2215, 231, 10178, 117, 11, 1541, 6627, 33188, 11, 40282, 15106, 44297, 1546, 26181, 44866, 1543, 36757, 8090, 11, 9254, 1486, 115, 28926, 6912, 2257, 34592, 21596, 11, 1654, 2930, 108, 1546, 38409, 16220, 20182, 20682, 16786, 14852, 10062, 4511, 11, 25040, 1960, 7093, 4275, 2289, 1543], "temperature": 0, "avg_logprob": -0.05609773454212007, "compression_ratio": 1.343980343980344, "no_speech_prob": 1.527736444295602e-11}, {"id": 4, "seek": 12000, "start": 120.0, "end": 150.0, "text": "去写回复,而是需要搞清楚几件事情。这个用户是在查物流,还是在准备退款。情绪是不是激烈,要不要去转给人工客服?JEFF呢,会从开发者给定的选项里面去做选择,并且告诉程序每一个答案有多大概率,那程序拿到结果以后,就可以继续执行对应的流程。所以,有一句话就可以很直观的去概括它了。CHATGPT负责写答案,JEFF负责去做判断。OK,那第二个问题,JEFF开源了没有?目前是没有的。Type Safe呢,没有公开JEFF的模型权重,", "tokens": [50365, 6734, 5676, 247, 8350, 1787, 235, 11, 11070, 1541, 35748, 41816, 38433, 6336, 254, 49459, 1543, 15368, 9254, 1486, 115, 1541, 3581, 42623, 23516, 27854, 11, 45726, 3581, 6336, 228, 1787, 229, 46361, 48798, 1543, 10570, 10115, 103, 23034, 26373, 222, 23661, 230, 11, 4275, 11962, 6734, 17819, 105, 23197, 4035, 23323, 32316, 27408, 30, 48474, 6345, 6240, 11, 12949, 35630, 18937, 28926, 12444, 23197, 12088, 1546, 2215, 231, 10178, 117, 15759, 8833, 6734, 10907, 2215, 231, 6852, 102, 11, 3509, 114, 20334, 44816, 29649, 6346, 237, 23664, 20182, 28223, 28899, 2412, 6392, 32044, 44866, 11, 4184, 29649, 6346, 237, 24351, 4511, 45641, 9319, 3588, 13547, 11, 36781, 10115, 100, 10115, 255, 3416, 100, 8082, 8713, 44297, 1546, 27854, 29649, 1543, 7239, 11, 32241, 34592, 21596, 36781, 4563, 16186, 42350, 1546, 6734, 26181, 40322, 11284, 2289, 1543, 5462, 2218, 38, 47, 51, 18464, 253, 18464, 96, 5676, 247, 28223, 28899, 11, 48474, 6345, 18464, 253, 18464, 96, 6734, 10907, 2437, 97, 4307, 255, 1543, 9443, 11, 4184, 19693, 7549, 34069, 11, 48474, 6345, 18937, 47402, 2289, 17944, 30, 39004, 1541, 17944, 1546, 1543, 22464, 494, 27030, 6240, 11, 17944, 13545, 18937, 48474, 6345, 1546, 41908, 39823, 4422, 225, 12624, 11], "temperature": 0, "avg_logprob": -0.10747670655203338, "compression_ratio": 1.2819905213270142, "no_speech_prob": 1.6965593860329697e-11}, {"id": 5, "seek": 15000, "start": 150.0, "end": 180.0, "text": "和核心实现,开发者呢,主要还是通过云端API去调用它。所以它和LAMA这一类可以去下载权重自己部署的开源模型,不是一回事。它更像是一种模型服务,你把材料和问题发给它,它把决策结果反回来。那第三问题,就是它能不能去携代码。首先,它不能替你去写Python,不能生成网页,也不能帮你写邮件或者营销文案。更准确的讲,它就没有自由文本生成这个能力。但呢,这不代表它和编程没有关系。一个编程agent在执行任务的时候,可能要不断的判断下一步该对,", "tokens": [50365, 12565, 8987, 116, 7945, 24726, 20204, 11, 18937, 28926, 12444, 6240, 11, 13557, 4275, 45726, 19550, 16866, 1369, 239, 11957, 107, 4715, 40, 6734, 8897, 225, 9254, 11284, 1543, 7239, 11284, 12565, 43, 38136, 5562, 2257, 22113, 119, 6723, 6734, 4438, 17819, 121, 4422, 225, 12624, 17645, 13470, 16469, 110, 1546, 18937, 47402, 41908, 39823, 11, 7296, 2257, 8350, 6973, 1543, 11284, 19002, 12760, 1541, 2257, 39810, 41908, 39823, 27408, 5087, 94, 11, 2166, 16075, 4422, 238, 33404, 12565, 34069, 28926, 23197, 11284, 11, 11284, 16075, 5676, 111, 7973, 244, 45641, 9319, 22138, 8350, 6912, 1543, 4184, 35878, 34069, 11, 5620, 11284, 8225, 28590, 6734, 22220, 118, 19105, 23230, 223, 1543, 36490, 11, 11284, 28590, 9531, 123, 2166, 6734, 5676, 247, 47, 88, 11943, 11, 28590, 8244, 11336, 16469, 239, 10178, 113, 11, 6404, 28590, 4845, 106, 2166, 5676, 247, 3023, 106, 20485, 31148, 10674, 98, 23049, 222, 17174, 28899, 1543, 19002, 6336, 228, 38114, 106, 1546, 39255, 11, 11284, 3111, 17944, 45233, 17174, 8802, 8244, 11336, 15368, 8225, 13486, 1543, 8395, 6240, 11, 5562, 1960, 19105, 17571, 11284, 12565, 38109, 244, 29649, 17944, 28053, 25368, 1543, 20182, 38109, 244, 29649, 559, 317, 3581, 3416, 100, 8082, 26443, 5087, 94, 49873, 11, 16657, 4275, 1960, 4307, 255, 1546, 2437, 97, 4307, 255, 45271, 31429, 44646, 8713, 11], "temperature": 0, "avg_logprob": -0.1155966841466895, "compression_ratio": 1.3363431151241536, "no_speech_prob": 1.6073089040280486e-11}, {"id": 6, "seek": 18000, "start": 180.0, "end": 210.0, "text": "读哪个文件,该掉用哪个工具,测试是不是通过,要不要继续修改。那Jeff不能生成代码,但可以去参与这些流程里的选择。所以他不是一个写代码的人,更像是站在流程里面去做不断分流的人。这也能解释他为什么就突然火起来。你看过去大家评价一个新的模型,习惯会看他会不会写代码,数学能力有多强,推理链到底有多长。而Jeff反过来去做了另一件事情,他主动放弃自由的文本生成,只优化判断,优化选择,优化打分。", "tokens": [50365, 5233, 119, 17028, 7549, 17174, 20485, 11, 44646, 29327, 9254, 17028, 7549, 23323, 39806, 11, 11038, 233, 5233, 243, 23034, 19550, 16866, 11, 4275, 11962, 10115, 100, 10115, 255, 7792, 106, 34490, 1543, 4184, 33887, 28590, 8244, 11336, 19105, 23230, 223, 11, 8395, 6723, 6734, 2129, 224, 940, 236, 5562, 13824, 27854, 29649, 15759, 1546, 2215, 231, 6852, 102, 1543, 7239, 5000, 7296, 20182, 5676, 247, 19105, 23230, 223, 29979, 11, 19002, 12760, 1541, 34155, 3581, 27854, 29649, 15759, 8833, 6734, 10907, 1960, 4307, 255, 6627, 27854, 29979, 1543, 5562, 6404, 8225, 17278, 5873, 232, 5000, 31006, 3111, 40859, 5823, 26171, 9147, 6912, 1543, 16529, 16866, 6734, 6868, 5233, 226, 1550, 115, 20182, 12560, 1546, 41908, 39823, 11, 2930, 254, 4799, 107, 12949, 4200, 5000, 12949, 47928, 5676, 247, 19105, 23230, 223, 11, 33188, 29618, 8225, 13486, 2412, 6392, 5702, 118, 11, 33597, 13876, 165, 241, 122, 33883, 2412, 6392, 32271, 1543, 11070, 33887, 22138, 16866, 6912, 6734, 10907, 2289, 22762, 2257, 49459, 11, 5000, 13557, 34961, 12744, 5702, 225, 45233, 1546, 17174, 8802, 8244, 11336, 11, 14003, 7384, 246, 23756, 2437, 97, 4307, 255, 11, 7384, 246, 23756, 2215, 231, 6852, 102, 11, 7384, 246, 23756, 12467, 6627, 1543], "temperature": 0, "avg_logprob": -0.06259891774394724, "compression_ratio": 1.4114583333333333, "no_speech_prob": 1.333915963797061e-11}, {"id": 7, "seek": 21000, "start": 210.0, "end": 240.0, "text": "他正好击中了agent正在遇到的一个现实问题。我们经常会看到一个agent最后生成了一份报告,好像他只完成了一次任务,但这个过程中他可能做了成千上百次的选择。比如说这条搜索结果有没有用,这两份资料是不是在重复表达同一个观点,现在的信息够不够,下一步应该掉用工具,还是开始去写作,这段结论有没有证据的支持。如果每一个小判断都需要去掉用一次非常强的通用大模型,延迟和成本会一层层累积。更麻烦的是,大模型明明只需要返回一个选项,但可能会顺手的减法,但他正好击", "tokens": [50365, 5000, 15789, 2131, 6336, 119, 5975, 2289, 559, 317, 15789, 3581, 3330, 229, 4511, 1546, 20182, 20204, 24726, 34069, 1543, 15003, 30276, 11279, 12949, 18032, 20182, 559, 317, 8661, 13547, 8244, 11336, 2289, 2257, 36266, 49817, 16846, 11, 33242, 5000, 14003, 41509, 2289, 27505, 26443, 5087, 94, 11, 8395, 15368, 16866, 29649, 5975, 5000, 16657, 10907, 2289, 11336, 20787, 5708, 31906, 9487, 1546, 2215, 231, 6852, 102, 1543, 36757, 8090, 5562, 48837, 162, 14987, 7732, 95, 45641, 9319, 2412, 17944, 9254, 11, 5562, 36257, 36266, 5266, 226, 33404, 23034, 3581, 12624, 1787, 235, 17571, 9830, 122, 13089, 20182, 42350, 12579, 11, 25040, 1546, 17665, 26460, 1787, 253, 1960, 1787, 253, 11, 45271, 31429, 44297, 44646, 29327, 9254, 23323, 39806, 11, 45726, 45213, 6734, 5676, 247, 11914, 11, 5562, 28427, 45641, 7422, 118, 2412, 17944, 5233, 223, 26075, 106, 1546, 35488, 1543, 13119, 23664, 20182, 7322, 2437, 97, 4307, 255, 7182, 35748, 6734, 29327, 9254, 27505, 14392, 5702, 118, 1546, 19550, 9254, 3582, 41908, 39823, 11, 17527, 114, 3316, 253, 12565, 11336, 8802, 12949, 2257, 9636, 224, 9636, 224, 43656, 8204, 107, 1543, 19002, 38999, 23661, 99, 24620, 11, 3582, 41908, 39823, 11100, 11100, 14003, 35748, 3316, 242, 8350, 20182, 2215, 231, 10178, 117, 11, 8395, 16657, 12949, 10178, 118, 11389, 1546, 6336, 237, 11148, 11, 8395, 5000, 15789, 2131, 6336, 119], "temperature": 0, "avg_logprob": -0.09256321860001226, "compression_ratio": 1.4260089686098654, "no_speech_prob": 1.3326023444448776e-11}, {"id": 8, "seek": 24000, "start": 240.0, "end": 269.18, "text": "写一段解释或者输出一个不完全符合要求的json。那jef呢,试图就是把这些小的判断单独拿出来解决。所以这么一看下来,其实我们就知道了,他第一个优势就是快。按照type safe公开的信息,jef处理一次决策通常需要70-500毫秒。对于浏览器的操作、实时审核、agent路由这类连续决策任务,这个差别会直接影响到产品体验了。好,第二个优势呢,就是便宜。官方价格大约是每百万输入的token是4.2美分啊。", "tokens": [50365, 5676, 247, 2257, 28427, 17278, 5873, 232, 31148, 9830, 241, 7781, 20182, 1960, 37100, 5437, 99, 14245, 4275, 32718, 1546, 73, 3015, 1543, 4184, 2884, 69, 6240, 11, 5233, 243, 3919, 122, 5620, 16075, 5562, 13824, 7322, 1546, 2437, 97, 4307, 255, 47446, 18637, 105, 24351, 44561, 17278, 5676, 111, 1543, 7239, 31889, 2257, 4200, 4438, 6912, 11, 41646, 15003, 3111, 7758, 2289, 11, 5000, 18049, 7549, 7384, 246, 5087, 123, 5620, 10251, 1543, 26613, 32150, 20467, 3273, 13545, 18937, 1546, 17665, 26460, 11, 2884, 69, 1787, 226, 13876, 27505, 5676, 111, 7973, 244, 19550, 11279, 35748, 5867, 12, 7526, 7256, 104, 20956, 1543, 8713, 37732, 11038, 237, 6758, 230, 34386, 1546, 13167, 235, 11914, 1231, 24726, 15729, 2415, 94, 8987, 116, 1231, 559, 317, 24658, 23786, 5562, 22113, 119, 3316, 252, 10115, 255, 5676, 111, 7973, 244, 26443, 5087, 94, 11, 15368, 21679, 18453, 12949, 43297, 16820, 3858, 235, 4511, 1369, 100, 30246, 29485, 49657, 234, 2289, 1543, 2131, 11, 19693, 7549, 7384, 246, 5087, 123, 6240, 11, 5620, 27364, 2415, 250, 1543, 31929, 9249, 1550, 115, 30921, 3582, 16853, 99, 1541, 23664, 31906, 23570, 9830, 241, 14028, 1546, 83, 8406, 1541, 19, 13, 17, 9175, 6627, 4905, 1543, 51824], "temperature": 0, "avg_logprob": -0.12471865672691196, "compression_ratio": 1.2285714285714286, "no_speech_prob": 1.5190832966527346e-11}, {"id": 9, "seek": 26918, "start": 269.18, "end": 299.18, "text": "单次调用省下来的钱可能没那么明显,但当系统每天都需要完成几万次几十万次判断时,成本差距就会无限的被放大。第三个优势就是输出稳定。如果开发者只给出了5个候选选项,他就只会在这5个选项里返回结果,不会突然说创造出第6个答案,也不需要程序再从一段自然语言里面去提取结论。但这里一定要区分两件事情,输出格式稳定不等于判断是一定正确的。jef不会在格式上去胡乱发挥,但他在这些什么情况下,就是说,如果开发者只给出了5个候选项,他", "tokens": [50365, 47446, 9487, 8897, 225, 9254, 2862, 223, 4438, 6912, 1546, 39623, 16657, 10062, 35693, 11100, 1431, 122, 11, 8395, 16233, 25368, 10115, 253, 23664, 6135, 7182, 35748, 41509, 6336, 254, 23570, 9487, 6336, 254, 20145, 23570, 9487, 2437, 97, 4307, 255, 15729, 11, 11336, 8802, 21679, 6563, 251, 3111, 12949, 31364, 43446, 1546, 23238, 12744, 3582, 1543, 35878, 7549, 7384, 246, 5087, 123, 5620, 9830, 241, 7781, 10415, 111, 12088, 1543, 13119, 18937, 28926, 12444, 14003, 23197, 7781, 2289, 20, 7549, 11238, 2215, 231, 2215, 231, 10178, 117, 11, 5000, 3111, 14003, 12949, 3581, 5562, 20, 7549, 2215, 231, 10178, 117, 15759, 3316, 242, 8350, 45641, 9319, 11, 47928, 40859, 5823, 8090, 2437, 249, 37583, 7781, 7536, 21, 7549, 28223, 28899, 11, 6404, 1960, 35748, 29649, 6346, 237, 8623, 35630, 2257, 28427, 9722, 5823, 5233, 255, 12009, 15759, 8833, 6734, 20949, 29436, 45641, 7422, 118, 1543, 8395, 35102, 48161, 9937, 118, 6627, 36257, 49459, 11, 9830, 241, 7781, 30921, 27584, 10415, 111, 12088, 1960, 10187, 37732, 2437, 97, 4307, 255, 1541, 24272, 15789, 38114, 106, 1546, 1543, 2884, 69, 47928, 3581, 30921, 27584, 5708, 6734, 43861, 2930, 109, 28926, 8501, 98, 11, 8395, 5000, 3581, 5562, 13824, 10440, 46514, 4438, 11, 5620, 8090, 11, 13119, 18937, 28926, 12444, 14003, 23197, 7781, 2289, 20, 7549, 11238, 2215, 231, 10178, 117, 11, 5000], "temperature": 0, "avg_logprob": -0.18526448476474916, "compression_ratio": 1.4254807692307692, "no_speech_prob": 1.435262846072316e-11}, {"id": 10, "seek": 29918, "start": 299.18, "end": 328.5, "text": "完全可能在合法选项里面去选错,所谓不产生自由闻闻幻觉,解决的是接口问题,不是准确性的问题。第4个优势就是它会返回概率。普通分类器往往只告诉你答案是什么,那jef希望去把置信程度也交给程序,这样系统就可以设计不同的策略,把握很高,而且操作风险很低,可以自动执行。把握一般,那就交给更强的大模型做复合,把握太低的,或者涉及到退款付款,商库这类高风险动作的,就转给人工。", "tokens": [50365, 37100, 16657, 3581, 14245, 11148, 2215, 231, 10178, 117, 15759, 8833, 6734, 2215, 231, 29900, 11, 5966, 8897, 241, 1960, 1369, 100, 8244, 45233, 8259, 119, 8259, 119, 3509, 119, 24447, 11, 17278, 5676, 111, 24620, 14468, 18144, 34069, 11, 7296, 6336, 228, 38114, 106, 21686, 1546, 34069, 1543, 7536, 19, 7549, 7384, 246, 5087, 123, 5620, 11284, 12949, 3316, 242, 8350, 26181, 44866, 1543, 29993, 19550, 6627, 22113, 119, 34386, 29510, 29510, 14003, 44816, 2166, 28223, 28899, 1541, 10440, 11, 4184, 2884, 69, 29955, 6734, 16075, 34719, 17665, 29649, 13127, 6404, 28455, 23197, 29649, 6346, 237, 11, 21209, 25368, 10115, 253, 36781, 7422, 122, 7422, 94, 47123, 1546, 7973, 244, 6904, 98, 11, 16075, 11673, 94, 4563, 12979, 11, 22942, 13167, 235, 11914, 47209, 8842, 102, 4563, 41377, 11, 6723, 9722, 34961, 3416, 100, 8082, 1543, 16075, 11673, 94, 2257, 49640, 11, 46887, 28455, 23197, 19002, 5702, 118, 1546, 3582, 41908, 39823, 10907, 1787, 235, 14245, 11, 16075, 11673, 94, 9455, 41377, 1546, 11, 31148, 35681, 231, 25703, 4511, 46361, 48798, 41469, 48798, 11, 45581, 6346, 241, 5562, 22113, 119, 12979, 47209, 8842, 102, 34961, 11914, 1546, 11, 3111, 17819, 105, 23197, 4035, 23323, 1543, 51831], "temperature": 0, "avg_logprob": -0.07608139514923096, "compression_ratio": 1.3061224489795917, "no_speech_prob": 1.6316547071792975e-11}, {"id": 11, "seek": 32850, "start": 328.5, "end": 341.3, "text": "另外,jef还可以读取一次材料,同时回答多个相互独立的问题。还是前面的客服消息,他可以同时判断用户意图,情绪,紧急程度,和是否需要人工介入,不用反复的去读取一段上下文。", "tokens": [50365, 26202, 11, 2884, 69, 14852, 6723, 5233, 119, 29436, 27505, 4422, 238, 33404, 11, 13089, 15729, 8350, 28223, 6392, 7549, 15106, 1369, 240, 18637, 105, 24409, 1546, 34069, 1543, 45726, 8945, 8833, 1546, 32316, 27408, 28837, 26460, 11, 5000, 6723, 13089, 15729, 2437, 97, 4307, 255, 9254, 1486, 115, 9042, 3919, 122, 11, 10570, 10115, 103, 11, 7732, 100, 36379, 29649, 13127, 11, 12565, 1541, 45334, 35748, 4035, 23323, 30312, 14028, 11, 24384, 22138, 1787, 235, 1546, 6734, 5233, 119, 29436, 2257, 28427, 5708, 4438, 17174, 1543, 51005], "temperature": 0, "avg_logprob": -0.11242147728248879, "compression_ratio": 1.3481308411214954, "no_speech_prob": 1.6388161661051726e-11}, {"id": 12, "seek": 32850, "start": 341.3, "end": 358.1, "text": "OK,那这条路线到底对还是不对?其实我的理解,在他适合的任务里面,这条路线肯定是成立的,大原模型通常要采用自回归生成,他需要一个token,接着一个token生成答案。这个机制呢,给予了他协作,编程和开放式推理的能力,但如果程序只需要一个不而直,或者一个没举项,", "tokens": [51005, 9443, 11, 4184, 5562, 48837, 24658, 16853, 123, 33883, 8713, 45726, 41639, 30, 41646, 14200, 13876, 17278, 11, 3581, 5000, 2215, 224, 14245, 1546, 26443, 5087, 94, 15759, 8833, 11, 5562, 48837, 24658, 16853, 123, 14356, 107, 12088, 1541, 11336, 24409, 1546, 11, 3582, 19683, 41908, 39823, 19550, 11279, 4275, 5873, 229, 9254, 9722, 8350, 7391, 240, 8244, 11336, 11, 5000, 35748, 20182, 83, 8406, 11, 14468, 20708, 20182, 83, 8406, 8244, 11336, 28223, 28899, 1543, 15368, 37960, 25491, 6240, 11, 23197, 1369, 230, 2289, 5000, 5322, 237, 11914, 11, 38109, 244, 29649, 12565, 18937, 12744, 27584, 33597, 13876, 1546, 8225, 13486, 11, 8395, 13119, 29649, 6346, 237, 14003, 35748, 20182, 1960, 11070, 16186, 11, 31148, 20182, 10062, 940, 122, 10178, 117, 11, 51845], "temperature": 0, "avg_logprob": -0.11242147728248879, "compression_ratio": 1.3481308411214954, "no_speech_prob": 1.6388161661051726e-11}, {"id": 13, "seek": 35810, "start": 358.1, "end": 373.1, "text": "完整走一遍文本生成的流程,确实比较重。而jef把输出空间提前的限制住,不再去生成任意文字,而是直接围绕候选答案去计算概率,他失去了自由表达能力,也因此换来了更低的延迟,更低的成本,和更容易被代码处理的结果。", "tokens": [50365, 14128, 27662, 9575, 2257, 3330, 235, 17174, 8802, 8244, 11336, 1546, 27854, 29649, 11, 38114, 106, 24726, 11706, 9830, 225, 12624, 1543, 11070, 2884, 69, 16075, 9830, 241, 7781, 23322, 31685, 20949, 8945, 1546, 43446, 25491, 21632, 11, 1960, 8623, 6734, 8244, 11336, 26443, 9042, 17174, 22381, 11, 11070, 1541, 43297, 3919, 112, 10115, 243, 11238, 2215, 231, 28223, 28899, 6734, 7422, 94, 19497, 26181, 44866, 11, 5000, 28564, 45190, 45233, 17571, 9830, 122, 8225, 13486, 11, 6404, 8698, 17947, 26075, 95, 34912, 19002, 41377, 1546, 17527, 114, 3316, 253, 11, 19002, 41377, 1546, 11336, 8802, 11, 12565, 19002, 49212, 23238, 19105, 23230, 223, 1787, 226, 13876, 1546, 45641, 9319, 1543, 51115], "temperature": 0, "avg_logprob": -0.06311465968256412, "compression_ratio": 1.1747967479674797, "no_speech_prob": 1.5808585343557446e-11}, {"id": 14, "seek": 37310, "start": 373.1, "end": 384.1, "text": "但他也不是所谓问题的答案,如果任务需要长链条推理,需要精确计算,需要理解复杂的日期关系,或者需要模型给出完整理由,jef都不是合适选择。", "tokens": [50365, 8395, 5000, 6404, 7296, 5966, 8897, 241, 34069, 1546, 28223, 28899, 11, 13119, 26443, 5087, 94, 35748, 32271, 165, 241, 122, 48837, 33597, 13876, 11, 35748, 34910, 38114, 106, 7422, 94, 19497, 11, 35748, 13876, 17278, 1787, 235, 4422, 224, 1546, 6890, 16786, 28053, 25368, 11, 31148, 35748, 41908, 39823, 23197, 7781, 14128, 27662, 13876, 23786, 11, 2884, 69, 7182, 7296, 14245, 2215, 224, 2215, 231, 6852, 102, 1543, 50915], "temperature": 0, "avg_logprob": -0.0872655889605949, "compression_ratio": 1.3180592991913747, "no_speech_prob": 1.492441240036335e-11}, {"id": 15, "seek": 37310, "start": 384.1, "end": 401.1, "text": "他的官方文档也明确提出,他在算术记住日期比较多部推理,以及包含大量无关信息的任务上存在短板,他同样可能也会受到提示词注入的影响。所以讨论这条路线的时候,我觉得不能问jef和cloud更聪明,他们解决的就不是同一种问题。", "tokens": [50915, 31309, 31929, 9249, 17174, 21416, 96, 6404, 11100, 38114, 106, 20949, 7781, 11, 5000, 3581, 19497, 1474, 107, 34756, 21632, 6890, 16786, 11706, 9830, 225, 6392, 13470, 33597, 13876, 11, 40282, 23305, 2392, 104, 3582, 26748, 31364, 28053, 17665, 26460, 1546, 26443, 5087, 94, 5708, 39781, 3581, 5881, 255, 43664, 11, 5000, 13089, 14496, 16657, 6404, 12949, 23151, 4511, 20949, 25696, 5233, 235, 26432, 14028, 1546, 16820, 3858, 235, 1543, 7239, 7422, 101, 7422, 118, 5562, 48837, 24658, 16853, 123, 49873, 11, 45681, 28590, 22064, 2884, 69, 12565, 44495, 19002, 8171, 103, 11100, 11, 47911, 17278, 5676, 111, 1546, 3111, 7296, 13089, 2257, 39810, 34069, 1543, 51765], "temperature": 0, "avg_logprob": -0.0872655889605949, "compression_ratio": 1.3180592991913747, "no_speech_prob": 1.492441240036335e-11}, {"id": 16, "seek": 40110, "start": 401.1, "end": 415.1, "text": "cloud和gbt擅长的是开放式任务,传统分类模型适合类别长期固定,数据充足,可以专门训练的任务,那jef想占据的,其实中间那一层,问题由开发者临时定义,但输出必须限制在明确的类型和选项里。", "tokens": [50365, 44495, 12565, 70, 4517, 13167, 227, 32271, 24620, 18937, 12744, 27584, 26443, 5087, 94, 11, 7384, 254, 10115, 253, 6627, 22113, 119, 41908, 39823, 2215, 224, 14245, 22113, 119, 18453, 32271, 16786, 3919, 118, 12088, 11, 33188, 26075, 106, 2347, 227, 37236, 11, 6723, 940, 241, 8259, 101, 7422, 255, 10115, 225, 1546, 26443, 5087, 94, 11, 4184, 2884, 69, 7093, 5322, 254, 26075, 106, 1546, 11, 41646, 5975, 31685, 4184, 2257, 9636, 224, 11, 34069, 23786, 18937, 28926, 12444, 940, 112, 15729, 12088, 2930, 231, 11, 8395, 9830, 241, 7781, 28531, 10178, 119, 43446, 25491, 3581, 11100, 38114, 106, 1546, 22113, 119, 39823, 12565, 2215, 231, 10178, 117, 15759, 1543, 51065], "temperature": 0, "avg_logprob": -0.07296816072507536, "compression_ratio": 1.2817955112219452, "no_speech_prob": 1.7646942934712584e-11}, {"id": 17, "seek": 40110, "start": 415.1, "end": 429.1, "text": "这一层以前经常交给大模型,现在终于有人开始为他单独去设计模型了。另外开发者也已经在尝试把jef放进不少的场景里面了,想客服系统,来判断用户意图,把消息转到查询订单,退款,投诉,或者人工客服。", "tokens": [51065, 5562, 2257, 9636, 224, 44996, 30276, 11279, 28455, 23197, 3582, 41908, 39823, 11, 25040, 10115, 230, 37732, 45820, 45213, 13992, 5000, 47446, 18637, 105, 6734, 7422, 122, 7422, 94, 41908, 39823, 2289, 1543, 26202, 18937, 28926, 12444, 6404, 49161, 3581, 1530, 251, 5233, 243, 16075, 2884, 69, 12744, 36700, 1960, 15686, 1546, 50255, 50218, 15759, 8833, 2289, 11, 7093, 32316, 27408, 25368, 10115, 253, 11, 6912, 2437, 97, 4307, 255, 9254, 1486, 115, 9042, 3919, 122, 11, 16075, 28837, 26460, 17819, 105, 4511, 42623, 5233, 95, 7422, 95, 47446, 11, 46361, 48798, 11, 37094, 43570, 11, 31148, 4035, 23323, 32316, 27408, 1543, 51765], "temperature": 0, "avg_logprob": -0.07296816072507536, "compression_ratio": 1.2817955112219452, "no_speech_prob": 1.7646942934712584e-11}, {"id": 18, "seek": 42910, "start": 429.1, "end": 435.1, "text": "其他系统可以用他先来判断检查回来的资料到底是不是相关的,先过滤掉无关内容,再把剩下的材料交给大模型。", "tokens": [50365, 9572, 5000, 25368, 10115, 253, 6723, 9254, 5000, 10108, 6912, 2437, 97, 4307, 255, 16407, 222, 42623, 8350, 6912, 1546, 5266, 226, 33404, 33883, 23034, 15106, 28053, 1546, 11, 10108, 16866, 15868, 97, 29327, 31364, 28053, 34742, 25750, 11, 8623, 16075, 5935, 102, 4438, 1546, 4422, 238, 33404, 28455, 23197, 3582, 41908, 39823, 1543, 50665], "temperature": 0, "avg_logprob": -0.1280564448630163, "compression_ratio": 1.2918287937743191, "no_speech_prob": 1.8747075072877983e-11}, {"id": 19, "seek": 42910, "start": 435.1, "end": 444.1, "text": "内容平台也可以用他来做垃圾信息,辱骂内容,风险文本的前置判断。agent可以用他来选择下一步调用哪个工具,或者决定任务,是不是已经完整。", "tokens": [50665, 34742, 25750, 16716, 15433, 47935, 9254, 5000, 6912, 10907, 33701, 225, 2523, 122, 17665, 26460, 11, 9830, 109, 49657, 224, 34742, 25750, 11, 47209, 8842, 102, 17174, 8802, 1546, 8945, 34719, 2437, 97, 4307, 255, 1543, 559, 317, 6723, 9254, 5000, 6912, 2215, 231, 6852, 102, 45271, 31429, 8897, 225, 9254, 17028, 7549, 23323, 39806, 11, 31148, 5676, 111, 12088, 26443, 5087, 94, 11, 23034, 49161, 14128, 27662, 1543, 51115], "temperature": 0, "avg_logprob": -0.1280564448630163, "compression_ratio": 1.2918287937743191, "no_speech_prob": 1.8747075072877983e-11}, {"id": 20, "seek": 44410, "start": 444.1, "end": 455.1, "text": "Blozer Use 社区还做了一个jef ultrafast浏览器agent。jef负责从网页元素去选择该点击的按钮和输入框,只有遇到真实需要填写的文字的时候,才去调用生成模型。", "tokens": [50365, 33, 752, 4527, 8278, 220, 27658, 9937, 118, 14852, 10907, 2289, 20182, 2884, 69, 14808, 7011, 11038, 237, 6758, 230, 34386, 559, 317, 1543, 2884, 69, 18464, 253, 18464, 96, 35630, 16469, 239, 10178, 113, 14812, 37204, 6734, 2215, 231, 6852, 102, 44646, 12579, 6336, 119, 1546, 26613, 25153, 106, 12565, 9830, 241, 14028, 21416, 228, 11, 35244, 3330, 229, 4511, 6303, 24726, 35748, 13331, 104, 5676, 247, 1546, 17174, 22381, 49873, 11, 18888, 6734, 8897, 225, 9254, 8244, 11336, 41908, 39823, 1543, 50915], "temperature": 0, "avg_logprob": -0.1267138675407127, "compression_ratio": 1.1830985915492958, "no_speech_prob": 1.6177146427986955e-11}, {"id": 21, "seek": 44410, "start": 455.1, "end": 460.1, "text": "他们的结果呢,一次疏离室到伦敦的航班搜索演示,只用了7.1秒。", "tokens": [50915, 47911, 1546, 45641, 9319, 6240, 11, 27505, 18140, 237, 18614, 119, 2415, 97, 4511, 7384, 99, 7017, 99, 1546, 10256, 103, 44695, 162, 14987, 7732, 95, 31382, 25696, 11, 14003, 9254, 2289, 22, 13, 16, 20956, 1543, 51165], "temperature": 0, "avg_logprob": -0.1267138675407127, "compression_ratio": 1.1830985915492958, "no_speech_prob": 1.6177146427986955e-11}, {"id": 22, "seek": 44410, "start": 460.1, "end": 471.1, "text": "Type Safe自己也演示过让jef去玩Doom。输入不是游戏画面,而是血量、敌人位置、弹药数量等已经整理好的游戏状态。jef不断从有限动作中去做选择,实现每秒的多次决策。", "tokens": [51165, 22464, 494, 27030, 17645, 6404, 31382, 25696, 16866, 33650, 2884, 69, 6734, 19912, 7653, 298, 1543, 9830, 241, 14028, 7296, 9592, 116, 1486, 237, 27126, 8833, 11, 11070, 1541, 44110, 26748, 1231, 7017, 234, 4035, 11160, 34719, 1231, 5702, 117, 31409, 107, 33188, 26748, 10187, 49161, 27662, 13876, 20715, 9592, 116, 1486, 237, 35276, 114, 3757, 223, 1543, 2884, 69, 1960, 4307, 255, 35630, 2412, 43446, 34961, 11914, 5975, 6734, 10907, 2215, 231, 6852, 102, 11, 24726, 20204, 23664, 20956, 1546, 6392, 9487, 5676, 111, 7973, 244, 1543, 51715], "temperature": 0, "avg_logprob": -0.1267138675407127, "compression_ratio": 1.1830985915492958, "no_speech_prob": 1.6177146427986955e-11}, {"id": 23, "seek": 47110, "start": 471.1, "end": 482.1, "text": "这些案例表面上差别很大,但他们其实有个共同点,任务边界清楚,后选动作明确,请求频率很高,而且程序不需要一段漂亮的文字,只需要知道下一步到底该做什么。", "tokens": [50365, 5562, 13824, 28899, 17797, 17571, 8833, 5708, 21679, 18453, 4563, 3582, 11, 8395, 47911, 41646, 2412, 7549, 15408, 13089, 12579, 11, 26443, 5087, 94, 40503, 21185, 38433, 11, 13547, 2215, 231, 34961, 11914, 11100, 38114, 106, 11, 27908, 32718, 39752, 44866, 4563, 12979, 11, 22942, 29649, 6346, 237, 1960, 35748, 2257, 28427, 29019, 1546, 17174, 22381, 11, 14003, 35748, 7758, 45271, 31429, 33883, 44646, 10907, 10440, 1543, 50915], "temperature": 0, "avg_logprob": -0.15252580790939727, "compression_ratio": 1.2193548387096773, "no_speech_prob": 1.8999435705269185e-11}, {"id": 24, "seek": 47110, "start": 482.1, "end": 484.1, "text": "这个才是jef真正适合的位置。", "tokens": [50915, 15368, 18888, 1541, 2884, 69, 6303, 15789, 2215, 224, 14245, 1546, 11160, 34719, 1543, 51015], "temperature": 0, "avg_logprob": -0.15252580790939727, "compression_ratio": 1.2193548387096773, "no_speech_prob": 1.8999435705269185e-11}, {"id": 25, "seek": 47110, "start": 484.1, "end": 492.1, "text": "如果你想自己上手,流程其实也不复杂,可以直接去用Type Safe提供的Pregground和API,也可以通过VSL AI Gateway或者CloudFair去调用。", "tokens": [51015, 13119, 38386, 17645, 5708, 11389, 11, 27854, 29649, 41646, 6404, 1960, 1787, 235, 4422, 224, 11, 6723, 43297, 6734, 9254, 22464, 494, 27030, 20949, 3254, 249, 1546, 47, 3375, 2921, 12565, 4715, 40, 11, 47935, 19550, 16866, 53, 47012, 7318, 48394, 31148, 32787, 37, 1246, 6734, 8897, 225, 9254, 1543, 51415], "temperature": 0, "avg_logprob": -0.15252580790939727, "compression_ratio": 1.2193548387096773, "no_speech_prob": 1.8999435705269185e-11}, {"id": 26, "seek": 47110, "start": 492.1, "end": 500.1, "text": "一次请求里主要可以放两个东西,一份是State,也就是模型做判断需要看到的材料,另一份是Questions,也就是你希望他回答了。", "tokens": [51415, 27505, 27908, 32718, 15759, 13557, 4275, 6723, 12744, 36257, 7549, 38409, 16220, 11, 2257, 36266, 1541, 4520, 473, 11, 6404, 5620, 41908, 39823, 10907, 2437, 97, 4307, 255, 35748, 18032, 1546, 4422, 238, 33404, 11, 22762, 2257, 36266, 1541, 8547, 377, 626, 11, 6404, 5620, 2166, 29955, 5000, 8350, 28223, 2289, 1543, 51815], "temperature": 0, "avg_logprob": -0.15252580790939727, "compression_ratio": 1.2193548387096773, "no_speech_prob": 1.8999435705269185e-11}, {"id": 27, "seek": 50010, "start": 500.1, "end": 503.1, "text": "也就是你希望他回答了问题,候选选项,或者评分范围。", "tokens": [50365, 6404, 5620, 2166, 29955, 5000, 8350, 28223, 2289, 34069, 11, 11238, 2215, 231, 2215, 231, 10178, 117, 11, 31148, 5233, 226, 6627, 27395, 225, 3919, 112, 1543, 50515], "temperature": 0, "avg_logprob": -0.07256103515625, "compression_ratio": 1.3078880407124682, "no_speech_prob": 1.466573043562569e-11}, {"id": 28, "seek": 50010, "start": 503.1, "end": 509.1, "text": "jef读完State以后,会返回每个问题的答案和概率,你的程序再根据概率去决定下一个动作。", "tokens": [50515, 2884, 69, 5233, 119, 14128, 4520, 473, 3588, 13547, 11, 12949, 3316, 242, 8350, 23664, 7549, 34069, 1546, 28223, 28899, 12565, 26181, 44866, 11, 18961, 29649, 6346, 237, 8623, 31337, 26075, 106, 26181, 44866, 6734, 5676, 111, 12088, 4438, 20182, 34961, 11914, 1543, 50815], "temperature": 0, "avg_logprob": -0.07256103515625, "compression_ratio": 1.3078880407124682, "no_speech_prob": 1.466573043562569e-11}, {"id": 29, "seek": 50010, "start": 509.1, "end": 515.1, "text": "不过真正困难的地方,往往不是API的接通,而是去把你的问题给定义清楚。", "tokens": [50815, 1960, 16866, 6303, 15789, 3919, 108, 46531, 1546, 30146, 11, 29510, 29510, 7296, 4715, 40, 1546, 14468, 19550, 11, 11070, 1541, 6734, 16075, 18961, 34069, 23197, 12088, 2930, 231, 38433, 1543, 51115], "temperature": 0, "avg_logprob": -0.07256103515625, "compression_ratio": 1.3078880407124682, "no_speech_prob": 1.466573043562569e-11}, {"id": 30, "seek": 50010, "start": 515.1, "end": 517.1, "text": "还是那句话,我买的东西一个星期没到,现在不想要了。", "tokens": [51115, 45726, 4184, 34592, 21596, 11, 1654, 2930, 108, 1546, 38409, 16220, 20182, 20682, 16786, 10062, 4511, 11, 25040, 1960, 7093, 4275, 2289, 1543, 51215], "temperature": 0, "avg_logprob": -0.07256103515625, "compression_ratio": 1.3078880407124682, "no_speech_prob": 1.466573043562569e-11}, {"id": 31, "seek": 50010, "start": 517.1, "end": 522.1, "text": "如果选项只有物流的延迟和申请退款,知道单选题本身就设计得太好。", "tokens": [51215, 13119, 2215, 231, 10178, 117, 35244, 23516, 27854, 1546, 17527, 114, 3316, 253, 12565, 3526, 111, 27908, 46361, 48798, 11, 7758, 47446, 2215, 231, 30716, 8802, 19847, 3111, 7422, 122, 7422, 94, 5916, 9455, 2131, 1543, 51465], "temperature": 0, "avg_logprob": -0.07256103515625, "compression_ratio": 1.3078880407124682, "no_speech_prob": 1.466573043562569e-11}, {"id": 32, "seek": 50010, "start": 522.1, "end": 526.1, "text": "物流延迟是原因,申请退款是诉求,它们可以同时成立。", "tokens": [51465, 23516, 27854, 17527, 114, 3316, 253, 1541, 42577, 11, 3526, 111, 27908, 46361, 48798, 1541, 43570, 32718, 11, 11284, 9497, 6723, 13089, 15729, 11336, 24409, 1543, 51665], "temperature": 0, "avg_logprob": -0.07256103515625, "compression_ratio": 1.3078880407124682, "no_speech_prob": 1.466573043562569e-11}, {"id": 33, "seek": 52610, "start": 526.1, "end": 533.1, "text": "模型不会提开发者修复一套混乱的业务标准,给他的信息不完整,问题不明确,或者选项互相重叠,速度再快也没有用。", "tokens": [50365, 41908, 39823, 47928, 20949, 18937, 28926, 12444, 7792, 106, 1787, 235, 2257, 1881, 245, 48640, 2930, 109, 1546, 940, 248, 5087, 94, 162, 3921, 6336, 228, 11, 23197, 31309, 17665, 26460, 1960, 14128, 27662, 11, 34069, 1960, 11100, 38114, 106, 11, 31148, 2215, 231, 10178, 117, 1369, 240, 15106, 12624, 2129, 254, 11, 31217, 13127, 8623, 10251, 6404, 17944, 9254, 1543, 50715], "temperature": 0, "avg_logprob": -0.05356796264648438, "compression_ratio": 1.2096219931271477, "no_speech_prob": 1.4896497230187933e-11}, {"id": 34, "seek": 52610, "start": 533.1, "end": 537.1, "text": "所以比较稳妥的使用方法就是先选择一个范围很小,答案比较清楚的任务。", "tokens": [50715, 7239, 11706, 9830, 225, 10415, 111, 5648, 98, 1546, 22982, 9254, 9249, 11148, 5620, 10108, 2215, 231, 6852, 102, 20182, 27395, 225, 3919, 112, 4563, 7322, 11, 28223, 28899, 11706, 9830, 225, 38433, 1546, 26443, 5087, 94, 1543, 50915], "temperature": 0, "avg_logprob": -0.05356796264648438, "compression_ratio": 1.2096219931271477, "no_speech_prob": 1.4896497230187933e-11}, {"id": 35, "seek": 52610, "start": 537.1, "end": 543.1, "text": "哪一批已经有人工处理过的历史数据,让jef在旁边运行,暂时不要参加真实操作。", "tokens": [50915, 17028, 2257, 3416, 117, 49161, 45820, 23323, 1787, 226, 13876, 16866, 1546, 5014, 228, 45399, 33188, 26075, 106, 11, 33650, 2884, 69, 3581, 4479, 223, 40503, 3316, 238, 8082, 11, 23826, 224, 15729, 11962, 2129, 224, 9990, 6303, 24726, 13167, 235, 11914, 1543, 51215], "temperature": 0, "avg_logprob": -0.05356796264648438, "compression_ratio": 1.2096219931271477, "no_speech_prob": 1.4896497230187933e-11}, {"id": 36, "seek": 54310, "start": 543.1, "end": 550.1, "text": "观察他在哪些任务上稳定、智信度和真实正确率是不是匹配,再根据风险设定自动通过,模型复合和人工处理的预知。", "tokens": [50365, 42350, 47550, 5000, 3581, 17028, 13824, 26443, 5087, 94, 5708, 10415, 111, 12088, 1231, 5094, 118, 17665, 13127, 12565, 6303, 24726, 15789, 38114, 106, 44866, 23034, 9937, 117, 38846, 11, 8623, 31337, 26075, 106, 47209, 8842, 102, 7422, 122, 12088, 9722, 34961, 19550, 16866, 11, 41908, 39823, 1787, 235, 14245, 12565, 4035, 23323, 1787, 226, 13876, 1546, 12501, 226, 6498, 1543, 50715], "temperature": 0, "avg_logprob": -0.12396110577529736, "compression_ratio": 1.2484848484848485, "no_speech_prob": 1.662541632141412e-11}, {"id": 37, "seek": 54310, "start": 550.1, "end": 554.1, "text": "低风险,高频,可设销的动作,可以更积极地去做状况。", "tokens": [50715, 41377, 47209, 8842, 102, 11, 12979, 39752, 11, 4429, 7422, 122, 23049, 222, 1546, 34961, 11914, 11, 6723, 19002, 8204, 107, 7360, 223, 10928, 6734, 10907, 35276, 114, 43943, 1543, 50915], "temperature": 0, "avg_logprob": -0.12396110577529736, "compression_ratio": 1.2484848484848485, "no_speech_prob": 1.662541632141412e-11}, {"id": 38, "seek": 54310, "start": 554.1, "end": 559.1, "text": "而退款,资金,权限和删除数据这些操作,仍然去保留代码规则和人工审批。", "tokens": [50915, 11070, 46361, 48798, 11, 5266, 226, 19117, 11, 4422, 225, 43446, 12565, 2437, 254, 32999, 33188, 26075, 106, 5562, 13824, 13167, 235, 11914, 11, 1550, 235, 5823, 6734, 24302, 24456, 19105, 23230, 223, 6758, 226, 2437, 247, 12565, 4035, 23323, 2415, 94, 3416, 117, 1543, 51165], "temperature": 0, "avg_logprob": -0.12396110577529736, "compression_ratio": 1.2484848484848485, "no_speech_prob": 1.662541632141412e-11}, {"id": 39, "seek": 54310, "start": 559.1, "end": 564.1, "text": "所以我认为,jef真正有意思的地方,不只是说把原本的一次调查做便宜了。", "tokens": [51165, 7239, 1654, 7422, 97, 13992, 11, 2884, 69, 6303, 15789, 2412, 16697, 1546, 30146, 11, 1960, 36859, 8090, 16075, 19683, 8802, 1546, 27505, 8897, 225, 42623, 10907, 27364, 2415, 250, 2289, 1543, 51415], "temperature": 0, "avg_logprob": -0.12396110577529736, "compression_ratio": 1.2484848484848485, "no_speech_prob": 1.662541632141412e-11}, {"id": 40, "seek": 56410, "start": 564.1, "end": 568.1, "text": "他可能让原本因为成本太高而没有发生的判断,开始大量出现。", "tokens": [50365, 5000, 16657, 33650, 19683, 8802, 34627, 11336, 8802, 9455, 12979, 11070, 17944, 28926, 8244, 1546, 2437, 97, 4307, 255, 11, 45213, 3582, 26748, 7781, 20204, 1543, 50565], "temperature": 0, "avg_logprob": -0.08369228627422068, "compression_ratio": 1.35656836461126, "no_speech_prob": 1.624050199877658e-11}, {"id": 41, "seek": 56410, "start": 568.1, "end": 574.1, "text": "当然更多判断不一定自动带来更好的结果,如果标准本身就错了,系统值在更快更便宜的重复错误。", "tokens": [50565, 40486, 19002, 6392, 2437, 97, 4307, 255, 1960, 24272, 9722, 34961, 4845, 99, 6912, 19002, 20715, 45641, 9319, 11, 13119, 162, 3921, 6336, 228, 8802, 19847, 3111, 29900, 2289, 11, 25368, 10115, 253, 40242, 3581, 19002, 10251, 19002, 27364, 2415, 250, 1546, 12624, 1787, 235, 29900, 5233, 107, 1543, 50865], "temperature": 0, "avg_logprob": -0.08369228627422068, "compression_ratio": 1.35656836461126, "no_speech_prob": 1.624050199877658e-11}, {"id": 42, "seek": 56410, "start": 574.1, "end": 579.1, "text": "但我始终认为jef至少提出了一个很值得观察的方向。", "tokens": [50865, 8395, 1654, 15476, 10115, 230, 7422, 97, 13992, 2884, 69, 20844, 15686, 20949, 7781, 2289, 20182, 4563, 40242, 5916, 42350, 47550, 1546, 9249, 24282, 1543, 51115], "temperature": 0, "avg_logprob": -0.08369228627422068, "compression_ratio": 1.35656836461126, "no_speech_prob": 1.624050199877658e-11}, {"id": 43, "seek": 56410, "start": 579.1, "end": 589.1, "text": "未来的软件未必需要一个模型包办所有事情,复杂推理可以交给强模型,内容生成交给生成模型,范围明确的小判断,再去交给jef这一类的决策模型,再配合代码规则和人工审核。", "tokens": [51115, 29954, 6912, 1546, 17819, 107, 20485, 29954, 28531, 35748, 20182, 41908, 39823, 23305, 39453, 39300, 24675, 11, 1787, 235, 4422, 224, 33597, 13876, 6723, 28455, 23197, 5702, 118, 41908, 39823, 11, 34742, 25750, 8244, 11336, 28455, 23197, 8244, 11336, 41908, 39823, 11, 27395, 225, 3919, 112, 11100, 38114, 106, 1546, 7322, 2437, 97, 4307, 255, 11, 8623, 6734, 28455, 23197, 2884, 69, 5562, 2257, 22113, 119, 1546, 5676, 111, 7973, 244, 41908, 39823, 11, 8623, 38846, 14245, 19105, 23230, 223, 6758, 226, 2437, 247, 12565, 4035, 23323, 2415, 94, 8987, 116, 1543, 51615], "temperature": 0, "avg_logprob": -0.08369228627422068, "compression_ratio": 1.35656836461126, "no_speech_prob": 1.624050199877658e-11}, {"id": 44, "seek": 58910, "start": 589.1, "end": 597.1, "text": "不是每一道题,都需要一皮A篇作文,有些时候程序员真正需要的只是一个足够快,带着概率,可以立刻执行的答案。", "tokens": [50365, 7296, 23664, 2257, 6025, 30716, 11, 7182, 35748, 2257, 34582, 32, 20878, 229, 11914, 17174, 11, 2412, 13824, 29111, 29649, 6346, 237, 3606, 246, 6303, 15789, 35748, 1546, 36859, 20182, 37236, 1787, 253, 10251, 11, 4845, 99, 20708, 26181, 44866, 11, 6723, 24409, 45500, 3416, 100, 8082, 1546, 28223, 28899, 1543, 50765], "temperature": 0, "avg_logprob": -0.16666371734054, "compression_ratio": 1.0138888888888888, "no_speech_prob": 1.7463933077443983e-11}], "language": "zh"}