diff --git a/public/subtitles/ep76.srt b/public/subtitles/ep76.srt
new file mode 100644
index 0000000..67c64d5
--- /dev/null
+++ b/public/subtitles/ep76.srt
@@ -0,0 +1,680 @@
+1
+00:00:00,000 --> 00:00:03,720
+为什么说像是这样一个听起来没有什么
+
+2
+00:00:03,740 --> 00:00:06,180
+智力听起来非常廉价的
+
+3
+00:00:06,200 --> 00:00:07,990
+一个模型它能够这么火呢
+
+4
+00:00:11,300 --> 00:00:12,200
+大家好 这里是
+
+5
+00:00:12,220 --> 00:00:13,280
+AsyncTalk 我是
+
+6
+00:00:13,300 --> 00:00:13,820
+AsyncTalk
+
+7
+00:00:13,840 --> 00:00:14,570
+主播 AnnatarHe
+
+8
+00:00:14,740 --> 00:00:17,800
+今天的话我想来聊一个前几天
+
+9
+00:00:17,820 --> 00:00:19,650
+非常火的一个模型叫Jev
+
+10
+00:00:20,020 --> 00:00:21,650
+首先呢 Jev是什么呢
+
+11
+00:00:21,800 --> 00:00:22,840
+Jev的话它就是
+
+12
+00:00:22,860 --> 00:00:25,260
+一个非常非常快的
+
+13
+00:00:25,280 --> 00:00:27,800
+只用来去判断非常简单的东西
+
+14
+00:00:27,820 --> 00:00:29,760
+比如说苹果能不能吃
+
+15
+00:00:29,780 --> 00:00:31,740
+就这种非常简单的问题
+
+16
+00:00:31,760 --> 00:00:33,620
+然后它可以比Flash
+
+17
+00:00:33,640 --> 00:00:36,440
+快更多更多的速度来去返回
+
+18
+00:00:36,460 --> 00:00:39,820
+这就是它和其他大模型非常不一样的地方
+
+19
+00:00:39,840 --> 00:00:42,260
+因为我们回忆一下传统的大模型的话
+
+20
+00:00:42,280 --> 00:00:44,640
+比如说像是目前最顶级的
+
+21
+00:00:44,660 --> 00:00:47,040
+像是GPT-6系列
+
+22
+00:00:47,060 --> 00:00:48,380
+比如说像是
+
+23
+00:00:48,400 --> 00:00:50,100
+Claude Opus Fable
+
+24
+00:00:50,120 --> 00:00:52,320
+他们所有的这些模型
+
+25
+00:00:52,340 --> 00:00:55,560
+最主要就是想要一步一步的提升智能
+
+26
+00:00:55,580 --> 00:00:57,620
+提升思考的深度
+
+27
+00:00:57,640 --> 00:00:58,540
+但Jev的话
+
+28
+00:00:58,560 --> 00:00:59,700
+它是走另一条路
+
+29
+00:00:59,720 --> 00:01:02,120
+它只做一些非常简单的判断
+
+30
+00:01:02,140 --> 00:01:04,040
+但它可以做得很快
+
+31
+00:01:04,060 --> 00:01:05,120
+可以做得很好
+
+32
+00:01:05,140 --> 00:01:07,620
+我相信如果你有观看我们
+
+33
+00:01:07,640 --> 00:01:08,560
+AsyncTalk节目的话
+
+34
+00:01:08,580 --> 00:01:09,820
+你大概就是一个程序员
+
+35
+00:01:09,840 --> 00:01:12,600
+那程序员我们里边可以说
+
+36
+00:01:12,620 --> 00:01:14,900
+最重要的一个语句就是if了
+
+37
+00:01:14,920 --> 00:01:17,560
+那if A我们执行A条件
+
+38
+00:01:17,580 --> 00:01:19,300
+执行A分支
+
+39
+00:01:19,320 --> 00:01:21,160
+那如果不是if的话
+
+40
+00:01:21,180 --> 00:01:22,520
+那我们就执行B条件
+
+41
+00:01:22,540 --> 00:01:24,720
+那其实我们按照这个思想的话
+
+42
+00:01:24,740 --> 00:01:25,500
+我们来做
+
+43
+00:01:25,520 --> 00:01:26,840
+Coding agent的时候
+
+44
+00:01:26,860 --> 00:01:30,220
+是会经常出现这样的情况的
+
+45
+00:01:30,240 --> 00:01:31,320
+那我举一个例子
+
+46
+00:01:31,340 --> 00:01:32,200
+那比如说用户
+
+47
+00:01:32,220 --> 00:01:33,640
+他有一个prompt是说
+
+48
+00:01:33,660 --> 00:01:34,880
+帮我拿个手机
+
+49
+00:01:34,900 --> 00:01:37,060
+那拿个手机是拿哪个手机呢
+
+50
+00:01:37,080 --> 00:01:38,780
+如果是拿A手机的话
+
+51
+00:01:38,800 --> 00:01:39,320
+应该怎么样
+
+52
+00:01:39,340 --> 00:01:40,600
+如果是拿B手机的话
+
+53
+00:01:40,620 --> 00:01:41,100
+应该怎么样
+
+54
+00:01:41,120 --> 00:01:42,500
+那这个如果的判断
+
+55
+00:01:42,520 --> 00:01:43,400
+是应该怎么做呢
+
+56
+00:01:43,420 --> 00:01:45,540
+那对于我们普通的大模型来说的话
+
+57
+00:01:45,560 --> 00:01:47,460
+你可能需要去定一个
+
+58
+00:01:47,480 --> 00:01:48,440
+output structure
+
+59
+00:01:48,460 --> 00:01:51,160
+然后对返回的一个JSON
+
+60
+00:01:51,180 --> 00:01:52,080
+做一个处理
+
+61
+00:01:52,100 --> 00:01:55,140
+然后再去看这个意向
+
+62
+00:01:55,160 --> 00:01:56,540
+到底是A还是B
+
+63
+00:01:56,560 --> 00:01:58,140
+你需要走一遍完整的大模型
+
+64
+00:01:58,160 --> 00:01:59,460
+其实一方面的话
+
+65
+00:01:59,480 --> 00:02:01,800
+它耗时会稍微有点久
+
+66
+00:02:01,820 --> 00:02:02,940
+那另一方面的话
+
+67
+00:02:02,960 --> 00:02:04,100
+其实成本也比较高
+
+68
+00:02:04,120 --> 00:02:05,940
+那对于我们传统的大模型来说的话
+
+69
+00:02:05,960 --> 00:02:07,220
+它是token by token的
+
+70
+00:02:07,240 --> 00:02:08,940
+这样一个计算和计费
+
+71
+00:02:08,960 --> 00:02:10,960
+对于Jev这种模式来说的话
+
+72
+00:02:10,980 --> 00:02:12,160
+它就变得不一样了
+
+73
+00:02:12,180 --> 00:02:14,100
+它就是用来去做这种简单的判断
+
+74
+00:02:14,120 --> 00:02:16,820
+它不再需要像我们传统的大模型一样
+
+75
+00:02:16,840 --> 00:02:18,570
+过一遍完整的
+
+76
+00:02:18,920 --> 00:02:21,620
+漫长的一个用户的意图识别
+
+77
+00:02:21,640 --> 00:02:23,000
+当然它也不是我们
+
+78
+00:02:23,020 --> 00:02:24,640
+纯工程上面的
+
+79
+00:02:24,660 --> 00:02:25,820
+一个if condition
+
+80
+00:02:25,840 --> 00:02:27,970
+a到一个a branch
+
+81
+00:02:28,300 --> 00:02:30,560
+这样一个纯工程的方面
+
+82
+00:02:30,580 --> 00:02:34,040
+毕竟我们中间还是需要去调用
+
+83
+00:02:34,060 --> 00:02:35,760
+AI来理解一下用户的意图
+
+84
+00:02:35,780 --> 00:02:38,300
+那这个时候Jev的话就非常的合适
+
+85
+00:02:38,320 --> 00:02:42,100
+除了这个比较简单的用户的意图的识别
+
+86
+00:02:42,120 --> 00:02:43,060
+还有什么场景呢
+
+87
+00:02:43,080 --> 00:02:45,250
+那放在我们的AI agent
+
+88
+00:02:45,560 --> 00:02:46,840
+因为现在其实我们的
+
+89
+00:02:46,860 --> 00:02:48,680
+agent已经变得非常的复杂了
+
+90
+00:02:48,700 --> 00:02:50,820
+他从最开始的一些分析
+
+91
+00:02:50,840 --> 00:02:52,260
+到中间的执行
+
+92
+00:02:52,280 --> 00:02:53,720
+到最后的结果展示
+
+93
+00:02:53,740 --> 00:02:56,380
+中间是有非常大量的这种
+
+94
+00:02:56,400 --> 00:02:58,440
+简单的if 逻辑在里面的
+
+95
+00:02:58,460 --> 00:02:59,560
+比如说我们的Coding agent
+
+96
+00:02:59,580 --> 00:03:01,620
+那这个文件它是不是
+
+97
+00:03:01,640 --> 00:03:02,680
+TypeScript的文件
+
+98
+00:03:02,700 --> 00:03:04,440
+这种判断的话
+
+99
+00:03:04,460 --> 00:03:05,420
+它也可以交给AI
+
+100
+00:03:05,440 --> 00:03:09,340
+所以交给这样这种快速决断的模型来说
+
+101
+00:03:09,360 --> 00:03:11,410
+就变得非常的合适
+
+102
+00:03:11,660 --> 00:03:13,980
+还有一个场景就是验证一切的场景
+
+103
+00:03:14,000 --> 00:03:15,920
+当我们在这调用agent
+
+104
+00:03:15,940 --> 00:03:17,340
+来去做一些事情的时候
+
+105
+00:03:17,360 --> 00:03:19,580
+在这些非常多的步骤完成之后的话
+
+106
+00:03:19,600 --> 00:03:21,060
+我们是有一个验证的过程
+
+107
+00:03:21,080 --> 00:03:24,660
+但是其实我们并不能验证所有的流程
+
+108
+00:03:24,680 --> 00:03:26,880
+因为它会把你的时间拉得非常长
+
+109
+00:03:26,900 --> 00:03:28,700
+同样的话它也非常消耗token
+
+110
+00:03:28,720 --> 00:03:31,220
+但是如果我们有Jev这样的模型的话
+
+111
+00:03:31,240 --> 00:03:32,540
+它就可以非常快速的
+
+112
+00:03:32,560 --> 00:03:35,380
+来去进行一些简单的校验
+
+113
+00:03:35,400 --> 00:03:37,320
+这样的话我们最终得到的
+
+114
+00:03:37,340 --> 00:03:38,620
+工作的效率也会更好
+
+115
+00:03:38,640 --> 00:03:41,600
+那接下来我来展示一下这个Jev模型
+
+116
+00:03:41,620 --> 00:03:43,200
+它有什么应用
+
+117
+00:03:43,220 --> 00:03:45,360
+以及它大概会是长什么样子
+
+118
+00:03:45,380 --> 00:03:47,450
+那这里我来问他一个问题就是
+
+119
+00:03:48,040 --> 00:03:49,450
+苹果可以被吃吗
+
+120
+00:03:51,080 --> 00:03:52,340
+好 我们问一下
+
+121
+00:03:52,360 --> 00:03:55,300
+他回答99%的概率可以被吃
+
+122
+00:03:55,320 --> 00:03:57,910
+那我们问另外一个问题 比如说
+
+123
+00:04:00,960 --> 00:04:03,270
+我可以买沪深300吗
+
+124
+00:04:05,980 --> 00:04:08,740
+他给的概率是大概46% 47%的样子
+
+125
+00:04:08,760 --> 00:04:11,530
+那我再问一下 我应该买
+
+126
+00:04:14,240 --> 00:04:16,640
+他也给个大概44的样子
+
+127
+00:04:16,660 --> 00:04:18,150
+会比沪深300要低一些
+
+128
+00:04:20,680 --> 00:04:23,300
+他给了中规中矩大概53那样子
+
+129
+00:04:23,320 --> 00:04:25,400
+那其实我的上下文非常的少
+
+130
+00:04:25,420 --> 00:04:27,460
+我只是问他说我是不是应该出去玩
+
+131
+00:04:27,480 --> 00:04:29,640
+我并没有告诉他我的地址在哪里
+
+132
+00:04:29,660 --> 00:04:30,560
+我的时间在哪里
+
+133
+00:04:30,580 --> 00:04:31,620
+我的假期在哪里
+
+134
+00:04:31,640 --> 00:04:33,040
+他能给出来就是
+
+135
+00:04:33,060 --> 00:04:35,900
+一个非常快速的一个简单的判断
+
+136
+00:04:35,920 --> 00:04:38,660
+当我们在增加更多的上下文的时候
+
+137
+00:04:38,680 --> 00:04:41,600
+他会得出来一些更加准确的判断
+
+138
+00:04:41,620 --> 00:04:43,240
+那当然Jev这么好的话
+
+139
+00:04:43,260 --> 00:04:46,340
+那他是不是可以用来去取代
+
+140
+00:04:46,360 --> 00:04:48,540
+所有的这种简单的判断呢
+
+141
+00:04:48,560 --> 00:04:50,240
+他是不是足够可靠呢
+
+142
+00:04:50,260 --> 00:04:51,860
+但其实并不是
+
+143
+00:04:51,880 --> 00:04:53,340
+他像所有的大模型一样
+
+144
+00:04:53,360 --> 00:04:54,620
+他同样是有幻觉的
+
+145
+00:04:54,640 --> 00:04:56,920
+他只是做的问题比较简单
+
+146
+00:04:56,940 --> 00:04:58,560
+他的反应速度比较快
+
+147
+00:04:58,580 --> 00:05:01,260
+但他其实没办法完全避免幻觉
+
+148
+00:05:01,280 --> 00:05:02,320
+因为幻觉这个东西
+
+149
+00:05:02,340 --> 00:05:04,340
+人类也非常经常的有幻觉
+
+150
+00:05:04,360 --> 00:05:06,320
+这是一个其实很难解决的问题
+
+151
+00:05:06,340 --> 00:05:08,520
+当然我们需要承认说
+
+152
+00:05:08,540 --> 00:05:11,440
+他的这条路线非常的有趣
+
+153
+00:05:11,460 --> 00:05:14,040
+和我们现在主流的AI发展并不一样
+
+154
+00:05:14,060 --> 00:05:16,800
+它给我们提供了另一种可能性
+
+155
+00:05:16,820 --> 00:05:18,140
+这点我们要支持
+
+156
+00:05:18,160 --> 00:05:19,540
+它可以用来去替代一些
+
+157
+00:05:19,560 --> 00:05:21,700
+比较简单的用户的判断
+
+158
+00:05:21,720 --> 00:05:24,540
+大家也可以下来去使用一下
+
+159
+00:05:24,560 --> 00:05:27,160
+然后它现在的这个Waitlist
+
+160
+00:05:27,180 --> 00:05:29,100
+应该是已经完全开放了
+
+161
+00:05:29,120 --> 00:05:31,720
+你可以拿来去试一试看一看
+
+162
+00:05:31,740 --> 00:05:34,300
+是否可以去接入一些简单的工作流
+
+163
+00:05:34,320 --> 00:05:37,250
+然后来提升一些效率
+
+164
+00:05:37,700 --> 00:05:38,970
+降低一些成本
+
+165
+00:05:39,390 --> 00:05:41,240
+那也欢迎在评论区来留下你的看法
+
+166
+00:05:41,260 --> 00:05:45,020
+你认为说这样一条非常规的AI发展路线
+
+167
+00:05:45,040 --> 00:05:46,700
+是不是另一种路
+
+168
+00:05:46,720 --> 00:05:48,000
+或者说你怎么样看
+
+169
+00:05:48,020 --> 00:05:50,300
+这样的一种快速决断的模型
+
+170
+00:05:50,320 --> 00:05:52,190
+那我们下期节目再见拜拜
+
diff --git a/src/assets/ep76/cover.jpeg b/src/assets/ep76/cover.jpeg
new file mode 100644
index 0000000..e7c7bdc
Binary files /dev/null and b/src/assets/ep76/cover.jpeg differ
diff --git a/src/content/posts/ep76.mdx b/src/content/posts/ep76.mdx
new file mode 100644
index 0000000..4c92f4f
--- /dev/null
+++ b/src/content/posts/ep76.mdx
@@ -0,0 +1,79 @@
+---
+type: podcast-episode
+status: published
+slug: /posts/ep76
+guid: 76
+title: "EP76 Jev: 只回答是或否的模型，为什么火了？"
+subtitle: "Jev: 只回答是或否的模型，为什么火了？"
+publicationDate: 2026-10-08
+author: AnnatarHe
+season: 3
+episodeNumber: 76
+episodeType: full
+excerpt: "TypeSafe 发布的 Jev 只回答是或否、多选一和打分，70–500 毫秒返回，输出免费。从程序员最熟悉的 if 讲起，聊它在 AI Agent 里的用法、现场实测，以及快和便宜之外的局限。"
+url: https://youtu.be/KZd31Ptt8Os
+size: 0
+duration: 0
+explicit: false
+xyzLink: https://www.xiaoyuzhoufm.com/episode/6ac71498ea9bf07d68b0a09a
+youtubeId: KZd31Ptt8Os
+biliUrl: //player.bilibili.com/player.html?isOutside=true&aid=117401001662343&bvid=BV11PHm6hEGP&cid=42549314494&p=1
+cover: ../../assets/ep76/cover.jpeg
+srt: /subtitles/ep76.srt
+categories:
+    - Jev
+    - TypeSafe
+    - AI
+    - 大模型
+    - AIAgent
+    - 程序员
+    - 编程
+    - 人工智能
+    - AI工具
+    - 科技播客
+    - AsyncTalk
+---
+
+### 📕 Shownotes
+
+这期 AsyncTalk 聊一个前几天突然火起来的模型：TypeSafe 发布的 Jev。
+
+它不聊天、不写长文，只回答简单的问题：是或否、多选一、打分。端到端 70–500 毫秒返回，输入每百万 token 0.042 美元，输出免费。
+
+我们从程序员最熟悉的 if 讲起：
+
+- 让大模型做一个「是或否」的判断，要定输出结构、解析 JSON、按 token 计费，又慢又贵
+- 纯代码的 if 又理解不了用户意图
+- Jev 正好卡在两者中间
+
+然后聊它在 AI Agent 里的用法：意图识别、「这是不是 TypeScript 文件」这类判断、每一步的快速校验。现场实测了几个问题（苹果能不能吃、该不该买沪深 300、今天要不要出去玩）。最后说说它的局限：快、便宜，但不一定更准，同样会有幻觉。
+
+你怎么看这种「快速决断」的模型？欢迎在评论区聊聊。
+
+### ⏰ Timeline
+
+00:00 开场：一个“没什么智力”的模型为什么火了？
+
+00:19 Jev 是什么
+
+01:05 程序员最重要的语句：if
+
+01:22 让大模型做 if，贵在哪？
+
+02:09 Jev 的位置
+
+02:38 Agent 里的大量 if
+
+03:38 实测 Jev
+
+04:41 它足够可靠吗？
+
+05:21 试一试 & 你怎么看
+
+### 📊 数据来源
+
+片中延迟、成本、准确率数据来自 TypeSafe 公布的 System One 基准，引自：
+
+- DataCamp: [https://www.datacamp.com/blog/system-one-models-jev](https://www.datacamp.com/blog/system-one-models-jev)
+- RuntimeWire: [https://runtimewire.com/article/typesafe-jev-system-one-ai-model-early-access](https://runtimewire.com/article/typesafe-jev-system-one-ai-model-early-access)
+- NYU Shanghai RITS: [https://rits.shanghai.nyu.edu/ai/typesafe-jev-system-one-model](https://rits.shanghai.nyu.edu/ai/typesafe-jev-system-one-model)