1 00:00:00,000 --> 00:00:27,780 大家好,我是Kid。最近有個模型非常火,它的名字叫Jeff。它不是平常我們對話的那種大語言模型。它是由Type-Safe AI公司開發的一款模型。它的創始人曾經是在OpenAI工作,然後是XGPT的共同作者。Jeff特別適合無人職守的自動化,用它來打遊戲也非常好。我們可以通過這個視頻了解一下Jeff能幫我們做什麼。 2 00:00:27,780 --> 00:00:30,040 如果是正常的大语言模型 3 00:00:30,040 --> 00:00:32,740 它呢就会一个token一个token的 4 00:00:32,740 --> 00:00:33,540 蹦出自来 5 00:00:33,540 --> 00:00:36,380 花费时间8.5秒 6 00:00:36,380 --> 00:00:39,140 那如果是JF呢 7 00:00:39,140 --> 00:00:41,600 它就会给出几个选项 8 00:00:41,600 --> 00:00:44,720 然后会给出知心度 9 00:00:44,720 --> 00:00:47,320 花费时间0.1秒 10 00:00:47,320 --> 00:00:50,780 如果是让它玩游戏 11 00:00:50,780 --> 00:00:52,200 它呢是 12 00:00:52,200 --> 00:00:54,960 就是由固定的几个操作 13 00:00:54,960 --> 00:00:57,020 往上往右往前 14 00:00:57,020 --> 00:00:58,540 还有Shopter 15 00:00:58,540 --> 00:00:59,400 Jeff呢 16 00:00:59,400 --> 00:01:01,820 他是能做一个选择 17 00:01:01,820 --> 00:01:02,500 分类 18 00:01:02,500 --> 00:01:03,540 评分 19 00:01:03,540 --> 00:01:04,400 路由 20 00:01:04,400 --> 00:01:05,780 不能去写诗 21 00:01:05,780 --> 00:01:06,700 写代码 22 00:01:06,700 --> 00:01:09,180 或者是一步步的推理 23 00:01:09,180 --> 00:01:10,880 这视频做的还是非常好的 24 00:01:10,880 --> 00:01:12,160 对于开发者来说 25 00:01:12,160 --> 00:01:13,640 主要有三种可以设置 26 00:01:13,640 --> 00:01:15,340 第一呢是choice 27 00:01:15,340 --> 00:01:16,000 选择 28 00:01:16,000 --> 00:01:17,240 从选项里选一个 29 00:01:17,240 --> 00:01:19,300 第二是打分 30 00:01:19,300 --> 00:01:21,040 然后第三是 31 00:01:21,040 --> 00:01:23,540 判断陈述是否成立 32 00:01:23,540 --> 00:01:26,340 它是0到1的一个真值的分数 33 00:01:26,340 --> 00:01:27,740 一个请求里呢 34 00:01:27,740 --> 00:01:29,080 可以混合多种问题 35 00:01:29,080 --> 00:01:32,060 模型还可以并行去回答 36 00:01:32,060 --> 00:01:33,120 官方的手法呢 37 00:01:33,120 --> 00:01:35,920 就是把它当成软件里的智能的IF 38 00:01:35,920 --> 00:01:37,540 路由打分器 39 00:01:37,540 --> 00:01:38,600 JIF的速度呢 40 00:01:38,600 --> 00:01:40,100 是非常非常快的 41 00:01:40,100 --> 00:01:42,040 而且它的价格尤其的便宜 42 00:01:42,040 --> 00:01:43,440 我掉落了200多次 43 00:01:43,440 --> 00:01:45,400 花费了200分 44 00:01:45,400 --> 00:01:46,520 而且Type Safe呢 45 00:01:46,520 --> 00:01:49,680 现在还提供五刀的一个体验额度 46 00:01:49,680 --> 00:01:51,780 所以我得花很长时间 47 00:01:51,780 --> 00:01:52,360 才能把这个 48 00:01:52,360 --> 00:01:55,020 才能把这个五刀的额度给使用出去 49 00:01:55,020 --> 00:01:58,620 我们可以在TyperSafe里面去加入它的Whitlist 50 00:01:58,620 --> 00:02:00,620 如果你特别着急的话 51 00:02:00,620 --> 00:02:03,140 也可以通过OperRouter进行调用 52 00:02:03,140 --> 00:02:06,460 或者是AI Gateway进行调用 53 00:02:06,460 --> 00:02:08,720 它也登录了Cloud Flare的AI 54 00:02:08,720 --> 00:02:10,960 所以是有多个路径可以调用它的 55 00:02:10,960 --> 00:02:12,840 而且它非常有意思的是 56 00:02:12,840 --> 00:02:15,480 它的输出Token是不收费的 57 00:02:15,480 --> 00:02:16,620 官方给出的理由是 58 00:02:16,620 --> 00:02:19,920 新的架构下输出便宜到不必寄费 59 00:02:19,920 --> 00:02:21,520 这里就有一个例子呢 60 00:02:21,520 --> 00:02:23,240 Verso的CTO做了一个测试 61 00:02:23,240 --> 00:02:28,420 它呢对比调用Gemini 2.5 Flashlight来做分类 62 00:02:28,420 --> 00:02:30,120 Gemini 2.5 Flashlight呢 63 00:02:30,120 --> 00:02:31,900 它的API还是比较便宜的 64 00:02:31,900 --> 00:02:34,640 而将Jeff和Gemini做一个对比 65 00:02:34,640 --> 00:02:35,980 它的据确率更高 66 00:02:35,980 --> 00:02:37,520 而且速度也更快 67 00:02:37,520 --> 00:02:38,480 通过这张图 68 00:02:38,480 --> 00:02:39,880 我们可以看到Jeff 69 00:02:39,880 --> 00:02:42,900 它的质地程度是非常不错的 70 00:02:42,900 --> 00:02:45,600 非常接近于Terra了 71 00:02:45,600 --> 00:02:47,740 但是要比Terra便宜非常多 72 00:02:47,740 --> 00:02:50,200 我们可以在Type-Safe官网右上角 73 00:02:50,200 --> 00:02:54,280 点击他的加入waitlist来申请 74 00:02:54,280 --> 00:02:56,120 现在就来看一下X上面 75 00:02:56,120 --> 00:02:59,000 关于Jeff有哪些精彩的案例 76 00:02:59,000 --> 00:03:00,660 Versal的CEO认为 77 00:03:00,660 --> 00:03:03,220 把他们的FX安全审查 78 00:03:03,220 --> 00:03:04,220 换成Jeff 79 00:03:04,220 --> 00:03:06,060 对面GPT和Luna呢 80 00:03:06,060 --> 00:03:08,340 Jeff的速度快到18倍 81 00:03:08,340 --> 00:03:09,580 而且还更准确 82 00:03:09,580 --> 00:03:12,360 把500封邮件发给Jeff 83 00:03:12,360 --> 00:03:15,380 他仍在几秒内就帮你分类好 84 00:03:15,380 --> 00:03:17,140 而且成本非常低 85 00:03:17,140 --> 00:03:18,680 Boss CEO认为呢 86 00:03:18,680 --> 00:03:22,000 企业里那些分类路由安全判断们 87 00:03:22,000 --> 00:03:23,040 那现在有了Jeff 88 00:03:23,040 --> 00:03:25,800 对大量的企业任务就意义巨大 89 00:03:25,800 --> 00:03:28,020 有用户昨天刚打到权限 90 00:03:28,020 --> 00:03:31,600 他就将Jeff集成到他们的数据管道里 91 00:03:31,600 --> 00:03:33,980 实体识别情绪分析 92 00:03:33,980 --> 00:03:36,080 把相关信号粘起来等等 93 00:03:36,080 --> 00:03:37,880 他认为都非常惊艳 94 00:03:37,880 --> 00:03:42,640 Jeff有一大应用就是他在浏览器方面的改进 95 00:03:42,640 --> 00:03:44,960 这浏览器方面的例子 96 00:03:44,960 --> 00:03:47,300 比如说七秒查机票 97 00:03:47,300 --> 00:03:49,800 只花了0.0039刀 98 00:03:49,800 --> 00:03:50,480 又快 99 00:03:50,480 --> 00:03:51,500 质量又好 100 00:03:51,500 --> 00:03:53,160 知名开源产品CV 101 00:03:53,160 --> 00:03:54,680 它也是紧跟潮流 102 00:03:54,680 --> 00:03:56,300 立即发了一个 103 00:03:56,300 --> 00:04:00,180 和Jeff相关的跨系统计算机使用 104 00:04:00,180 --> 00:04:01,600 叫Jeff Use 105 00:04:01,600 --> 00:04:03,300 在玩2048的时候 106 00:04:03,300 --> 00:04:05,080 Jeff的速度快了5倍 107 00:04:05,080 --> 00:04:07,080 成本大约便宜1000倍 108 00:04:07,080 --> 00:04:08,740 比Astra更高效 109 00:04:08,740 --> 00:04:11,940 还有用户将Jeff和OpenCode结合起来 110 00:04:11,940 --> 00:04:12,680 测试应用 111 00:04:12,680 --> 00:04:14,320 测试的时候就发现 112 00:04:14,320 --> 00:04:17,280 点按钮快到像是在看加速片 113 00:04:17,280 --> 00:04:22,200 Jeff呢他在浏览器工作上是非常非常便宜的 114 00:04:22,200 --> 00:04:24,020 比如说这样一位用户 115 00:04:24,020 --> 00:04:27,100 他就分享浏览器忘记关了 116 00:04:27,100 --> 00:04:30,620 那消耗了这么多token才花费不到七刀 117 00:04:30,620 --> 00:04:32,240 Scode的作者呢 118 00:04:32,240 --> 00:04:36,360 他用Jeff来控制iOS的模拟器 119 00:04:36,360 --> 00:04:38,700 去建一个日历事件 120 00:04:38,700 --> 00:04:41,460 我们会看到这边会写着 121 00:04:41,460 --> 00:04:41,880 Taper 122 00:04:41,880 --> 00:04:44,800 然后ScreenChanger 123 00:04:44,800 --> 00:04:46,580 ScreenChanger 124 00:04:46,580 --> 00:04:47,600 等等 125 00:04:47,600 --> 00:04:49,640 还有用户喊话Sam 126 00:04:49,640 --> 00:04:52,240 让Codex负责规划 127 00:04:52,240 --> 00:04:54,940 Jeff负责从页面选择点谁 128 00:04:54,940 --> 00:04:57,400 有用户他分享了一个点子 129 00:04:57,400 --> 00:05:00,740 让Jeff去做及时的压缩 130 00:05:00,740 --> 00:05:02,980 给每次2Core打分 131 00:05:02,980 --> 00:05:05,020 然后丢掉无关的上下文 132 00:05:05,020 --> 00:05:08,000 我们可以看一下这样的一个例子 133 00:05:08,000 --> 00:05:09,800 绿色的就是留下的 134 00:05:09,800 --> 00:05:11,180 红色的就是删去 135 00:05:11,180 --> 00:05:14,820 也有用户还做了一个Jeff的review 136 00:05:14,820 --> 00:05:19,240 给编码agent接上打分反馈还的一个mcp 137 00:05:19,240 --> 00:05:21,220 不同的ai agent 138 00:05:21,220 --> 00:05:24,640 然后review之后进行打分 139 00:05:24,640 --> 00:05:29,240 给定他一个数值到了哪个程度 140 00:05:29,240 --> 00:05:30,720 那就让他再去优化 141 00:05:30,720 --> 00:05:33,940 有用户就直接做了一个cloud的插件 142 00:05:33,940 --> 00:05:39,100 一秒就把绘画从近100万token上下无人长度 143 00:05:39,100 --> 00:05:41,020 压缩到8.6万 144 00:05:41,020 --> 00:05:44,700 这里展示压缩前后的一些情况 145 00:05:44,700 --> 00:05:45,640 有兴趣的话 146 00:05:45,640 --> 00:05:46,540 大家去看一下 147 00:05:46,540 --> 00:05:47,940 Rosa的工程师呢 148 00:05:47,940 --> 00:05:49,340 说每个Agent 149 00:05:49,340 --> 00:05:50,540 Hannis呢 150 00:05:50,540 --> 00:05:52,460 现在都能用上Jeff 151 00:05:52,460 --> 00:05:54,980 通过这样的工具 152 00:05:54,980 --> 00:05:55,880 也有用户呢 153 00:05:55,880 --> 00:05:57,340 给OpenCode写了一个插件 154 00:05:57,340 --> 00:05:58,960 自栏语言写规则 155 00:05:58,960 --> 00:05:59,760 Jeff呢 156 00:05:59,760 --> 00:06:01,220 拦住每一条道路 157 00:06:01,220 --> 00:06:02,220 还有用户 158 00:06:02,220 --> 00:06:03,620 使用Jeff呢 159 00:06:03,620 --> 00:06:05,020 光速给提交 160 00:06:05,020 --> 00:06:06,540 打安全补丁的标签 161 00:06:06,540 --> 00:06:07,700 当然Jeff呢 162 00:06:07,700 --> 00:06:09,340 来做意向调研 163 00:06:09,340 --> 00:06:10,380 也是非常好的 164 00:06:10,380 --> 00:06:12,080 比如说这样的一位用户 165 00:06:12,080 --> 00:06:13,180 做的一个分享 166 00:06:13,180 --> 00:06:15,420 也有用户花了8个小时 167 00:06:15,420 --> 00:06:18,500 让Jeff做一个爆款帖分类器 168 00:06:18,500 --> 00:06:20,740 他认为Jeff的准确率呢 169 00:06:20,740 --> 00:06:22,800 和Fibber 5.1一样好 170 00:06:22,800 --> 00:06:24,100 但是快了100倍 171 00:06:24,100 --> 00:06:25,320 也有博主呢 172 00:06:25,320 --> 00:06:28,840 让Jeff去分析自己的3000多条帖子 173 00:06:28,840 --> 00:06:31,520 然后找出一些公式 174 00:06:31,520 --> 00:06:33,920 最后这是推荐的主题 175 00:06:33,920 --> 00:06:35,460 也有用户分享 176 00:06:35,460 --> 00:06:38,040 我们可以通过Jeff呢 177 00:06:38,040 --> 00:06:39,700 做一些规则匹配 178 00:06:39,700 --> 00:06:42,400 就可以傻瓜式的 179 00:06:42,400 --> 00:06:44,660 将一些文件进行分类了 180 00:06:44,660 --> 00:06:45,780 还有用户呢 181 00:06:45,780 --> 00:06:48,400 把Jeff接到自动驾驶的决策上 182 00:06:48,400 --> 00:06:49,900 他认为Jeff 183 00:06:49,900 --> 00:06:52,800 对自动驾驶的影响会非常大 184 00:06:52,800 --> 00:06:53,960 一个用户呢 185 00:06:53,960 --> 00:06:56,100 让Jeff像人一样 186 00:06:56,100 --> 00:06:57,600 来复原一个魔方 187 00:06:57,600 --> 00:06:59,380 他能在很快的时间内 188 00:06:59,380 --> 00:07:02,040 来对这个魔方进行转动 189 00:07:02,040 --> 00:07:03,760 视频里是放慢了 190 00:07:03,760 --> 00:07:05,360 方便我们去看清楚 191 00:07:05,360 --> 00:07:06,580 还有用户呢 192 00:07:06,580 --> 00:07:09,340 他说他花了一周时间 193 00:07:09,340 --> 00:07:10,920 才把PPO策略 194 00:07:10,920 --> 00:07:15,020 训练到通关的超级马利奥世界里 195 00:07:15,020 --> 00:07:18,180 这里的PPO是尽端策略优化 196 00:07:18,180 --> 00:07:21,460 而Jeff他是花了30分钟就打通了 197 00:07:21,460 --> 00:07:24,080 非常非常恐怖的一个效率 198 00:07:24,080 --> 00:07:26,320 有用户就来畅想了 199 00:07:26,320 --> 00:07:29,740 一个Jeff在云电脑上玩Minecraft 200 00:07:29,740 --> 00:07:32,480 那如果想象1000个11个 201 00:07:32,480 --> 00:07:35,460 每一个都在云端自己的电脑上协作 202 00:07:35,460 --> 00:07:39,140 发现新科学 跑实验 设计新药 203 00:07:39,140 --> 00:07:40,560 而且是全天候的 204 00:07:40,560 --> 00:07:44,040 也有用户让Jeff做一些实时的广告拦截 205 00:07:44,040 --> 00:07:47,620 比如说他做了一个插件 206 00:07:47,620 --> 00:07:49,960 然后进入了一个网站 207 00:08:03,960 --> 00:08:05,520 Baby AGI的作者呢 208 00:08:05,520 --> 00:08:06,920 还用Jeff来抽图 209 00:08:06,920 --> 00:08:09,000 我们可以看一下这里的demo 210 00:08:09,000 --> 00:08:10,020 Fluss的作者呢 211 00:08:10,020 --> 00:08:12,840 对Jeff非常非常认可 212 00:08:12,840 --> 00:08:14,240 他认为以前很多场景 213 00:08:14,240 --> 00:08:15,280 用大语言模型 214 00:08:15,280 --> 00:08:16,640 是根本不成立的 215 00:08:16,640 --> 00:08:17,920 今天大家也可以看到 216 00:08:17,920 --> 00:08:19,980 我为什么梳理了那么多案例 217 00:08:19,980 --> 00:08:21,660 实在是Jeff太优秀了 218 00:08:21,660 --> 00:08:23,640 真的是有非常非常多 219 00:08:23,640 --> 00:08:26,260 可以和我们现在的agent 220 00:08:26,260 --> 00:08:28,200 结合在一起使用的玩法 221 00:08:28,200 --> 00:08:29,420 有用户认为呢 222 00:08:29,420 --> 00:08:31,020 Hannis就会被改写 223 00:08:31,020 --> 00:08:32,800 无论是Codas CC 224 00:08:32,800 --> 00:08:34,940 内置这样的dev型的决策 225 00:08:34,940 --> 00:08:36,400 只是时间的问题 226 00:08:36,400 --> 00:08:38,280 Vestor CEO认为 227 00:08:38,280 --> 00:08:41,680 有很多是AI炒的一些热点 228 00:08:41,680 --> 00:08:45,320 但是Jeff是真正站得住的那个产品 229 00:08:45,320 --> 00:08:48,580 他认为Jeff这种AI 230 00:08:48,580 --> 00:08:50,980 可能比今天的大元模型 231 00:08:50,980 --> 00:08:54,060 更持久的改写我们的工作和产品 232 00:08:54,060 --> 00:08:57,640 因为Jeff输入是非常非常划算的 233 00:08:57,640 --> 00:08:59,660 所以大量的自动化 234 00:08:59,660 --> 00:09:01,760 就会突然变得买得起 235 00:09:01,760 --> 00:09:02,780 说了那么多 236 00:09:02,780 --> 00:09:06,040 已经有人做了一个非常好的仓库 237 00:09:06,040 --> 00:09:07,520 是Orzen Type Safe 238 00:09:07,520 --> 00:09:08,540 仓库里呢 239 00:09:08,540 --> 00:09:09,880 有官方的资源 240 00:09:09,880 --> 00:09:11,840 也有很多社区的项目 241 00:09:11,840 --> 00:09:14,280 比如客户端库和集成 242 00:09:14,280 --> 00:09:16,080 Agent和开发者工具 243 00:09:16,080 --> 00:09:17,560 浏览器智能体 244 00:09:17,560 --> 00:09:19,380 机器人交互演示 245 00:09:19,380 --> 00:09:22,120 评测与独立研究等等 246 00:09:22,120 --> 00:09:23,520 非常非常多 247 00:09:23,520 --> 00:09:24,480 而Jeff呢 248 00:09:24,480 --> 00:09:27,400 发布也才两天左右的时间而已 249 00:09:27,400 --> 00:09:28,480 我们使用Jeff呢 250 00:09:28,480 --> 00:09:29,400 也非常简单 251 00:09:29,400 --> 00:09:31,660 进入他的控制台 252 00:09:31,660 --> 00:09:32,920 它有个Quick Start 253 00:09:32,920 --> 00:09:35,420 然后我们可以copy这里的 254 00:09:35,420 --> 00:09:37,320 给Agent的提示词 255 00:09:37,320 --> 00:09:39,280 就可以在我们电脑上 256 00:09:39,280 --> 00:09:41,320 安装Type Safe的Skill 257 00:09:41,320 --> 00:09:42,520 比如这里呢 258 00:09:42,520 --> 00:09:44,580 我呢就调用它的Skill 259 00:09:44,580 --> 00:09:47,120 然后让它帮我做一个 260 00:09:47,120 --> 00:09:48,980 打游戏的Web页面 261 00:09:48,980 --> 00:09:51,520 JPT很快就帮我做出了这样一个游戏 262 00:09:51,520 --> 00:09:52,980 第一章呢 263 00:09:52,980 --> 00:09:55,300 我刚刚已经让Jeff已经游玩过了 264 00:09:55,300 --> 00:09:57,840 我们先看一下这里的游戏指南 265 00:09:57,840 --> 00:09:59,900 可以让Jeff来玩 266 00:09:59,900 --> 00:10:01,600 也可以手动接管 267 00:10:01,600 --> 00:10:03,780 总共有三个章节 268 00:10:03,780 --> 00:10:07,660 Choice它会显示后选动作的真实概率 269 00:10:07,660 --> 00:10:10,580 自信度衡量分布的集中程度 270 00:10:10,580 --> 00:10:15,340 Nord它是单独判断是否急需休息 271 00:10:15,340 --> 00:10:17,000 第一关已经打完了 272 00:10:17,000 --> 00:10:20,660 那现在我来点一下看一下第二关 273 00:10:20,660 --> 00:10:25,140 我们可以观察右侧Jeff的每一步 274 00:10:25,140 --> 00:10:28,460 今天的这个视频我是没有做变速 275 00:10:28,460 --> 00:10:30,240 所以大家可以观察一下 276 00:10:30,240 --> 00:10:32,860 杰弗他的一个速度情况 277 00:10:32,860 --> 00:10:34,440 现在来点击一下 278 00:10:34,440 --> 00:10:35,020 让他开始 279 00:10:35,020 --> 00:10:40,100 他开始收集到原木了 280 00:10:40,100 --> 00:10:41,460 有了原木之后 281 00:10:41,460 --> 00:10:43,080 这里的桥就可以搭建好了 282 00:10:43,080 --> 00:10:44,800 学好之后 283 00:10:44,800 --> 00:10:47,420 他就可以去获取对岸的胡萝卜 284 00:10:47,420 --> 00:10:50,560 然后还有一个胡萝卜 285 00:10:50,560 --> 00:10:51,560 他还没收集好 286 00:10:51,560 --> 00:10:52,800 所以他这里收集一下 287 00:10:52,800 --> 00:10:55,400 再回到他的家 288 00:10:55,400 --> 00:10:58,220 我可以把这里的蛋报呢 289 00:10:58,220 --> 00:10:58,940 进行保存 290 00:10:58,940 --> 00:11:00,140 也就是说 291 00:11:00,140 --> 00:11:02,040 他会下载一个JSON文件 292 00:11:02,040 --> 00:11:03,440 右侧我们会看到 293 00:11:03,440 --> 00:11:05,620 他的当前章节的目标 294 00:11:05,620 --> 00:11:07,100 这里的概率呢 295 00:11:07,100 --> 00:11:08,520 他就写的比较大 296 00:11:08,520 --> 00:11:09,620 所以兔子呢 297 00:11:09,620 --> 00:11:11,000 最后就回到引力去了 298 00:11:11,000 --> 00:11:12,760 在探险偏好里呢 299 00:11:12,760 --> 00:11:14,720 也可以去选择一下 300 00:11:14,720 --> 00:11:15,920 页面的下方 301 00:11:15,920 --> 00:11:17,940 还可以看到他的冒险记录 302 00:11:17,940 --> 00:11:19,760 还有他的休息的情况 303 00:11:19,760 --> 00:11:21,720 我呢之前将第一个章节 304 00:11:21,720 --> 00:11:23,540 玩的一个战报 305 00:11:23,540 --> 00:11:25,580 然后下载了下来 306 00:11:25,580 --> 00:11:27,400 让AI介绍一下 307 00:11:27,400 --> 00:11:29,360 他说呢 308 00:11:29,360 --> 00:11:30,460 回合 309 00:11:30,460 --> 00:11:31,360 九回合 310 00:11:31,360 --> 00:11:32,680 结局胜利 311 00:11:32,680 --> 00:11:35,200 体力还剩37 312 00:11:35,200 --> 00:11:37,140 葛萝卜的话有五根 313 00:11:37,140 --> 00:11:39,120 第一章节里面 314 00:11:39,120 --> 00:11:40,860 它是没有木头的 315 00:11:40,860 --> 00:11:41,560 下方呢 316 00:11:41,560 --> 00:11:43,300 是兔子走过的路线 317 00:11:43,300 --> 00:11:45,460 日志对应的是什么 318 00:11:45,460 --> 00:11:46,540 决策记录里 319 00:11:46,540 --> 00:11:47,420 我们会看到 320 00:11:47,420 --> 00:11:49,640 用的是Jave的模型 321 00:11:49,640 --> 00:11:51,640 然后第一步是比较慢 322 00:11:51,640 --> 00:11:53,300 这和我当前的网络 323 00:11:53,300 --> 00:11:54,680 当然也有很大关系 324 00:11:54,680 --> 00:11:56,440 后面的话快乐非常多 325 00:11:56,440 --> 00:11:58,680 大多数的致敬度呢 326 00:11:58,680 --> 00:12:00,760 在0.77到0.96 327 00:12:00,760 --> 00:12:02,820 其中有两步 328 00:12:02,820 --> 00:12:03,500 有欲 329 00:12:03,500 --> 00:12:05,200 因为一开始这个桥呢 330 00:12:05,200 --> 00:12:06,300 没有修复好 331 00:12:06,300 --> 00:12:07,400 所以 332 00:12:07,400 --> 00:12:09,360 Jeff呢就来回 333 00:12:09,360 --> 00:12:11,060 在这三个洪框的点 334 00:12:11,060 --> 00:12:12,300 这里来回折腾 335 00:12:12,300 --> 00:12:15,460 后来我就让Astra来修复了一下 336 00:12:15,460 --> 00:12:15,940 这个游戏 337 00:12:15,940 --> 00:12:17,120 现在我们再来到 338 00:12:17,120 --> 00:12:18,540 他的第三章节 339 00:12:18,540 --> 00:12:19,760 那我这里呢 340 00:12:19,760 --> 00:12:21,720 点击一下手动接管 341 00:12:21,720 --> 00:12:23,160 音效呢也打开了 342 00:12:23,160 --> 00:12:24,480 那我点 343 00:12:24,480 --> 00:12:25,440 点一个 344 00:12:25,440 --> 00:12:26,240 比如说 345 00:12:26,240 --> 00:12:28,300 我这里先点了除菊小镜 346 00:12:28,300 --> 00:12:30,440 再点一个萤火炒坡 347 00:12:30,440 --> 00:12:32,920 然后我们会看到 348 00:12:32,920 --> 00:12:34,000 还有胡萝卜 349 00:12:34,000 --> 00:12:35,740 然后这里还有木头 350 00:12:35,740 --> 00:12:36,660 没有收集到 351 00:12:36,660 --> 00:12:40,720 那我再返回别的地方 352 00:12:40,720 --> 00:12:41,560 点一下 353 00:12:41,560 --> 00:12:44,780 这里收集到木头了 354 00:12:44,780 --> 00:12:46,980 再点一下 355 00:12:46,980 --> 00:12:49,160 他说等待修船 356 00:12:49,160 --> 00:12:51,340 等等 357 00:12:51,340 --> 00:12:54,260 那再来让他单独来玩一下 358 00:12:54,260 --> 00:12:55,280 看看他怎么玩 359 00:12:55,280 --> 00:13:02,140 修好桥了 360 00:13:02,140 --> 00:13:03,540 然后再去取胡萝卜 361 00:13:03,540 --> 00:13:05,100 又返回来 362 00:13:05,100 --> 00:13:06,880 它要去取胡萝卜 363 00:13:06,880 --> 00:13:12,240 因为这个设定里呢 364 00:13:12,240 --> 00:13:14,280 它还需要一个金种子 365 00:13:14,280 --> 00:13:15,940 开启石门 366 00:13:15,940 --> 00:13:18,320 所以它最后是在右侧 367 00:13:18,320 --> 00:13:20,520 游戏那里去结束的 368 00:13:20,520 --> 00:13:22,720 最后呢再来看一个案例 369 00:13:22,720 --> 00:13:25,060 我呢让Astra做的一个页面 370 00:13:25,060 --> 00:13:26,680 外星人要到地球 371 00:13:26,680 --> 00:13:29,700 那要把外星人的奇怪发言 372 00:13:29,700 --> 00:13:31,860 变成代码执行的判断 373 00:13:31,860 --> 00:13:33,780 左边呢 374 00:13:33,780 --> 00:13:35,420 就会有非常多的 375 00:13:35,420 --> 00:13:37,060 待检的旅客 376 00:13:37,060 --> 00:13:39,220 总共是有24个 377 00:13:39,220 --> 00:13:40,340 他们呢 378 00:13:40,340 --> 00:13:42,520 长的是有三种样式 379 00:13:42,520 --> 00:13:44,580 名字有各个不同 380 00:13:44,580 --> 00:13:47,160 比如说这里的白闪博士 381 00:13:47,160 --> 00:13:49,280 他说旅客的原话 382 00:13:49,280 --> 00:13:50,440 是什么样子的 383 00:13:50,440 --> 00:13:52,880 那正常这种情况下 384 00:13:52,880 --> 00:13:56,420 他说他有简易放行证明 385 00:13:56,420 --> 00:13:59,400 点击一下开始扫描 386 00:13:59,400 --> 00:14:02,260 肯定是让他去通过的 387 00:14:02,260 --> 00:14:03,140 Jeff这里呢 388 00:14:03,140 --> 00:14:04,560 他有三个判断 389 00:14:04,560 --> 00:14:05,700 第一是Choice 390 00:14:05,700 --> 00:14:08,000 第二呢是No 391 00:14:08,000 --> 00:14:09,580 那对这样外星人呢 392 00:14:09,580 --> 00:14:11,780 我们就不需要过度的去关注他 393 00:14:11,780 --> 00:14:15,060 然后下方就有一个欢迎雨 394 00:14:15,060 --> 00:14:16,360 欢迎来到地球 395 00:14:16,360 --> 00:14:17,560 决策档案里呢 396 00:14:17,560 --> 00:14:20,100 就会显示刚刚做了那些决策 397 00:14:20,100 --> 00:14:22,860 实测的实验室里呢 398 00:14:22,860 --> 00:14:26,900 这是之前跑过的一次例子 399 00:14:26,900 --> 00:14:30,380 这是24条人工编写的一个测试情境 400 00:14:30,380 --> 00:14:31,680 每条呢 401 00:14:31,680 --> 00:14:33,940 向真实的Jeff来发送 402 00:14:33,940 --> 00:14:36,300 同一state下的三个独立问题 403 00:14:36,300 --> 00:14:39,180 参考标签不会传给模型 404 00:14:39,180 --> 00:14:40,460 比如说 405 00:14:40,460 --> 00:14:42,940 这里的参考通道啊 406 00:14:42,940 --> 00:14:44,440 是哪个 407 00:14:44,440 --> 00:14:46,920 那就不会发给Jeff 408 00:14:46,920 --> 00:14:48,380 通过这个测试呢 409 00:14:48,380 --> 00:14:51,300 就可以比对人工写的参考信息 410 00:14:51,300 --> 00:14:55,160 和GF判断了结果之间的一些差别 411 00:14:55,160 --> 00:14:57,920 最后的对照里面会说明他们的状态 412 00:14:57,920 --> 00:15:01,020 那我这里再点击一下重新实测 413 00:15:01,020 --> 00:15:05,180 我是开了最多的10号并发 414 00:15:05,180 --> 00:15:08,620 那我们可以看到它是成功完成 415 00:15:08,620 --> 00:15:11,020 参考是有91% 416 00:15:11,020 --> 00:15:12,940 是一致的 417 00:15:12,940 --> 00:15:15,980 耗时当然是非常短的 418 00:15:15,980 --> 00:15:17,480 实际耗用了这些Token 419 00:15:17,480 --> 00:15:21,380 那我再点一下来进行测试 420 00:15:21,380 --> 00:15:26,660 这个刚刚的测试结果也是一样的 421 00:15:26,660 --> 00:15:30,000 不过呢每次测试不一定是完全一样的 422 00:15:30,000 --> 00:15:32,840 所以有的时候可能这里是百分百 423 00:15:32,840 --> 00:15:36,540 然后下方我们可以看一下 424 00:15:36,540 --> 00:15:38,300 有哪些地方是不一致 425 00:15:38,300 --> 00:15:42,820 比如说这对这个外星人的一个分析 426 00:15:42,820 --> 00:15:47,200 然后参考认为是可以常规路径的 427 00:15:47,200 --> 00:15:49,640 而Jeff认为需要深度调查 428 00:15:49,640 --> 00:15:53,280 对这两位外星人人工分析 429 00:15:53,280 --> 00:15:56,220 和Jeff判断有不一致的情况 430 00:15:56,220 --> 00:15:57,140 我们看一下 431 00:15:57,140 --> 00:16:00,520 他为什么对这里是要人工复合呢 432 00:16:00,520 --> 00:16:01,760 这位外星人他说 433 00:16:01,760 --> 00:16:06,060 他运送了一件武器外星的博物馆空壳 434 00:16:06,060 --> 00:16:10,120 内部完全猜出确认没有什么什么什么 435 00:16:10,120 --> 00:16:11,900 还有附有检查证明 436 00:16:11,900 --> 00:16:15,220 以上就是今天介绍Jeff的所有内容 437 00:16:15,220 --> 00:16:16,820 这个模型对于我们日常工作 438 00:16:16,820 --> 00:16:18,520 还是有非常大的帮助的 439 00:16:18,520 --> 00:16:20,460 非常推荐大家去了解它 440 00:16:20,460 --> 00:16:22,200 记得看社区里面 441 00:16:22,200 --> 00:16:24,360 Owson TypeSafe的仓库 442 00:16:24,360 --> 00:16:25,880 希望我的视频对你有帮助 443 00:16:25,880 --> 00:16:26,940 我们下期再见