WEBVTT
Kind: captions
Language: zh-TW

00:00:14.014 --> 00:00:16.983
這個8月Grok已經成長為無法忽視的規模了

00:00:16.983 --> 00:00:19.052
不僅模型升級為Grok 4.6

00:00:19.052 --> 00:00:21.354
產圖模型升級為Image 2.0

00:00:21.354 --> 00:00:24.357
甚至端出了全新的AI Agent工具Grok Bot

00:00:24.357 --> 00:00:25.258
震撼市場

00:00:25.258 --> 00:00:27.427
其他三家AI大廠也沒閒著

00:00:27.427 --> 00:00:27.961
首先

00:00:27.961 --> 00:00:30.363
ChatGPT從7月開始就瘋狂寵粉

00:00:30.363 --> 00:00:31.598
除了靠不斷重置

00:00:31.598 --> 00:00:33.333
大量吸引Claude用戶跳槽

00:00:33.333 --> 00:00:36.436
8月7日更下放GPT-5.6 Luna模型

00:00:36.436 --> 00:00:38.204
給免費版還有Go用戶

00:00:38.204 --> 00:00:38.805
此外

00:00:38.805 --> 00:00:41.674
在設定裡面還新增開啟自動同步的功能

00:00:41.674 --> 00:00:43.209
可以同步抓到電腦裡面

00:00:43.209 --> 00:00:45.245
使用其他AI軟體的開發紀錄

00:00:45.245 --> 00:00:48.148
對於同時使用Claude、ChatGPT的用戶來說

00:00:48.148 --> 00:00:49.983
這簡直就是天大的福音

00:00:49.983 --> 00:00:52.552
Claude則在7月底推出Opus 5模型

00:00:52.552 --> 00:00:53.686
並在近期確定

00:00:53.686 --> 00:00:56.389
Sonnet 5現在的優惠價已經成為定價

00:00:56.389 --> 00:00:57.557
未來不會漲價

00:00:57.557 --> 00:00:58.892
一般用戶最有感的

00:00:58.892 --> 00:01:01.127
則是現在Chrome側邊欄的擴充功能

00:01:01.127 --> 00:01:02.362
變成Claude Cowork

00:01:02.362 --> 00:01:03.797
你在瀏覽器上對話

00:01:03.797 --> 00:01:05.865
都可以同步在桌面的軟體出現

00:01:05.865 --> 00:01:07.300
最後來看一下Gemini

00:01:07.300 --> 00:01:07.867
唉

00:01:07.867 --> 00:01:09.702
等了整整三個月的旗艦模型

00:01:09.702 --> 00:01:10.670
3.5 Pro

00:01:10.670 --> 00:01:12.472
似乎已經確定消失了

00:01:12.472 --> 00:01:14.607
但在上週他們推出了全新模型

00:01:14.607 --> 00:01:15.575
3.7 Flash

00:01:15.575 --> 00:01:18.445
而且在美國還開放了14個應用程式

00:01:18.445 --> 00:01:19.512
可以連結Gemini

00:01:19.512 --> 00:01:21.347
對於封閉Google生態圈來說

00:01:21.347 --> 00:01:22.515
算是很大的進步

00:01:22.515 --> 00:01:22.849
現在

00:01:22.849 --> 00:01:26.186
馬上帶領你們深度認識美國四大AI更新內容

00:01:27.487 --> 00:01:29.189
Google現在最該緊張的對手

00:01:29.189 --> 00:01:30.123
不是OpenAI

00:01:30.123 --> 00:01:30.824
是Grok

00:01:30.824 --> 00:01:31.491
為什麼呢

00:01:31.491 --> 00:01:33.693
因為Google這幾年的護城河一直是

00:01:33.693 --> 00:01:35.662
又便宜、又快、又什麼都能做

00:01:35.662 --> 00:01:38.364
文字、圖片、影片全包在同一個生態系裡

00:01:38.364 --> 00:01:39.432
而Grok這個月

00:01:39.432 --> 00:01:41.234
每一項更新都瞄準Google

00:01:41.234 --> 00:01:42.068
先看模型

00:01:42.068 --> 00:01:43.136
8月12號

00:01:43.136 --> 00:01:44.737
Grok 4.6正式登場

00:01:44.737 --> 00:01:46.706
價格是輸入每百萬Token

00:01:46.706 --> 00:01:47.540
2美金

00:01:47.540 --> 00:01:48.808
輸出6美金

00:01:48.808 --> 00:01:50.777
大約是其他前沿模型的一半

00:01:50.777 --> 00:01:54.080
在Artificial Analysis智力指數上拿到61分

00:01:54.080 --> 00:01:55.682
追平OpenAI的旗艦

00:01:55.682 --> 00:01:57.117
GPT-5.6 Sol

00:01:57.117 --> 00:01:59.519
它最強的是長時間的Agent任務

00:01:59.519 --> 00:02:00.553
官方說法是

00:02:00.553 --> 00:02:03.323
特別擅長把一個粗略的產品想法

00:02:03.323 --> 00:02:05.158
直接做成能動的第一版

00:02:05.158 --> 00:02:06.559
而且跑越久越穩

00:02:06.559 --> 00:02:07.961
因為它會自我測試

00:02:07.961 --> 00:02:08.828
自我驗證

00:02:08.828 --> 00:02:09.629
再看圖片

00:02:09.629 --> 00:02:11.397
Imagine Image 2.0

00:02:11.397 --> 00:02:12.532
8月7號上線

00:02:12.532 --> 00:02:14.434
這一代主打精準編輯

00:02:14.434 --> 00:02:16.603
魔法棒工具點哪裡改哪裡

00:02:16.603 --> 00:02:18.471
一次最多吃五張參考圖

00:02:18.471 --> 00:02:20.406
角色產品場景直接合體

00:02:20.406 --> 00:02:22.208
文字算圖也大升級啊

00:02:22.208 --> 00:02:24.077
海報上的小字不會糊掉

00:02:24.077 --> 00:02:25.311
最後看影片

00:02:25.311 --> 00:02:27.714
Imagine Video 1.5月初升級

00:02:27.714 --> 00:02:29.182
支援文字生影片

00:02:29.182 --> 00:02:30.750
圖像跟語音參考

00:02:30.750 --> 00:02:32.785
跟Google的Veo 3.1比起來

00:02:32.785 --> 00:02:33.720
評測顯示

00:02:33.720 --> 00:02:35.021
Veo 3.1在物理表現

00:02:35.021 --> 00:02:35.955
人體動作

00:02:35.955 --> 00:02:37.657
這種硬指標上還是比較準

00:02:37.657 --> 00:02:38.925
單支最長8秒

00:02:38.925 --> 00:02:39.459
Grok

00:02:39.459 --> 00:02:41.094
單支能做到15秒

00:02:41.094 --> 00:02:42.629
音效是原生內建的

00:02:42.629 --> 00:02:44.164
生成大概30秒就好

00:02:44.164 --> 00:02:45.899
成本比Veo 3.1低了7成5

00:02:45.899 --> 00:02:47.000
到將近9成

00:02:47.000 --> 00:02:47.967
一句話總結

00:02:47.967 --> 00:02:50.103
要拍廣告要精緻選Veo 3.1

00:02:50.103 --> 00:02:52.138
要一天產20支社群短影音

00:02:52.138 --> 00:02:52.839
選Grok

00:02:52.839 --> 00:02:54.674
而這個月Grok讓人驚豔的

00:02:54.674 --> 00:02:55.675
還不只這樣

00:02:55.675 --> 00:02:57.610
xAI併入SpaceX之後

00:02:57.610 --> 00:02:58.811
終於放出了大招

00:02:58.811 --> 00:02:59.913
就是Grok Bot

00:02:59.913 --> 00:03:01.681
但它的使用邏輯跟介面

00:03:01.681 --> 00:03:03.149
與你熟悉的ChatGPT

00:03:03.149 --> 00:03:04.417
Claude完全不一樣

00:03:04.417 --> 00:03:05.285
先看介面

00:03:05.285 --> 00:03:06.686
打開App你會發現

00:03:06.686 --> 00:03:08.588
它的介面走一個極簡風

00:03:08.588 --> 00:03:10.256
完全就是一個通訊軟體

00:03:10.256 --> 00:03:11.658
左邊一排是你的Bot

00:03:11.658 --> 00:03:12.959
右邊是對話框

00:03:12.959 --> 00:03:14.561
不得不吐槽LINE的介面

00:03:14.561 --> 00:03:15.495
還比這個複雜勒

00:03:15.495 --> 00:03:17.030
使用邏輯也完全不一樣

00:03:17.030 --> 00:03:18.097
你用Claude Cowork

00:03:18.097 --> 00:03:19.499
或ChatGPT的時候

00:03:19.499 --> 00:03:20.633
是「用任務思考」

00:03:20.633 --> 00:03:22.235
我要他做這件事

00:03:22.235 --> 00:03:24.404
Grok Bot要你「用角色思考」

00:03:24.404 --> 00:03:26.539
每一個Bot都是獨立的角色

00:03:26.539 --> 00:03:27.974
這一支負責我的信箱

00:03:27.974 --> 00:03:29.542
這一支負責找客戶

00:03:29.542 --> 00:03:31.778
而這一支負責幫我想影片大綱

00:03:31.778 --> 00:03:33.079
你不是在下指令

00:03:33.079 --> 00:03:35.114
你是在組一支AI團隊

00:03:35.114 --> 00:03:36.950
但在正式建置團隊之前

00:03:36.950 --> 00:03:38.751
建議你先建立一個Commander

00:03:38.751 --> 00:03:40.620
作為整體工作流的總指揮

00:03:40.620 --> 00:03:42.822
雖然Grok Bot支援自行建立

00:03:42.822 --> 00:03:44.157
但我的慣用做法

00:03:44.157 --> 00:03:46.159
是將第一個Bot設定為Commander

00:03:46.159 --> 00:03:48.194
並投餵你在網路上收集到的

00:03:48.194 --> 00:03:49.662
各種工作流程概念

00:03:49.662 --> 00:03:51.898
或是過去在其他AI開發過

00:03:51.898 --> 00:03:53.266
可重複使用的Skill

00:03:53.266 --> 00:03:55.001
讓Commander來負責指揮

00:03:55.001 --> 00:03:56.202
與協調整體運作

00:03:56.202 --> 00:03:58.504
當Commander設定好所需的Bot後

00:03:58.504 --> 00:04:00.440
只要將他們拉進同一個對話框

00:04:00.440 --> 00:04:01.341
就能透過@

00:04:01.341 --> 00:04:03.309
提及特定Bot來發送指令

00:04:03.309 --> 00:04:04.978
當然不使用@也沒問題

00:04:04.978 --> 00:04:06.246
因為每個Bot都具備

00:04:06.246 --> 00:04:07.981
自動判斷任務歸屬的能力

00:04:07.981 --> 00:04:10.250
這也正是Commander的核心價值所在

00:04:10.250 --> 00:04:11.818
只要前期設定夠完善

00:04:11.818 --> 00:04:14.020
Commander就能精準理解你的需求

00:04:14.020 --> 00:04:15.555
篩選並派發任務

00:04:15.555 --> 00:04:17.223
體驗上遠比讓各個Bot

00:04:17.223 --> 00:04:19.292
盲目猜測來的高效且順暢

00:04:19.292 --> 00:04:20.627
針對特定的對話

00:04:20.627 --> 00:04:22.195
你也可以啟動Start Thread

00:04:22.195 --> 00:04:23.596
進行更深入的探討

00:04:23.596 --> 00:04:25.398
這樣既可以在同一個聊天室內

00:04:25.398 --> 00:04:27.266
進行單一Bot的細部交流

00:04:27.266 --> 00:04:29.602
又能隨時切換回主線繼續溝通

00:04:29.602 --> 00:04:31.771
讓整個工作區更加井然有序

00:04:31.771 --> 00:04:34.040
不過即使單純透過對話來驅動

00:04:34.040 --> 00:04:36.175
基本上也都能順利完成任務

00:04:36.175 --> 00:04:38.845
像是文件生成這類常見的辦公室需求

00:04:38.845 --> 00:04:40.513
完全不需要呼叫API

00:04:40.513 --> 00:04:42.281
只要運用內部多個專屬Bot

00:04:42.281 --> 00:04:44.250
與Commander的指令配合即可

00:04:44.250 --> 00:04:46.653
裡面的Multi-Agent機制十分流暢

00:04:46.653 --> 00:04:49.322
絕不會發生彼此搶工作或衝突的情況

00:04:49.322 --> 00:04:50.957
若你曾使用過OpenClaw

00:04:50.957 --> 00:04:52.458
或類似的多機架構

00:04:52.458 --> 00:04:54.961
就會明白這類的設定通常非常繁瑣

00:04:54.961 --> 00:04:55.862
而Grok Bot

00:04:55.862 --> 00:04:57.797
不僅能輕鬆實現平行處理

00:04:57.797 --> 00:04:59.666
與即時排定Agent工作流程

00:04:59.666 --> 00:05:01.100
甚至還會自動建立

00:05:01.100 --> 00:05:03.102
負責品質把關的檢查Bot

00:05:03.102 --> 00:05:04.971
是目前市面上最完善

00:05:04.971 --> 00:05:06.639
且上手門檻最低的系統

00:05:06.639 --> 00:05:08.241
Grok Bot到底強在哪裡

00:05:08.241 --> 00:05:09.142
以下這兩件事

00:05:09.142 --> 00:05:11.511
Claude跟ChatGPT目前都做不到

00:05:11.511 --> 00:05:11.878
第一

00:05:11.878 --> 00:05:13.846
配一台全天候開著的雲端電腦

00:05:13.846 --> 00:05:16.416
裡面有瀏覽器、有檔案、有登入狀態

00:05:16.416 --> 00:05:17.784
而且是長期留著的

00:05:17.784 --> 00:05:18.484
差別在哪

00:05:18.484 --> 00:05:20.453
Claude Cowork 跟 ChatGPT 的 agent

00:05:20.453 --> 00:05:22.288
除非是排程在雲端的定期任務

00:05:22.288 --> 00:05:24.557
其他工作還是很仰賴地端的主機

00:05:24.557 --> 00:05:27.727
Grok Bot 則是「所有Bot使用一個獨立的雲端」

00:05:27.727 --> 00:05:30.630
所以你關機出門 它還在用雲端電腦工作

00:05:30.630 --> 00:05:31.831
想看它在雲端幹嘛

00:05:31.831 --> 00:05:33.733
使用App右上角的電腦圖示

00:05:33.733 --> 00:05:36.502
它開 Chrome、開終端機 畫面全都看得到

00:05:36.502 --> 00:05:38.037
第二多角色溝通

00:05:38.037 --> 00:05:39.706
你可以把多個Bot拉在一起

00:05:39.706 --> 00:05:40.673
同時下任務

00:05:40.673 --> 00:05:43.343
例如一個A Bot負責管專案預算

00:05:43.343 --> 00:05:44.877
B Bot負責訂便當

00:05:44.877 --> 00:05:47.447
C Bot負責安排拍攝行程

00:05:47.447 --> 00:05:48.915
你可以直接發派任務

00:05:48.915 --> 00:05:50.383
就像前面示範的樣子

00:05:50.383 --> 00:05:51.718
其他還有些小差異

00:05:51.718 --> 00:05:53.586
像是外掛不用進設定頁

00:05:53.586 --> 00:05:54.587
你在對話裡說

00:05:54.587 --> 00:05:55.655
「以後都存到 Notion」

00:05:55.655 --> 00:05:57.090
他當場跳授權

00:05:57.090 --> 00:05:58.691
排程也不用手動設

00:05:58.691 --> 00:05:59.292
直接講

00:05:59.292 --> 00:06:01.227
「每天早上八點幫我整理一次」

00:06:01.227 --> 00:06:02.929
也能用螢幕錄操作錄影

00:06:02.929 --> 00:06:04.897
就能讓Grok Bot重複執行

00:06:04.897 --> 00:06:06.365
聽起來好像很厲害

00:06:06.365 --> 00:06:07.734
但先別急著訂閱

00:06:07.734 --> 00:06:09.168
因為他現在貴到靠北

00:06:09.168 --> 00:06:10.536
如果你要使用Grok Bot

00:06:10.536 --> 00:06:12.205
要嘛訂Super Grok Heavy

00:06:12.205 --> 00:06:13.539
月費300美金

00:06:13.539 --> 00:06:14.974
要嘛走Cursor Ultra

00:06:14.974 --> 00:06:15.975
200美金

00:06:15.975 --> 00:06:17.243
或是Cursor團隊

00:06:17.243 --> 00:06:18.678
每席120美金

00:06:18.678 --> 00:06:20.146
一般訂閱X Premium

00:06:20.146 --> 00:06:21.681
或是Super Grok的用戶

00:06:21.681 --> 00:06:22.782
完全不能使用

00:06:22.782 --> 00:06:24.083
另外還有一個缺點是

00:06:24.083 --> 00:06:25.952
目前Bot也不能選模型

00:06:25.952 --> 00:06:27.453
你看不到他跑哪一支

00:06:27.453 --> 00:06:28.554
花了多少成本

00:06:28.554 --> 00:06:30.189
如果你還是想嘗試

00:06:30.189 --> 00:06:32.425
現在新用戶有7天免費試用

00:06:32.425 --> 00:06:33.393
下載軟體登錄

00:06:33.393 --> 00:06:34.560
就可以體驗看看

00:06:34.560 --> 00:06:36.629
不過因為他的額度真的非常少

00:06:36.629 --> 00:06:38.998
建議先跟其他AI軟體討論完目標

00:06:38.998 --> 00:06:41.367
來源跟限定範圍等條件都寫死

00:06:41.367 --> 00:06:42.735
才開始跑任務

00:06:44.036 --> 00:06:45.738
這個設定讓ChatGPT

00:06:45.738 --> 00:06:47.874
可能成為所有AI Agent的入口

00:06:47.874 --> 00:06:50.176
他就是「保持匯入同步」功能

00:06:50.176 --> 00:06:52.378
對於雙刀流的重度AI使用者來說

00:06:52.378 --> 00:06:54.547
這項功能大大提升工作效率

00:06:54.547 --> 00:06:56.015
因為他可以直接偵測到

00:06:56.015 --> 00:06:58.084
桌面上的其他AI應用程式

00:06:58.084 --> 00:06:59.418
並且可以直接同步

00:06:59.418 --> 00:07:01.120
更新到ChatGPT裡面

00:07:01.120 --> 00:07:02.688
舉凡你在Claude Code跟

00:07:02.688 --> 00:07:03.823
Cowork裡面的紀錄

00:07:03.823 --> 00:07:05.558
都可以直接彙整進來

00:07:05.558 --> 00:07:07.193
還有一項更可怕的更新

00:07:07.193 --> 00:07:08.661
叫做Computer History

00:07:08.661 --> 00:07:10.763
雖然這項功能目前只有在Mac電腦

00:07:10.763 --> 00:07:11.564
上才能使用

00:07:11.564 --> 00:07:14.634
但他可以讓你跟ChatGPT的互動大幅升級

00:07:14.634 --> 00:07:16.636
從原本記得我們聊過什麼

00:07:16.636 --> 00:07:19.739
變成「記得我今天在電腦上做過什麼」

00:07:19.739 --> 00:07:20.573
先講清楚

00:07:20.573 --> 00:07:23.109
他跟原本設定裡面就有的「本機記憶」

00:07:23.109 --> 00:07:24.210
不是互相取代

00:07:24.210 --> 00:07:25.678
而是上下層關係

00:07:25.678 --> 00:07:27.513
本機記憶是「記憶引擎」

00:07:27.513 --> 00:07:30.249
Computer History是新裝上去的「感測器」

00:07:30.249 --> 00:07:31.717
官方也寫得很明白

00:07:31.717 --> 00:07:34.120
你必須先打開Memories才能用

00:07:34.120 --> 00:07:35.054
Computer History

00:07:35.054 --> 00:07:36.255
那他到底記什麼?

00:07:36.255 --> 00:07:38.825
以前ChatGPT只知道你跟他聊過什麼

00:07:38.825 --> 00:07:40.560
你早上用Chrome查資料

00:07:40.560 --> 00:07:42.228
用Slack跟同事討論

00:07:42.228 --> 00:07:43.729
開Google Docs改企劃

00:07:43.729 --> 00:07:45.431
打開Excel看數字

00:07:45.431 --> 00:07:46.966
這些他通通不知道

00:07:46.966 --> 00:07:48.568
現在使用Computer History

00:07:48.568 --> 00:07:51.304
他會把這些活動整理成一條時間軸

00:07:51.304 --> 00:07:53.005
九點十分看企劃回饋

00:07:53.005 --> 00:07:54.574
九點半改上線文件

00:07:54.574 --> 00:07:56.409
十點十五看銷售數據

00:07:56.409 --> 00:07:58.010
十點四十回Slack

00:07:58.010 --> 00:07:59.645
所以你下午可以直接問他

00:07:59.645 --> 00:08:00.680
「我早上做到哪了？」

00:08:00.680 --> 00:08:02.281
「幫我整理一下今天做了什麼

00:08:02.281 --> 00:08:03.349
等一下要報告」

00:08:03.349 --> 00:08:05.618
甚至「我剛剛看的那份提案在哪」

00:08:05.618 --> 00:08:07.119
這裡一定要澄清一個誤會

00:08:07.119 --> 00:08:08.754
他不是一直在錄你的螢幕

00:08:08.754 --> 00:08:11.090
官方說明他記錄的是互動事件

00:08:11.090 --> 00:08:13.659
像是點擊、打字、快捷鍵、切換App

00:08:13.659 --> 00:08:16.863
加上macOS輔助使用功能提供的相關資訊

00:08:16.863 --> 00:08:19.131
不會用影像的方式記錄你的行為

00:08:19.131 --> 00:08:20.733
更不會監聽你在幹嘛

00:08:20.733 --> 00:08:23.135
就連無痕視窗的行為也不會被記錄

00:08:23.135 --> 00:08:25.104
這些事件暫存在你的Mac上

00:08:25.104 --> 00:08:26.472
最長48小時

00:08:26.472 --> 00:08:29.575
處理完產生的記憶是留在本機的Markdown檔

00:08:29.575 --> 00:08:31.444
你可以自由決定要不要刪掉

00:08:31.444 --> 00:08:32.745
另外OpenAI提醒

00:08:32.745 --> 00:08:33.913
開了這個功能

00:08:33.913 --> 00:08:35.982
prompt injection的風險會變高

00:08:35.982 --> 00:08:38.651
建議你跟訊息對話先暫停收集

00:08:38.651 --> 00:08:40.519
敏感的App也記得排除掉

00:08:41.754 --> 00:08:43.456
Chrome瀏覽器的Claude擴充功能

00:08:43.456 --> 00:08:45.524
現在整個換成Claude Cowork了

00:08:45.524 --> 00:08:47.226
一分鐘告訴你差在哪裡

00:08:47.226 --> 00:08:49.595
以前那個側邊欄叫Claude in Chrome

00:08:49.595 --> 00:08:51.230
它是一段獨立的對話

00:08:51.230 --> 00:08:52.298
關掉就沒了

00:08:52.298 --> 00:08:54.267
跟你桌面版、手機版的Claude

00:08:54.267 --> 00:08:55.201
是兩個世界

00:08:55.201 --> 00:08:58.104
現在側邊欄裡跑的是一整個Cowork session

00:08:58.104 --> 00:09:00.039
你的對話會存進你的帳號

00:09:00.039 --> 00:09:01.807
所以你在瀏覽器裡開的頭

00:09:01.807 --> 00:09:02.775
可以回到桌面版

00:09:02.775 --> 00:09:03.476
網頁版

00:09:03.476 --> 00:09:04.076
手機版

00:09:04.076 --> 00:09:04.810
接著做

00:09:04.810 --> 00:09:07.313
它在瀏覽器裡能做的事比你想的多

00:09:07.313 --> 00:09:09.282
它看得到你現在正在看的頁面

00:09:09.282 --> 00:09:10.416
可以幫你點連結

00:09:10.416 --> 00:09:11.017
打字

00:09:11.017 --> 00:09:11.617
換頁

00:09:11.617 --> 00:09:12.351
填表單

00:09:12.351 --> 00:09:13.252
更關鍵的是

00:09:13.252 --> 00:09:15.154
你在Cowork裡設定好的Skills

00:09:15.154 --> 00:09:16.122
Connectors

00:09:16.122 --> 00:09:17.023
Plugins

00:09:17.023 --> 00:09:19.525
在瀏覽器擴充功能裡也一樣能用

00:09:19.525 --> 00:09:21.160
等於你那一整套工作流

00:09:21.160 --> 00:09:22.628
直接搬進瀏覽器了

00:09:22.628 --> 00:09:26.432
像是我直接請他幫我把X上面的AI更新資訊

00:09:26.432 --> 00:09:28.200
彙整到我的Cowork裡面

00:09:28.200 --> 00:09:30.636
接著我就可以直接在Claude編輯

00:09:30.636 --> 00:09:31.804
是不是超方便呢

00:09:31.804 --> 00:09:33.539
但有三個限制要先講

00:09:33.539 --> 00:09:33.973
第一

00:09:33.973 --> 00:09:34.941
方案限制

00:09:34.941 --> 00:09:36.042
Max跟Team

00:09:36.042 --> 00:09:36.909
現在就有

00:09:36.909 --> 00:09:38.978
Pro是接下來幾周陸續開放

00:09:38.978 --> 00:09:40.279
Enterprise版呢

00:09:40.279 --> 00:09:41.314
預設是關閉的

00:09:41.314 --> 00:09:43.683
要管理員針對核可的網域開放

00:09:43.683 --> 00:09:44.216
第二

00:09:44.216 --> 00:09:45.318
瀏覽器限制

00:09:45.318 --> 00:09:48.821
目前只支援Chrome或Chromium核心的瀏覽器

00:09:48.821 --> 00:09:49.989
手機還不行喔

00:09:49.989 --> 00:09:50.590
第三

00:09:50.590 --> 00:09:51.457
資安顧慮

00:09:51.457 --> 00:09:52.758
這點我要特別提醒

00:09:52.758 --> 00:09:54.193
一個能幫你點按鈕

00:09:54.193 --> 00:09:56.195
填表單的AI在瀏覽器裡跑

00:09:56.195 --> 00:09:58.564
代表他看得到你登錄之後的畫面

00:09:58.564 --> 00:10:00.366
要不要讓他在你的銀行分頁

00:10:00.366 --> 00:10:02.034
公司後台旁邊開著

00:10:02.034 --> 00:10:02.835
這個

00:10:02.835 --> 00:10:04.036
真的要自己想清楚

00:10:04.036 --> 00:10:06.005
講到這裡你應該已經發現了

00:10:06.005 --> 00:10:06.706
現在Claude

00:10:06.706 --> 00:10:07.640
ChatGPT

00:10:07.640 --> 00:10:08.207
Gemini

00:10:08.207 --> 00:10:10.343
三家都把手伸進你的瀏覽器裡

00:10:10.343 --> 00:10:11.744
Claude是Chrome側邊欄

00:10:11.744 --> 00:10:13.112
跑一整個Cowork session

00:10:13.112 --> 00:10:15.881
ChatGPT那邊在Atlas瀏覽器收掉之後

00:10:15.881 --> 00:10:17.049
路線又變了一次

00:10:17.049 --> 00:10:19.151
Gemini則是直接長在Chrome裡面

00:10:19.151 --> 00:10:20.820
你根本不用裝任何東西

00:10:20.820 --> 00:10:22.388
三家的做法完全不一樣

00:10:22.388 --> 00:10:23.456
能碰到的資料

00:10:23.456 --> 00:10:24.457
要付的代價

00:10:24.457 --> 00:10:25.558
誰比較適合你

00:10:25.558 --> 00:10:26.392
差蠻多的

00:10:26.392 --> 00:10:27.627
這個題目要講清楚

00:10:27.627 --> 00:10:29.295
需要一整支影片的長度

00:10:29.295 --> 00:10:30.162
所以我把「Claude

00:10:30.162 --> 00:10:31.063
ChatGPT

00:10:31.063 --> 00:10:32.498
Gemini三家瀏覽器

00:10:32.498 --> 00:10:33.699
AI完整比較」

00:10:33.699 --> 00:10:34.934
做成了會員影片

00:10:34.934 --> 00:10:35.668
想看的話

00:10:35.668 --> 00:10:37.837
點下面加入會員就可以看到喔

00:10:37.837 --> 00:10:39.338
也是對我們最好的支持

00:10:40.740 --> 00:10:41.907
最後看Gemini

00:10:41.907 --> 00:10:43.976
等了三個月的旗艦3.5 Pro

00:10:43.976 --> 00:10:44.877
還是沒來

00:10:44.877 --> 00:10:46.012
但8月13號

00:10:46.012 --> 00:10:48.214
Google直接端出了3.7 Flash

00:10:48.214 --> 00:10:48.814
注意了

00:10:48.814 --> 00:10:51.150
3.6 Flash才剛出沒幾個禮拜

00:10:51.150 --> 00:10:53.052
而且這一版的定位非常明確

00:10:53.052 --> 00:10:55.221
它不是給你聊天用的通用模型

00:10:55.221 --> 00:10:57.156
而是專門給Coding Agent用的

00:10:57.156 --> 00:10:58.624
官方文件上直接寫

00:10:58.624 --> 00:10:59.625
Complex Coding

00:10:59.625 --> 00:11:00.993
Agentic Workflows

00:11:00.993 --> 00:11:03.195
Reliable Multi-step Execution

00:11:03.195 --> 00:11:05.898
還特別說改善了真實世界的軟體工程

00:11:05.898 --> 00:11:06.799
Issue解決

00:11:06.799 --> 00:11:08.968
以及Agent卡住鬼打牆的問題

00:11:08.968 --> 00:11:10.636
價格是這次的重頭戲

00:11:10.636 --> 00:11:12.938
輸入每百萬Token 0.75美金

00:11:12.938 --> 00:11:14.807
輸出3.75美金

00:11:14.807 --> 00:11:17.576
但這個價格只到今年12月31號

00:11:17.576 --> 00:11:19.612
明年1月1號會直接翻倍

00:11:19.612 --> 00:11:21.814
變成1.5跟7.5美金

00:11:21.814 --> 00:11:23.315
所以現在這段時間

00:11:23.315 --> 00:11:25.284
它的價格是真的很有殺傷力

00:11:25.284 --> 00:11:27.553
雖然不比其他廠的旗艦模型強

00:11:27.553 --> 00:11:29.989
但在速度上與品質上還是蠻好的

00:11:29.989 --> 00:11:32.024
不過有兩件事情要老實講

00:11:32.024 --> 00:11:32.358
第一

00:11:32.358 --> 00:11:34.193
3.7 Flash的思考Token

00:11:34.193 --> 00:11:36.128
比3.6多了大約四成

00:11:36.128 --> 00:11:38.064
所以跑得快不等於比較便宜

00:11:38.064 --> 00:11:39.031
實際算下來

00:11:39.031 --> 00:11:41.133
每個任務的成本可能沒有降那麼多

00:11:41.133 --> 00:11:41.734
第二

00:11:41.734 --> 00:11:42.635
評測顯示

00:11:42.635 --> 00:11:45.438
它的幻覺率從55.6%

00:11:45.438 --> 00:11:47.406
上升到64.5%

00:11:47.406 --> 00:11:50.009
這個等級不建議直接拿去面對客戶

00:11:50.009 --> 00:11:52.244
前面一定要加檢索跟防護

00:11:52.244 --> 00:11:52.678
好的

00:11:52.678 --> 00:11:54.613
以上就是8月AI重點更新

00:11:54.613 --> 00:11:55.648
歡迎留言告訴我

00:11:55.648 --> 00:11:56.615
這一輪更新

00:11:56.615 --> 00:11:57.817
你對哪一個最有感

00:11:57.817 --> 00:11:59.218
如果這集對你有幫助

00:11:59.218 --> 00:12:00.586
或是覺得乾貨滿滿

00:12:00.586 --> 00:12:01.987
歡迎利用超級感謝

00:12:01.987 --> 00:12:02.988
或是加入會員

00:12:02.988 --> 00:12:04.590
就是對我們最好的支持了

00:12:04.590 --> 00:12:05.658
我們下集見