工作清單

I’m using Jev more than Opus 5.5 or GPT-6. Here’s why.

逐字稿文字預覽

Jev, Jev, Jev. Welcome to Jev week onhow I AI. We have seen a lot of newmodels be released in the last 5 days.We saw Opus 55. We saw GPT6 Soul, GPT6
Luna. Muse is blowing up the timeline.Everybody still loves their Grock bots.And yet there is one thing that I want
o talk about in AI right now and thatis this fast cheap doesn't speak systemone decision model from Typesafe AI. As
oon as I saw this trending on X, asoon as I saw it launched, I immediatelystarted testing it. Now, what I will say
is more than any other model I'vexperienced lately, Jev has been the onethat has explodeduse cases, personal productivity use
cases, code use cases, product usecases. Today I'm going to give you avery quick whirlwind tour of what Jebis, what it will give you back and what
it won't, and then how I have used itover the last week to do work that Ithink is worth hundreds of thousands if
not millions of dollars. And I haveprobably spent sub $10 on Jev tokens. Alot of it has been subsidized because
Jeb is currently free on the AI gatewayby Versel. But even at list price, it isa very inexpensive model. It is a very
effective model and it is going to be inthe middle of almost everything I buildfrom here on out. [music] So, let's get
o it. This episode is brought to you byOpen Art Arena, the global leaderboardfor creative intelligence. Every week,
new AI models launch, and everyoneclaims to be the best. But best at what?
Open Art Arena is built to answer thequestion that actually matters. Whichmodel is [music] best for your specific
job? Instead of one overall winner, Open
Art Arena ranks models across realcreative tasks from advertising and filmto animation, product, graphic design,
editing, and lip-sync [music]covering both image and video. And theserankings aren't based on hype. They're
judged by professionals, industryleaders, and working creators throughblind evaluations. So judges never know
hich model produced which output. Thatmeans you can see how models actuallyperform when it comes to the creative
work you're doing. So stop guessingwhich model to use. Explore rankingsbased on real creative work and find the
right model for your project and savetime and cost. See the rankings at OpenArt Arena. So if I were to explain Jev
to you, I would go to this table on thetype blog post announcing Jev. And itbasically compares normal LLMs on the
left, Jev style LLMs on the right. Bothtake in unstructured data as inputs.Both take in text as inputs. Jev does
not take in images, but it takes in textand it takes in text descriptions ofimages if you really need to get there.
The outputs though are very different.With the standard LLMs that you're usedto working with, you are getting strings
and generated text out. So you'regetting text in, text out. With Jev,you're getting text in, type safe values
out. And what I mean by type safe valuesis these are values that are predefinedthat then Jev picks from and selects and
returns to you. We will show you whathose values are, but essentiallythey're like it's this or that. It's yes
or no. It's 1 through 10. Like it'spretty simple. Now it sounds simple, butit is incredibly powerful when you put
it against the right problem.The other thing that you will noticeabout Jeff is it is cheap AF and it isfast AF. And so if you look at the
current cost of input and output tokens,it can be pennies to tens or hundreds ofdollars per million output tokens.
Jev only charges you on input tokensbecause it barely outputs anything. Andit is 4 cents per million input tokens.
It is like dirt freaking cheap. Andbecause they output basically nothing,they don't even charge you for output
okens. Um whereas the standard LM aregoing to charge you tons for the outputokens. And then I'll go into use cases,
which is how you would use like astandard LLM versus why you would useJev. You know, standard LMS, chat botswhere you want text in, you want text
out. Coding where you want, you know,code to be produced. Um, and so it'sgreat for things where you needstuff generated, but Jev is really great
for making decisions. That's ultimatelyI've been calling this it's like adecision model. So you give it adecision and it makes a decision and and
it's high agency I would say. So if youneed smart if statements, should I goleft or right? If it's good, put it to
this person. If it's bad, send it tosupport. All that kind of stuff Jeff isreally good at. It's really good at
classifying data. That's a lot of whatI'm going to talk to you about today.And it's really good at real time. So,
it is fast, fast, fast. So, you can putan LLM in a real-time loop in a way thatwas not performant enough with these
other models. I do want to explainexactly what Jev outputs. And so, Iwould highly recommend if you're going
to use Jev, go to their docs. The twoplaces that I'm using the most are theprimitives, so what it returns. And then
I use some of the patterns in cookbooks.So what will Jev return to you? It will
essentially return one of three things.It will return a choice which means youcan give it text, you can give it a list
of choices and it will pick a choice.So, if my question is, what should Iwear on this date and my choice is adress, jeans, workout clothes, ski gear,
or nothing,it'll like maybe pick the dress, right?
And so that is choice. If I selectscore, it's going to say like how severeis this or on a score from one to five,
what it is. This is really good forseverity rankings. So say the examplehere is a great one which is like if you
get a bug, you want to rank it likecosmetic, broken or blocking, Jev willtriage that bug and give it a score. And
then finally, there's a new. It's aversion of a boolean. It basically tellsyou what the likelihood that the answer
to a question is yes. So if the answerto a question is yes or no, is Claire apodcaster? it's going to give you like a
99% nule because there's a 99% chancethat I am a podcaster. Okay, so this ishow you know you would think about the
three things it can return. And agai

來源：transcript.txt

本次執行流程與問題處理

  1. 下達指令

    指令：videonote https://youtu.be/-KIBgpGA_XI?si=pWvYszUvtgySliyE --language en --bilingual --learning --auto-source-subtitles

    來源：https://youtu.be/-KIBgpGA_XI?si=pWvYszUvtgySliyE

    辨識語言：en ｜ 參數：--language en --bilingual --learning --auto-source-subtitles

  2. 預期流程

    依指令產生的處理順序：

    1. 自動字幕優先，Whisper 作為備援
    2. 來源選擇與品質檢查
    3. 文字校正與句子單元
    4. 雙語翻譯
    5. 雙語時間軸與學習單元
    6. 筆記與發布
    7. 公開網頁發布

  3. 1. 下載與語音辨識

    工具：videonote-transcribe-url（主辨識管線：下載、Whisper 辨識與選稿）

    處理：下載音訊、以 Whisper 辨識並選出最佳逐字稿

    產出：yt_dlp_meta.json（影片元資料）、audio_16k_mono.wav（16kHz 單聲道音訊）、transcript.whisper.json（Whisper 辨識結果）、transcript.json（正式逐字稿）、transcript_selection_report.json（來源選擇報告）

    耗時：15.3 秒

    關係：依序執行

    結果：完成

  4. 2. 全文快速校讀

    工具：videonote_rapid_transcript_review.py

    處理：先速讀完整逐字稿，再修正高可信突兀詞句並列出待確認項目

    產出：rapid_transcript_review_report.json（產出檔案）

    耗時：47.7 秒

    關係：依序執行

    結果：完成

  5. 3. 建立句子單元

    工具：videonote-build-sentence-units

    處理：將逐字稿切成帶時間軸的句子單元

    產出：sentence_units.json（句子單元）

    耗時：0.8 秒

    關係：依序執行

    結果：完成

  6. 4. 時間軸稽核

    工具：videonote_timeline_audit.py

    處理：檢查時間軸與音訊對齊

    產出：timeline_audit_report.json（時間軸稽核）

    耗時：0.8 秒

    關係：依序執行

    結果：完成

  7. 5. 品質檢查

    工具：videonote_controlled_quality.py（檢查時間軸與品質）

    處理：檢查語言、時間軸與可讀性品質

    產出：controlled_quality_report.json（品質報告）、language_sanity_report.json（語言檢查）

    耗時：1.0 秒

    關係：依序執行

    結果：完成

  8. 6. 播放時間軸回填

    工具：videonote_rebase_display_timeline.py

    處理：將網頁播放時間軸回填為校正字幕時間

    產出：display_timeline_rebase_report.json（時間軸回填報告）

    耗時：1.0 秒

    關係：依序執行

    結果：完成

  9. 7. 雙語翻譯

    工具：vnt-sentence（句子雙語翻譯）

    處理：將句子翻譯成繁體中文

    產出：sentence_bilingual_transcript.json（雙語逐字稿）

    耗時：197.0 秒

    關係：依序執行

    結果：完成

  10. 8. 雙語時間軸與學習單元

    工具：videonote_learning_timeline_v2.py（建立雙語時間軸與學習單元）

    處理：建立雙語時間軸與學習內容

    產出：sentence_bilingual_timeline.html（雙語時間軸頁）、learning.json（學習單元）、japanese_learning.json（日文學習單元）

    耗時：127.8 秒

    關係：依序執行

    結果：完成

  11. 9. 生成筆記摘要

    工具：videonote-generate-note（生成筆記摘要）

    處理：生成筆記摘要與重點

    產出：note.md（筆記 Markdown）、note.html（筆記 HTML）

    耗時：165.1 秒

    關係：依序執行

    結果：完成

  12. 10. 建立筆記頁面

    工具：videonote-render-html（產生筆記 HTML）

    處理：產生可閱讀的筆記 HTML 頁面

    產出：note.html（筆記 HTML）

    耗時：0.3 秒

    關係：依序執行

    結果：完成

  13. 11. 發布公開頁面

    工具：videonote-publish（發布公開頁面）

    處理：發布到公開網站與每日列表

    產出：公開筆記網址、result.json（管線結果）

    耗時：0.4 秒

    關係：依序執行

    結果：完成

手動來源字幕

自動來源字幕

Whisper 字幕

Qwen 字幕

Parakeet 字幕

MOSS 字幕

聲學時間證據（PyAnnote / WhisperX）

校正前後逐字稿

目前待解決問題與初步改善方向

進一步精進字幕與筆記
使用 Workbench 上傳校正字幕、中文參考稿或回饋 SRT，重新產生雙語與學習單元。
開啟 Workbench