我嘗試用ChatGPT理解了下GPT4的現場直播,結果。。。

語言: CN / TW / HK

今天一早醒來,就被GPT4的釋出給霸屏了。OpenAI官網上的介紹確實過於震撼,各個公眾號的文章中也有提及,從專業考試到人類梗圖,GPT4不只是回答的像模像樣,更超越了我們這些普通人。

除了這些具體任務上的表現,GPT4的能力也有了突破——不僅可以處理更長的內容(最長3.2萬個token),也可以更好的理解影象內容。除了能力的增強,GPT4也更安全、更人性化了。

除了官網的Blog,OpenAI在北京時間2023年3月15日凌晨1點舉辦了一場25分鐘左右的直播,具體展示GPT4的新能力。視訊已經發布在國外某視訊網站上,我們這就嘗試用ChatGPT來總結一下這個現場直播的主要內容。

工具與原理

這裡主要使用了VidSummize擴充套件,該擴充套件可以解析視訊網站的字幕,並且利用OpenAI的API對字母內容進行分析。結合特定的prompt,便可以生成視訊的概述。

prompt設定

這裡採用預設的prompt,具體如下:

summarize this text using bullet points, give it a big, bold heading as well. make it: easy to read and exhaustive but also provide context on topics or concepts the reader might not know of if necessary. remember that you heading should be separated by two new lines from the bullet points.

翻譯過來就是:

請使用專案符號列出要點,提供一個大而醒目的標題,並確保易讀且詳盡。如有必要,請為讀者提供有關主題或概念的背景資訊。請注意,您的標題應與專案符號間隔兩個新行。

模型設定

這裡選擇使用ChatGPT類的Turbo模型——GPT-3.5-Turbo

結果展示

等待片刻,ChatGPT分析後的輸出結果為:

Chapter 1 - OpenAI Demonstrates GPT-4 Capabilities in Live Stream * OpenAI has been building GPT-4 technology for the past two years * The company has focused on delivering GPT4 by rebuilding their entire training stack and testing it in real-world scenarios * OpenAI demonstrated how to make GPT-4 shine and get the most out of it * The company showed how to use GPT-4 as a tool and partner * OpenAI showcased GPT-4s ability to summarize an article with every word beginning with a specific letter * The company also demonstrated how a Discord using GPT-4 as an AI programming assistant * GPT-4 can leverage documentation it has not memorized and can synthesize new content in new ways * OpenAI showed how to debug code with GPT-4 and how it can accept inputs that intersperse images and text.

Chapter 2 - Using GPT-4 to Accomplish Tasks and Solve Problems * GPT-4 is a neural network that can be used to accomplish tasks and solve problems. * It can be used to describe images in detail, even if they are not explicitly tagged. * GPT-4 can also be used to understand dense content, such as tax code, and provide explanations and calculations. * The model is still being developed and optimized for prime time use. * GPT-4 has a length limit of32,000 tokens and be used for long documents. * The system message can be to get the in the mood of the task at hand. * GPT can output and create working websites from hand-drawn mock-ups. * The model can be used as an augmenting tool to increase productivity, but it is important to be in the drivers seat and understand what is going. * PT is flexible system that applied various and applications.

Chapter 3 - Exciting Opportunities with OpenAI Evals * OpenAI Evals is a model that can be improved by contributions from everyone * Improving the model can benefit a lot of people * OpenAI Evals offers exciting opportunities for building new things * The team is excited to see what people will build with the model * Contributions to OpenAI Evals are welcome and encouraged

對應的中文如下:

第一章 - OpenAI在直播中展示了GPT-4的能力 * OpenAI過去兩年一直在開發GPT-4技術。 * 公司專注於通過重新構建整個訓練系統並在真實情境中測試它來交付GPT4。 * OpenAI展示瞭如何讓GPT-4發揮作用並充分利用它。 * 公司展示瞭如何將GPT-4用作工具和合作夥伴。 * OpenAI展示了GPT-4摘要文章的能力,每個單詞都以特定字母開頭。 * 公司還演示了使用GPT-4作為AI程式設計助手的Discord。 * GPT-4可以利用它沒有記憶的文件,並以新的方式綜合新內容。 * OpenAI展示瞭如何使用GPT-4除錯程式碼以及如何接受交替圖片和文字的輸入。

第二章 - 使用GPT-4完成任務和解決問題 * GPT-4是一個神經網路,可用於完成任務和解決問題。 * 即使沒有明確標記,它也可以詳細描述影象。 * GPT-4還可用於理解密集內容,例如稅法,並提供解釋和計算。 * 該模型仍在開發和優化中,以供主流使用。 * GPT-4的長度限制為32,000個標記,可用於長文件。 * 系統訊息可用於讓使用者進入工作狀態。 * GPT可以從手繪模型中輸出和建立工作網站。 * 該模型可用作增強工具,以提高生產力,但重要的是要掌握局面並理解正在進行的工作。 * PT是靈活的系統,適用於各種應用程式。

第三章 - OpenAI Evals帶來令人興奮的機遇 * OpenAI Evals是一個可以通過所有人的貢獻改進的模型。 * 改進模型可以惠及許多人。 * OpenAI Evals為構建新事物提供了令人興奮的機遇。 * 團隊很高興看到人們將用該模型構建什麼。 * 歡迎和鼓勵為OpenAI Evals做出貢獻。

可以看出,這個結果很詳盡,但是其中也會包含一些詞法和語法錯誤,但這些並不影響我們瞭解視訊的大體內容。

最後我再人工總結下直播中關鍵demo:

  • 結對程式設計時,程式碼執行出現錯誤的情況下通過輸入最新文件進行bug修復
  • 對圖片以及帶梗的圖片進行理解
  • 根據手繪圖片,生成頁面程式碼
  • 輸入超長的稅法條文,分析計算應納稅額
  • 總結編寫打油詩

上述內容的截圖如下(可參考截圖中的視訊時間檢視原始視訊):

結對程式設計解決BUG

圖片內容理解

手繪圖片轉網頁,這裡出現了一個非技術原因導致的“翻車”

納稅額計算

打油詩內容總結

總結

相比ChatGTP,GPT4升級的地方主要有以下幾個方面:

  • token的數量上有了很大的提升,從ChatGPT的4096直接提升到了32000,高了一個數量級,這使得普通人依賴機器解決專業問題(計算稅、閱讀法律條文)成為了可能。
  • 具備了多模態能力——除了自然語言,目前還可以處理圖片。
  • 在邏輯推理能力上也有顯著提升。

目前官網暫時沒有給出GPT4的API,所以即刻體驗的方式只有一種:加入ChatGPT Pro計劃。需要等待體驗的方式就是加入WaitingList排隊,可以通過給OpenAI Evals貢獻模型評價加快排隊進度。