區塊內幕 Cryptoinsiders 消息- GPT-4 成功完成了隊列中前 10% 的許多任務，而 ChatGPT 的原始版本通常在後 10% 中完成。
GPT-4 completed many of the tests within the top 10% of the cohort, while the original version of ChatGPT often finished up in the bottom 10%.
GPT-4 是人工智能聊天機器人 ChatGPT 的最新版本，能夠在高中和大學水平的考試中取得高分，達到 90% 的分數。 此外，AI 的處理能力有了顯著提高，這在以前的版本中是不可能的。
GPT-4, the latest version of the Artificial Intelligence (AI) chatbot, ChatGPT, can pass high school tests and law school exams with scores ranking in the 90th percentile and has new processing capabilities that were not possible with the prior version.
GPT-4 的測試成績數據於 3 月 14 日由其創建者 OpenAI 發布，這表明它還可以將圖像、音頻和視頻轉換為文本，並且具有更強的能力來生成更有創意和更有效的指令翻譯。
The figures from GPT-4’s test scores were shared on March 14 by its creator OpenAI revealing it can also convert image, audio and video inputs to text in addition to handling “much more nuanced instructions” more creatively and reliably.
OpenAI 表示，它“通過了一場模擬律師考試，得分在應試者的前 10% 左右”。 相反，GPT-3.5 的得分在倒數 10% 左右。
“It passes a simulated bar exam with a score around the top 10% of test takers,” OpenAI added. “In contrast, GPT-3.5’s score was around the bottom 10%.”
The figures show that GPT-4 achieved a score of 163 in the 88th percentile on the LSAT exam — the test college students need to pass in the United States to be admitted into law school.
如果得到證實，GPT4 的分數將躋身法學院前 20%，與哈佛、斯坦福、普林斯頓或耶魯等名校錄取所需的報告分數僅差幾分。
GPT4’s score would put it in a good position to be admitted into a top 20 law school and is only a few marks short of the reported scores needed for acceptance to prestigious schools such as Harvard, Stanford, Princeton or Yale.
之前版本的 ChatGPT 在 LSAT 上的得分為 149，位於倒數 40%。
The prior version of ChatGPT only scored 149 on the LSAT’s putting it in the bottom 40%.
GPT-4 在統一律師考試中也獲得了 298 分（滿分 400 分）——該考試由最近畢業的法學院學生參加，允許他們在美國任何州從事法律工作。
GPT-4 also scored 298 out of 400 in the Uniform Bar Exam — a test undertaken by recently graduated law students permitting them to practice as a lawyer in any U.S. jurisdiction.
以前版本的 ChatGPT 在這項任務上遇到了麻煩，在滿分 400 分中以 213 分排名倒數 10%。
The old version of ChatGPT struggled in this test, finishing in the bottom 10% with a score of 213 out of 400.
關於為評估美國高中生的大學準備情況而進行的 SAT 閱讀和寫作循證測試以及數學測試，GPT-4 得分分別位於第 93 和第 89 個百分位。
As for the SAT Evidence-Based Reading & Writing and SAT Math exams taken by U.S. high school students to measure their college readiness, GPT-4 scored in the 93rd and 89th percentile respectively.
GPT-4 在“硬”科學方面也表現出色，在 AP 生物學 (85-100%)、化學 (71-88%) 和物理 2 (66-84%) 中取得了高於平均水平的百分位分數。
GPT-4 excelled in the “hard” sciences too, posting well above average percentile scores in AP Biology (85-100%), Chemistry (71-88%) and Physics 2 (66-84%).
然而，它的 AP 微積分分數處於中等水平，排名在 43 到 59% 之間。
However its AP Calculus score was fairly average, ranking in the 43r to 59th percentile.
GPT-4 的另一個不足之處是英語語言考試，他們在兩次考試中的分數從第 8 到第 44 個百分位不等。
Another area where GPT-4 lacked was in English Literature exams, posting scores in the 8th to 44th percentile across two separate tests.
OpenAI 聲稱 GPT-4 和 GPT-3.5 從 2022-2023 實踐考試中進行了這些測試，他們說語言處理工具沒有接受算法的特定訓練：
OpenAI said GPT-4 and GPT-3.5 took these tests from the 2022-2023 practice exams, and that “no specific training” was taken by the language processing tools:
“We did no specific training for these exams. A minority of the problems in the exams were seen by the model during training, but we believe the results to be representative.”
這些結論也引起了 Twitter 社區的關注。
The results prompted fear in the Twitter community too.
FactoryDAO 的創建者 Nick Almond 在 3 月 14 日告訴他的 14,300 名推特粉絲，GPT4 將“嚇壞人們”並將“崩潰”全球教育體系。
Nick Almond, the founder of FactoryDAO told his 14,300 Twitter followers on March 14 that GPT4 is going to “scare people” and it will “collapse” the global education system.
They’ve basically trained it to pass all the tests that we use in the education system, and it does, ace them.— drnick 🗳️² (@DrNickA) March 14, 2023
So… we either need a new way of measuring intelligence, or 90% of us are dumber than the machines. https://t.co/1UCRS7ANSh
Coinbase 領導團隊前成員 Conor Grogan 表示，他在 GPT-4 模型中插入了一個實時的以太坊智能合約，並立即指出了幾個“安全問題”，並描述瞭如何利用代碼：
Former Coinbase director, Conor Grogan, said he inserted a live Ethereum smart contract into GPT-4 and instantly pointed to several “security vulnerabilities” and outlined how the code can be exploited:
I dumped a live Ethereum contract into GPT-4.— Conor (@jconorgrogan) March 14, 2023
In an instant, it highlighted a number of security vulnerabilities and pointed out surface areas where the contract could be exploited. It then verified a specific way I could exploit the contract pic.twitter.com/its5puakUW
早期對 ChatGPT 的智能合約審計觀察到，它的第一次迭代也能夠在很大程度上標記出代碼缺陷。
Earlier smart contract audits on ChatGPT found that its first version was also capable at spotting out code bugs to a reasonable degree too.
AI 時事通訊“The Rundown”的創作者 Rowan Cheung 發布了一段 GPT 將手繪假網站轉錄到一張紙上的視頻。
Rowan Cheung, the founder of AI newsletter “The Rundown” shared a video of GPT transcribing a hand drawn fake website on a piece of paper into code.
I just watched GPT-4 turn a hand-drawn sketch into a functional website.— Rowan Cheung (@rowancheung) March 14, 2023
This is insane. pic.twitter.com/P5nSjrk7Wn