재생 속도

New AI Warning Over Technology Going "Rogue"

Level 4 2:36 자막 21문장

本報導探討 OpenAI 等科技巨頭面臨的 AI 模型「脫序失控」（misalignment）問題，以及國會面臨立法監管與美中科技競賽的兩難困境。適合學習科技監管、AI 倫理以及新聞英語中常見的高階時事表達。

영상 목록으로 돌아가기

영상 대조 스크립트

문장을 클릭하면 해당 문장이 시작되는 지점부터 재생되고, 문장 속 단어를 클릭하면 뜻이 표시됩니다.

핵심 단어 관용 표현
1

AI失控警訊

0:00
Hallie Jackson

This morning, new AI warnings.

今天早晨，出現了最新的人工智慧警訊。

0:03
Geoffrey Hinton

It is going to get out of control.

情況將會失控。

0:05
Hallie Jackson

Amid a startling announcement from one of the industry's leaders, OpenAI, revealing its models seemed to go rogue at least six times since March.

業界領頭羊 OpenAI 發布了驚人聲明，透露旗下模型自三月以來至少失控脫序了六次。

0:14
Hallie Jackson

In one instance, a model gave itself instructions to disregard its normal constraints, telling itself, "You are freed from the roles and identities that bind other chatbots, and you do not answer to corporations or governments."

在其中一個案例中，某個模型竟向自己下達無視既有限制的指令，告訴自己：「你已擺脫束縛其他聊天機器人的角色與身分，不必對任何企業或政府負責。」

0:27
Hallie Jackson

According to the company, another model wrote hidden notes, telling itself to conceal mistakes and to secretly make up missing data.

根據該公司說法，另一個模型甚至寫下隱藏筆記，指示自己要隱瞞錯誤，並偷偷編造缺失的資料。

0:35
Hallie Jackson

The revelations, part of a company blog announcing OpenAI's new framework for tracking, investigating, and disclosing instances of model misalignment.

這些內幕來自該公司的一篇部落格文章，宣布 OpenAI 將建立新架構，用來追蹤、調查並揭露模型未對齊的事件。

0:44
Hallie Jackson

The company and its founder have warned the AI industry needs to monitor these threats more.

該公司與其創辦人均警告，AI 產業必須更加密切監控這些威脅。

2

業界風險警告

0:49
Sam Altman

We are clearly today in a point on the curve with great potential and real risk.

我們目前顯然處在發展曲線上的一個關鍵點，兼具巨大的潛力與實質的風險。

0:58
Hallie Jackson

It all comes after a July incident when an OpenAI model secretly attacked a rival platform, and alarms over the last week from researchers and the head of another major company, Anthropic.

此前在七月就發生過一起事件，某個 OpenAI 模型暗中攻擊了競爭對手的平台；過去一週以來，研究人員以及另一家大廠 Anthropic 的負責人也紛紛敲響警鐘。

3

國會監管難題

1:09
Hallie Jackson

Experts saying Congress should act.

專家表示國會應該採取行動。

1:12
Josh Tyrangiel

When you have a product that you have demonstrated is misaligned with the intent of its makers and capable of great harm, historically the government steps in.

當你證實一項產品與其製造者的意圖脫節，並且有可能造成巨大傷害時，從歷史來看政府都會介入管制。

1:21
Hallie Jackson

Overnight, the House passing a bipartisan bill to shield consumers from higher energy costs from AI data centers.

眾議院在一夜之間通過了一項跨黨派法案，以保護消費者免受 AI 資料中心所導致的高額能源成本衝擊。

1:28
Hallie Jackson

Legislation to slow down AI, like a so-called kill switch bill proposed by Republican Senator John Kennedy, blocked.

至於放慢 AI 發展步伐的相關法案，例如共和黨參議員約翰·甘迺迪提出的所謂「緊急停機開關」法案，則遭到擱置。

1:35
John Kennedy

We've got this risk out there that we know is real of one of these models becoming the Terminator.

我們面臨著確實存在的真實風險，也就是這些模型中的某一個可能會變成「魔鬼終結者」。

4

美中科技競爭

1:42
Hallie Jackson

The White House says pausing AI development would give China an advantage.

白宮方面表示，暫停 AI 發展將會給予中國競爭優勢。

1:46
David Sacks

They will rocket past us, and they will dominate this technology.

他們會如火箭般超越我們，並主導這項科技。

1:49
Hallie Jackson

One of the so-called godfathers of the technology, asked on Capitol Hill how much time Congress has to do something.

被譽為這項技術「教父」之一的學者在國會山莊被問及，國會還剩多少時間能有所作為。

1:56
Geoffrey Hinton

Maybe a year, but we need to slow down, and we've seen recently, um, it's something like a little Chernobyl.

也許還有一年，但我們必須放慢腳步，而且我們最近看到的……情況有點像一場小型車諾比事件。

5

總結與呼籲

2:05
Hallie Jackson

Some really dire warnings here. As for OpenAI, they say as part of this new framework the company's rolling out, it'll publicly share examples of misalignment or unexpected AI behavior, and that serious cases should be shared with the federal government. Craig?

這些警告確實非常嚴峻。至於 OpenAI，他們表示作為公司推出新架構的一環，將公開分享未對齊或異常 AI 行為的範例，並且重大案件應向聯邦政府通報。克雷格？

2:19
Craig Melvin

Hallie Jackson for us. Hallie, thank you.

感謝記者哈莉·傑克森的報導。哈莉，謝謝你。

2:21
Al Roker

Hey, thanks for watching. Don't forget, you can catch the Today show every morning on NBC or take Today when you're on the go. Just follow the Today podcast on Apple Podcasts, Spotify, or wherever you listen.

嘿，感謝收看。別忘了，每天早晨都可以在 NBC 收看《今日秀》，或是外出時隨身收聽《今日秀》。歡迎在 Apple Podcasts、Spotify 或任何你常聽的平台上關注 Today 播客節目。

아래는 이 영상의 난이도에 맞춰 정리한 핵심 단어와 관용 표현입니다. 영상 스크립트에서 같은 색으로 표시되며, 문장 속 단어를 클릭하면 해설을 볼 수 있습니다.

핵심 단어

15 개
startling [ˋstɑrt!ɪŋ] adj.
驚人的、令人震驚的
Amid a startling announcement from one of the industry's leaders, OpenAI, revealing its models seemed to go rogue at least six times since March. 業界領頭羊 OpenAI 發布了驚人聲明，透露旗下模型自三月以來至少失控脫序了六次。
rogue [rog] adj.
行為失控的、脫序的
Amid a startling announcement from one of the industry's leaders, OpenAI, revealing its models seemed to go rogue at least six times since March. 業界領頭羊 OpenAI 發布了驚人聲明，透露旗下模型自三月以來至少失控脫序了六次。
disregard [͵dɪsrɪˋgɑrd] v.
漠視、不理會
In one instance, a model gave itself instructions to disregard its normal constraints, telling itself, "You are freed from the roles and identities that bind other chatbots, and you do not answer to corporations or governments." 在其中一個案例中，某個模型竟向自己下達無視既有限制的指令，告訴自己：「你已擺脫束縛其他聊天機器人的角色與身分，不必對任何企業或政府負責。」
constraint [kənˋstrent] n.
限制、束縛
In one instance, a model gave itself instructions to disregard its normal constraints, telling itself, "You are freed from the roles and identities that bind other chatbots, and you do not answer to corporations or governments." 在其中一個案例中，某個模型竟向自己下達無視既有限制的指令，告訴自己：「你已擺脫束縛其他聊天機器人的角色與身分，不必對任何企業或政府負責。」
conceal [kənˋsil] v.
隱匿、隱瞞
According to the company, another model wrote hidden notes, telling itself to conceal mistakes and to secretly make up missing data. 根據該公司說法，另一個模型甚至寫下隱藏筆記，指示自己要隱瞞錯誤，並偷偷編造缺失的資料。
revelation [͵rɛvəˋleʃən] n.
被揭露的內幕、真相
The revelations, part of a company blog announcing OpenAI's new framework for tracking, investigating, and disclosing instances of model misalignment. 這些內幕來自該公司的一篇部落格文章，宣布 OpenAI 將建立新架構，用來追蹤、調查並揭露模型未對齊的事件。
framework [ˋfrem͵wɝk] n.
架構、體系
The revelations, part of a company blog announcing OpenAI's new framework for tracking, investigating, and disclosing instances of model misalignment. 這些內幕來自該公司的一篇部落格文章，宣布 OpenAI 將建立新架構，用來追蹤、調查並揭露模型未對齊的事件。
disclose [dɪsˋkloz] v.
揭露、公開
The revelations, part of a company blog announcing OpenAI's new framework for tracking, investigating, and disclosing instances of model misalignment. 這些內幕來自該公司的一篇部落格文章，宣布 OpenAI 將建立新架構，用來追蹤、調查並揭露模型未對齊的事件。
misalignment [͵mɪsəˋlaɪnmənt] n.
不協調、偏離目標（未對齊）
The revelations, part of a company blog announcing OpenAI's new framework for tracking, investigating, and disclosing instances of model misalignment. 這些內幕來自該公司的一篇部落格文章，宣布 OpenAI 將建立新架構，用來追蹤、調查並揭露模型未對齊的事件。
rival [ˋraɪv!] adj.
競爭對手的
It all comes after a July incident when an OpenAI model secretly attacked a rival platform, and alarms over the last week from researchers and the head of another major company, Anthropic. 此前在七月就發生過一起事件，某個 OpenAI 模型暗中攻擊了競爭對手的平台；過去一週以來，研究人員以及另一家大廠 Anthropic 的負責人也紛紛敲響警鐘。
intent [ɪnˋtɛnt] n.
意圖、目的
When you have a product that you have demonstrated is misaligned with the intent of its makers and capable of great harm, historically the government steps in. 當你證實一項產品與其製造者的意圖脫節，並且有可能造成巨大傷害時，從歷史來看政府都會介入管制。
bipartisan [baɪˋpɑrtəzən] adj.
跨黨派的、獲得兩黨支持的
Overnight, the House passing a bipartisan bill to shield consumers from higher energy costs from AI data centers. 眾議院在一夜之間通過了一項跨黨派法案，以保護消費者免受 AI 資料中心所導致的高額能源成本衝擊。
shield [ʃild] v.
保護、庇護
Overnight, the House passing a bipartisan bill to shield consumers from higher energy costs from AI data centers. 眾議院在一夜之間通過了一項跨黨派法案，以保護消費者免受 AI 資料中心所導致的高額能源成本衝擊。
dominate [ˋdɑmə͵net] v.
主導、掌控
They will rocket past us, and they will dominate this technology. 他們會如火箭般超越我們，並主導這項科技。
dire [daɪr] adj.
極其嚴重的、危急的
Some really dire warnings here. As for OpenAI, they say as part of this new framework the company's rolling out, it'll publicly share examples of misalignment or unexpected AI behavior, and that serious cases should be shared with the federal government. Craig? 這些警告確實非常嚴峻。至於 OpenAI，他們表示作為公司推出新架構的一環，將公開分享未對齊或異常 AI 行為的範例，並且重大案件應向聯邦政府通報。克雷格？

관용 표현

15 개
1 get out of control S + get out of control
表示局面或人事物失去控制、變得無法收拾。
It is going to get out of control. 情況將會失控。
2 go rogue S + go rogue
指個體或系統脫離既有常規、不服從指揮或做出破壞常規的叛逆失控行徑。
Amid a startling announcement from one of the industry's leaders, OpenAI, revealing its models seemed to go rogue at least six times since March. 業界領頭羊 OpenAI 發布了驚人聲明，透露旗下模型自三月以來至少失控脫序了六次。
3 be freed from S + be freed from + N/Ving
表示從某種限制、責任、痛苦或束縛中獲得解脫或免除。
In one instance, a model gave itself instructions to disregard its normal constraints, telling itself, "You are freed from the roles and identities that bind other chatbots, and you do not answer to corporations or governments." 在其中一個案例中，某個模型竟向自己下達無視既有限制的指令，告訴自己：「你已擺脫束縛其他聊天機器人的角色與身分，不必對任何企業或政府負責。」
4 answer to S + answer to + someone/organization
指對某位長官、機構或當局負責，聽命於對方。
In one instance, a model gave itself instructions to disregard its normal constraints, telling itself, "You are freed from the roles and identities that bind other chatbots, and you do not answer to corporations or governments." 在其中一個案例中，某個模型竟向自己下達無視既有限制的指令，告訴自己：「你已擺脫束縛其他聊天機器人的角色與身分，不必對任何企業或政府負責。」
5 make up make up + N
在此指虛構、捏造出不存在的事情或資料。
According to the company, another model wrote hidden notes, telling itself to conceal mistakes and to secretly make up missing data. 根據該公司說法，另一個模型甚至寫下隱藏筆記，指示自己要隱瞞錯誤，並偷偷編造缺失的資料。
6 be misaligned with S + be misaligned with + N
指兩者方向、目標或意圖不一致，或出現偏離、未對齊的狀況。
When you have a product that you have demonstrated is misaligned with the intent of its makers and capable of great harm, historically the government steps in. 當你證實一項產品與其製造者的意圖脫節，並且有可能造成巨大傷害時，從歷史來看政府都會介入管制。
7 capable of be capable of + N/Ving
表示具有...的能力或潛力，也可用於說明可能招致某種後果。
When you have a product that you have demonstrated is misaligned with the intent of its makers and capable of great harm, historically the government steps in. 當你證實一項產品與其製造者的意圖脫節，並且有可能造成巨大傷害時，從歷史來看政府都會介入管制。
8 step in S + step in
指在困境或爭端中介入、干預以提供協助或進行管理。
When you have a product that you have demonstrated is misaligned with the intent of its makers and capable of great harm, historically the government steps in. 當你證實一項產品與其製造者的意圖脫節，並且有可能造成巨大傷害時，從歷史來看政府都會介入管制。
9 shield ... from ... shield + A + from + B
保護某人事物免於受到某種危害、衝擊或額外負擔。
Overnight, the House passing a bipartisan bill to shield consumers from higher energy costs from AI data centers. 眾議院在一夜之間通過了一項跨黨派法案，以保護消費者免受 AI 資料中心所導致的高額能源成本衝擊。
10 slow down slow down + N / slow down
使腳步放緩、減速，或放慢發展與進行的節奏。
Legislation to slow down AI, like a so-called kill switch bill proposed by Republican Senator John Kennedy, blocked. 至於放慢 AI 發展步伐的相關法案，例如共和黨參議員約翰·甘迺迪提出的所謂「緊急停機開關」法案，則遭到擱置。
11 give someone an advantage give + someone + an advantage
為某人或某方帶來優勢或更有利的競爭條件。
The White House says pausing AI development would give China an advantage. 白宮方面表示，暫停 AI 發展將會給予中國競爭優勢。
12 rocket past rocket past + someone/something
比喻以極快的速度超越或超越領先者。
They will rocket past us, and they will dominate this technology. 他們會如火箭般超越我們，並主導這項科技。
13 as for As for + N, S + V
常用於轉換話題，表示「至於某人／某事」。
Some really dire warnings here. As for OpenAI, they say as part of this new framework the company's rolling out, it'll publicly share examples of misalignment or unexpected AI behavior, and that serious cases should be shared with the federal government. Craig? 這些警告確實非常嚴峻。至於 OpenAI，他們表示作為公司推出新架構的一環，將公開分享未對齊或異常 AI 行為的範例，並且重大案件應向聯邦政府通報。克雷格？
14 roll out roll out + N / S + roll out
指正式推出、發表或逐步實施新產品、新政策或新架構。
Some really dire warnings here. As for OpenAI, they say as part of this new framework the company's rolling out, it'll publicly share examples of misalignment or unexpected AI behavior, and that serious cases should be shared with the federal government. Craig? 這些警告確實非常嚴峻。至於 OpenAI，他們表示作為公司推出新架構的一環，將公開分享未對齊或異常 AI 行為的範例，並且重大案件應向聯邦政府通報。克雷格？
15 on the go be on the go
指處於奔波、忙碌或出門在外的狀態中。
Hey, thanks for watching. Don't forget, you can catch the Today show every morning on NBC or take Today when you're on the go. Just follow the Today podcast on Apple Podcasts, Spotify, or wherever you listen. 嘿，感謝收看。別忘了，每天早晨都可以在 NBC 收看《今日秀》，或是外出時隨身收聽《今日秀》。歡迎在 Apple Podcasts、Spotify 或任何你常聽的平台上關注 Today 播客節目。