實際影片長度:14:49.000。原文、繁中、雙語可點擊句子跳轉影片。
0:00.000–0:04.080
Why are we trying to build AI systemsthat are duplicative of humans? They
0:04.080–0:09.400
said that we should be building AGI, sowe say that we should be building AGI,but I would like to ask like why are we
0:09.400–0:12.240
doing that? Why is it that we are
0:12.240–0:19.760
building a technology that is ultimatelydesigned to replace and automate peopleaway? We should be building technology
0:19.760–0:28.320
and the purpose of technology throughouthistory has been to improve humanflourishing, not to replace people. Is
0:28.320–0:30.280
this the right goal? What like why are
0:30.280–0:33.160
we doing this? Why can't we just build
0:33.160–0:39.520
AI systems, which are systems that havenothing to do with the statisticalengines that they're trying to build to
0:39.520–0:41.360
duplicate the human brain?
0:41.360–0:48.640
>> That might be the most uncomfortablequestion I've heard about AI this year.Not can AI become smarter than humans,
0:48.640–0:56.440
not will AI destroy humanity, butsomething much simpler. Why are wetrying to build machines that replacepeople instead of helping people? Think
0:56.440–1:02.640
about it for a second. When cars wereinvented, they helped us travel faster.When smartphones became common, they
1:02.640–1:08.800
connected us to the whole world.Technology has always been about givinghumans more power, but for the first
1:08.800–1:15.040
ime, millions of people are asking acompletely different question. Whathappens when technology starts replacing
1:15.040–1:19.440
the person instead of just helping theperson? And that's exactly why I wanted
1:19.440–1:26.560
to react to this podcast. A 2-hourconversation between Steven Bartlett andjournalist Karen Hao, a woman who spent
1:26.560–1:35.320
8 years and over 300 interviewsinvestigating what's really happeninginside companies like OpenAI.And I need to be honest with you. Some
1:35.320–1:41.880
of what she says made me angry. Some ofit made me question things I use everyday, and one part about a small town in
1:41.880–1:50.480
Tennessee genuinely disturbed me.Let's walk through it and give my honestopinion about it.I want to give you some background first
1:50.480–1:59.080
because you can't understand the fightwithout understanding the field.>> So, this was back in 1956and [music] there were a group of
1:59.080–2:05.040
scientists that gathered at DartmouthUniversity to start a new discipline, ascientific discipline, to try and chase
2:05.040–2:11.400
an ambition. And specifically, anassistant professor at DartmouthUniversity, John McCarthy, decided to
2:11.400–2:18.320
name this discipline artificialintelligence.This was not the first name that hetried. The previous year, he tried to
2:18.320–2:25.040
name it Automata Studies. And the reasonwhy some of his colleagues wereconcerned about this name was because it
2:25.040–2:31.400
pegged the idea of this discipline torecreating human intelligence.>> His colleagues were deeply concerned
2:31.400–2:38.520
about the term artificial intelligencebecause it tied the field to the goal ofrecreating human intelligence. Both then
2:38.520–2:46.000
and now, there is absolutely noscientific consensus on what humanintelligence actually is. There is nosingle definition from biology,
2:46.000–2:54.040
psychology, or neurology, which means,as Karen points out, when today's AIcompanies talk about building AGI,
2:54.040–3:00.200
artificial general intelligence, they'rechasing a finish line that literallydoesn't have fixed coordinates.
3:00.200–3:07.280
>> Well,there are no goalposts for this field,and there are no goalposts for theindustry when they say that they are
3:07.280–3:15.320
ultimately trying to recreate AI systemsthat would be as smart as humans. How dowe even define what that means? And when
3:15.320–3:20.600
are we going to get there if we don'tknow how to define the destination?
3:20.600–3:28.840
>> Karen shows how OpenAI has redefined AGIdepending on who's in the room. When SamAltman talks to Congress, AGI is the
3:28.840–3:35.640
thing that will cure cancer, solveclimate change, cure poverty. When he'selling to regular customers like us,
3:35.640–3:42.600
it's the most amazing digital assistantyou'll ever have. When Microsoft waswriting the investment contract, AGI was
3:42.600–3:52.160
legally defined as a system thatgenerates $100 billion in revenue.>> They can use it however they want to,and they can define and redefine it
3:52.160–3:58.200
based on what is convenient for them.So, in OpenAI's history, it has definedand redefined it many times. When Sam
3:58.200–4:05.840
Altman is talking with Congress, AGI isa system that's going to cure cancer,solve climate change, cure poverty. When
4:05.840–4:12.200
he's talking with consumers that he'strying to sell his products to, it's themost amazing digital assistant that
4:12.200–4:18.600
you're going to ever going to have. Whenhe was talking with Microsoft, you know,in the deal that OpenAI and Microsoft
4:18.600–4:25.480
struck where Microsoft invested in thecompany, it was defined as a system thatwill generate a hundred billion dollars
4:25.480–4:33.560
of revenue. And on OpenAI's own website,they define it as highly autonomousystems that outperform humans in most
4:33.560–4:42.040
economically valuable work. This is likenot a coherent vision of one technology.These are very different definitions
4:42.040–4:50.720
that are spoken out loud to the audiencethat needs to be mobilized to ward offregulation or get more consumer buy-in
4:50.720–4:57.360
to the the industry's quest.>> Read that again. Same three letters,completely different meaning depending
4:57.360–5:04.160
on who needs convincing. I don't thinkthat's a coincidence. I think that's astrategy.Here's the clip that genuinely made me
5:04.160–5:13.440
sit up.>> In 2015,in a blog post that Sam Altman wrotebefore OpenAI was officially announced,he explicitly outlined the existential
5:13.440–5:21.200
risk by saying, "Development ofsuperhuman machine intelligence isprobably the greatest threat to thecontinued existence of humanity. There
5:21.200–5:28.560
are other threats that I think are morecertain to happen, for example, anengineered virus,but AI is probably the most likely way
5:28.560–5:37.160
to destroy everything.">> Back in 2015, before OpenAI officiallyexisted, Sam Altman wrote a blog posthat said, and I'm paraphrasing
5:37.160–5:44.360
carefully here, that a superhuman AI wasprobably the single biggest threat tohumanity's survival, bigger than almost
5:44.360–5:51.400
anything else.Karen's reporting shows that Altmanwrote that specific line right aroundthe time he was trying to convince Elon
5:51.400–5:59.960
Musk to co-found OpenAI with him. AndMusk, at that exact same time, was doingfamous talks, including one at MIT,
5:59.960–6:08.440
where he compared building advanced AIto summoning the demon.>> And so, in that blog post, if you lookat the the language that Altman uses
6:08.440–6:15.000
ide by side with the language that Muskwas using at the time, it mirrors theall the things that Musk was saying.
6:15.000–6:21.640
>> Yeah.>> years ago Musk was going on podcastsaying, tweeting, whatever, that thegreatest existential risk to humanity
6:21.640–6:29.360
was AI.>> Musk did end up co-founding OpenAI. Healso ended up leaving, and by his ownaccount, feeling used. The two are now
6:29.360–6:36.800
suing each other.But this next part is where Karen'sreporting feels like something you needto actually know before you open your
6:36.800–6:43.000
chatbot app again.>> And that is why I call these companiesempires of AI.>> What do you mean by an imperial agenda?
6:43.000–6:44.280
What What does that term mean?
6:44.280–6:51.840
>> Empire is the only metaphor that I've[music] ever found to fully encapsulateall of the dimensions of what these
6:51.840–7:00.640
companies do,and the scale that they operate, andwhat motivates them to do what they do.And there are many parallels that you
7:00.640–7:09.080
see between what I call the empires ofAI and the empires of old.>> Karen's whole book rests on one boldclaim: These are not just companies. She
7:09.080–7:16.720
calls them empires of AI, and she laysout four specific parallels to empiresof old. One, resource grabbing.
7:16.720–7:22.360
She says these companies lay claim toresources that aren't theirs, thepersonal data of ordinary people, the
7:22.360–7:28.720
intellectual property of artists andwriters, and land for building massivecomputing facilities. Two, labor
7:28.720–7:36.160
exploitation.Hundreds of thousands of workers aroundthe world, including in the US, arecontracted to help build these systems,
7:36.160–7:43.240
often in ways that quietly strip awaylabor protections once the tools aredeployed.Three, controlling knowledge.
7:43.240–7:48.640
These corporate empires project thefalse narrative that they are the onlyones pure enough or smart enough to
7:48.640–7:56.000
understand and guide AI safely. Theyactively gaslight the public and policymakers. Even worse, they spend hundreds
7:56.000–8:01.880
of millions of dollars to kill publicsafety legislation and ruthlessly censorany internal researchers whose findings
8:01.880–8:09.040
threaten their corporate profits.Take the case of Dr. Timnit Gebru atGoogle.She was hired to lead an ethical AI
8:09.040–8:15.640
team. But when she co-authored abrilliant paper highlighting the severebiases and environmental harms of large
8:15.640–8:22.840
language models, Google executives triedto suppress it and ultimately fired herand her co-lead, Margaret Mitchell.
8:22.840–8:28.960
Four, myth-making. This is the one thatgot me. Karen says these companies needthe public to believe two things at
8:28.960–8:35.120
once.That AI could be catastrophic and thatit could bring utopia. Why both? Because
8:35.120–8:40.880
if it's dangerous, only they should betrusted to control it. And if it'sutopian, we should hand them our trust,
8:40.880–8:48.160
our data, and our money to get therefaster. Here's a real quote she bringsup from 2017.>> the record on podcasts in their blog
8:48.160–8:54.240
posts to see how it's evolved over time.And Dario, who was the former VP ofresearch at OpenAI, and has now moved on
8:54.240–9:00.360
to Anthropic, who are taking a slightlydifferent approach to developing AI,said back in 2017 while he was still at
9:00.360–9:08.720
OpenAI, that I think at the extreme endis the Nick Bostrom style of fear thatan AGI could destroy humanity. I can't
9:08.720–9:16.160
see any reason in principle why thatcouldn't happen. My chance thatsomething goes really quitecatastrophically wrong on the scale of
9:16.160–9:26.760
human civilizationmight be somewhere between 10% and 25%.>> 10 to 25%. Read that number again.That's not a fringe blogger. That's a
9:26.760–9:35.360
man now running one of the most fundedAI labs on Earth.I thought the empire argument was themost shocking part of this conversation.
9:35.360–9:42.120
I was wrong. The next section about asmall community in Tennessee is the parthat actually stayed with me overnight.
9:42.120–9:51.040
We talk about AI like it lives in thecloud, like it's weightless. It is not.>> OpenAI's largest, one of its largest
9:51.040–9:56.800
data center projects, is being built inAbilene, Texas as part of the Stargateinitiative, which was an effort
9:56.800–10:03.840
announced at the beginning of Trump'second administration to spend $500billion on AI computing infrastructure.
10:03.840–10:14.760
That one would be the size of CentralPark.And it would run a million computerchips, and it wouldrequire the power of more than 20% of
10:14.760–10:22.640
New York City.>> Then there's a meta facility being builtin Louisiana that's reportedly fourtimes the size of that Texas one, using
10:22.640–10:31.000
about half the average power demand ofthe entire city of New York.Karen talks about Memphis, Tennessee,where Elon Musk built Colossus, the
10:31.000–10:38.200
supercomputer used to train Grok,powered partly by 35 methane gasturbines.The community living there, a working
10:38.200–10:44.880
class, reportedly wasn't formally toldthey'd be hosting this facility.According to Karen, residents first
10:44.880–10:51.920
realized something had changed when theysmelled what seemed like a gas leakinside their own homes.That community, she says, already has
10:51.920–10:58.440
ome of the highest rates of asthma andrespiratory illness in the area.Now add industrial-scale gas turbines
10:58.440–11:05.520
running next door, 24 hours a day, topower a machine most of those residentswill never directly benefit from.
11:05.520–11:12.160
And in a cruel twist, the same wave ofautomation this facility helps build isalso the thing most likely to replace
11:12.160–11:22.160
the jobs many of them do.I want to be fair here, because I thinkit matters. AI data centers also createconstruction jobs, local tax revenue,
11:22.160–11:29.160
and in some cases genuinely usefultechnology for medicine and science.This isn't a simple villain story.
11:29.160–11:34.240
Karen says these companies pick some ofthe most vulnerable communities forthese projects. And when you look at
11:34.240–11:40.920
where facilities like this actuallyland, it's hard to call that acoincidence, either.Now, here's the part where you might
11:40.920–11:48.760
expect me to just tell you to log offand boycott AI forever.That's not what Karen says. Her answeris more interesting and honestly more
11:48.760–11:57.160
useful.Karen's suggestion isn't delete every AIapp. It's smarter than that. She callsall of us data donors, meaning every
11:57.160–12:03.800
time we use these tools, we're feedingthe very systems these companies need tokeep growing. Her point, you have more
12:03.800–12:09.440
leverage than you think. She lists a fewconcrete things, and I want to walkthrough them simply. Withhold what you
12:09.440–12:15.600
can. Artists and writers suing thesecompanies over their work are, in herwords, trying to withhold the resource
12:15.600–12:23.640
these companies need most, originalhuman data. Get loud locally.>> And that's what we've seen with all ofthese people that are now literally
12:23.640–12:29.280
protesting the streets against datacenters and having an enormous effect,by the way, actually stalling data
12:29.280–12:36.320
center projects and also completelybanning data centers from beingdeveloped in their localities.>> She points out that data centers are
12:36.320–12:42.760
popping up in ordinary towns right now,and in several places, local communitieshave already succeeded in stalling or
12:42.760–12:50.280
outright banning new data centerprojects through public pressure. Pushat your school or workplace.If your company or university is
12:50.280–12:56.720
drafting an AI adoption policy, thatconversation is one small piece of amuch bigger puzzle, and your voice in
12:56.720–13:05.200
that room genuinely counts. Supportalternatives.Karen says there's research showing thesame useful AI capabilities can be built
13:05.200–13:11.800
with far more efficient methods and farless resource consumption, what shecalls the bicycles of AI as opposed to
13:11.800–13:20.200
the resource-hungry rockets.She also cites a striking number.>> 80% of Americans in the most recent pollthink that the AI industry need to be
13:20.200–13:25.790
regulated.>> Yeah, I see.>> When was the last time that 80% ofAmericans were on the same side of anissue?
13:25.800–13:31.520
>> A shift she says happened largelybecause of stories like the ones in herbook reaching the public. I want to end
13:31.520–13:38.200
this fairly because Karen herself does.Late in the conversation, Steven pushesback a little saying that as an
13:38.200–13:45.280
entrepreneur and investor, he genuinelyfinds AI incredible. Safer driving,tools that help him create things he
13:45.280–13:52.400
couldn't before. Real value for over abillion users. And Karen doesn't dismissthat. Her answer is basically, you can
13:52.400–14:01.320
hold both truths at once.AI can be genuinely useful and thecurrent way it's being built and rushedto market can be genuinely harmful.
14:01.320–14:07.120
Those aren't contradictions. They're twohonest observations about the sameindustry. That's actually the most
14:07.120–14:14.640
balanced moment in the whole two hours.And it's the note I want to end on, too.So, here's my question for you and I
14:14.640–14:20.280
genuinely want to know. Knowing what younow know about where the water for yourchatbot's training run might be coming
14:20.280–14:24.760
from, does it change how you'll usethese tools tomorrow? Not never. I'm not
14:24.760–14:32.520
asking you to swear off technology. Justpay attention. Ask questions.Support the version of AI that doesn't
14:32.520–14:40.000
need a small town in Tennessee to smellgas in their living room to make itwork. If this made you thinkdifferently, even a little, share it
14:40.000–14:46.200
with one person who uses AI every daywithout ever asking these questions.That's genuinely how change like this
14:46.200–14:48.800
tarts.
0:00.000–0:04.080
為什麼我們要試圖建構與人類重複的AI系統?他們
0:04.080–0:09.400
說我們應該建構AGI,所以我們說我們應該建構AGI,但我想要問的是,為什麼我們
0:09.400–0:12.240
在做這件事?為什麼我們
0:12.240–0:19.760
建構的技術最終是為了取代和自動化人們?我們應該建構技術
0:19.760–0:28.320
而技術的目的在歷史上一直是為了提升人類繁榮,而不是取代人們。這是
0:28.320–0:30.280
正確的目標嗎?為什麼我們
0:30.280–0:33.160
在做這件事?為什麼我們不能只是建構
0:33.160–0:39.520
人工智慧系統,這些系統與它們試圖建構的統計引擎毫無關係,
0:39.520–0:41.360
複製人腦?
0:41.360–0:48.640
>> 這可能是我今年聽過關於AI最不舒服的問題。不是AI會不會比人類更聰明,
0:48.640–0:56.440
不僅是人工智慧會毀滅人類,而是更簡單的問題。為什麼我們試圖建造取代人類而非幫助人類的機器?想想
0:56.440–1:02.640
看。當汽車被發明時,它們幫助我們更快地旅行。當智慧型手機普及時,它們
1:02.640–1:08.800
想一下。當汽車被發明時,它們幫助我們更快地旅行。當智慧型手機普及時,它們
1:08.800–1:15.040
將我們與全世界連接起來。技術一直是關於賦予人類更多權力,但這是第一次
1:15.040–1:19.440
數百萬人問了一個完全不同的問題。當技術開始取代
1:19.440–1:26.560
人們而不是僅僅幫助人們時,會發生什麼事?這正是我想要
1:26.560–1:35.320
回應這個播客的原因。史蒂文·巴特利特(Steven Bartlett)和記者凱倫·豪(Karen Hao)之間兩小時的對話,這位女性花了
1:35.320–1:41.880
8年和超過300次採訪,調查OpenAI等公司內部真正發生的事情。我必須對你誠實。她說的某些話讓我感到憤怒。有些話讓我質疑我每天使用的東西,還有一部分關於田納西州一個小鎮的內容讓我真正感到不安。讓我們一起探討它,並給出我對此的誠實意見。我想先給你一些背景
1:41.880–1:50.480
因為如果不了解領域,就無法理解這場爭鬥。>> 所以,這是在1956年,[音樂]有一群
1:50.480–1:59.080
科學家聚集在達特茅斯學院,開始一個新的學科,一個科學學科,試圖追求
1:59.080–2:05.040
一個抱負。具體來說,達特茅斯學院的一位助理教授約翰·麥卡錫(John McCarthy)決定
2:05.040–2:11.400
將這個學科命名為人工智慧。這不是他嘗試的第一個名字。前一年,他嘗試
2:11.400–2:18.320
將其命名為自動機研究。他的某些同事對這個名字感到擔憂的原因是,因為它
2:18.320–2:25.040
命名為自動化研究。而他的一些同事對這個名稱感到擔憂的原因是,因為它
2:25.040–2:31.400
人工智慧這個詞深感擔憂,因為它將該領域與複製人類智慧的目標聯繫起來。無論當時
2:31.400–2:38.520
還是現在,對於人類智慧究竟是什麼,絕對沒有科學共識。沒有一個來自生物學、
2:38.520–2:46.000
心理學或神經學的單一定義,這意味著,正如凱倫所指出的,當今天的AI公司談論建構AGI時,
2:46.000–2:54.040
人工通用智慧,他們正在追逐一個根本沒有固定座標的終點線。
2:54.040–3:00.200
人工通用智慧,他們追逐的終點線根本沒有固定的座標。
3:00.200–3:07.280
他們最終試圖建構與人類一樣聰明的人類智慧系統時,行業也沒有球門柱。我們甚至如何定義這意味著什麼?如果我們不知道如何定義目的地,
3:07.280–3:15.320
最終試圖重建與人類一樣聰明的人工智慧系統。我們該如何定義這意味著什麼?並且當
3:15.320–3:20.600
>> 凱倫展示了OpenAI如何根據房間裡的人重新定義AGI。當山姆·阿爾特曼(Sam Altman)對國會說話時,AGI是
3:20.600–3:28.840
將治癒癌症、解決氣候變化、治癒貧窮的東西。當他向像我們這樣的普通客戶推銷時,
3:28.840–3:35.640
它將是你擁有過的最驚人的數位助手。當微軟撰寫投資合約時,AGI在法律上被定義為
3:35.640–3:42.600
產生1000億美元收入的系統。>> 他們可以隨心所欲地使用它,並且可以根據對他們方便的情況來定義和重新定義它。所以,在OpenAI的歷史中,它已經定義和重新定義了它很多次。當山姆
3:42.600–3:52.160
阿爾特曼與國會談話時,AGI是一個將治癒癌症、解決氣候變化、治癒貧窮的系統。當
3:52.160–3:58.200
基於對他們方便的事。因此,在 OpenAI 的歷史中,它定義並重新定義了許多次。當山姆
3:58.200–4:05.840
將擁有過的最驚人的數位助手。當他與微軟談話時,你知道,在OpenAI和微軟的交易中
4:05.840–4:12.200
當他與消費者對話,也就是他試圖銷售產品的對象時,它是最令人驚嘆的數位助理
4:12.200–4:18.600
當你即將擁有的。當他與微軟對話時,你知道,在 OpenAI 和微軟的交易中
4:18.600–4:25.480
微軟投資該公司時,將其定義為將產生一千億美元營收的系統。
4:25.480–4:33.560
而在 OpenAI 自己的網站上,他們將其定義為在大多數經濟上有價值的工作中表現優於人類的、高度自主的系統。
4:33.560–4:42.040
這就像沒有關於一項技術的連貫願景。這些是非常不同的定義。
4:42.040–4:50.720
這些定義是針對需要動員起來以抵禦監管或獲得更多消費者支持的受眾說出口的。
4:50.720–4:57.360
>> 再讀一次。同樣的三個字母,根據需要說服對象的不同,意義完全相反。
4:57.360–5:04.160
我不認為這是巧合。我認為這是一種策略。這裡有一段讓我真正坐起來的影片。
5:04.160–5:13.440
>> 在 2015 年,Sam Altman 在 OpenAI 正式宣佈之前撰寫的一篇部落格文章中,他明確概述了存在性風險,他說:「開發出超人類的機器智能,可能是人類持續存在最大的威脅。」
5:13.440–5:21.200
「我認為其他威脅更有可能發生,例如製造的病毒,但 AI 可能是摧毀一切最可能的方式。」
5:21.200–5:28.560
>> 回到 2015 年,在 OpenAI 正式成立之前,Sam Altman 寫了一篇部落格文章,我說,這裡我謹慎地轉述,
5:28.560–5:37.160
「超人類的 AI 可能是人類生存的最大威脅,比幾乎任何其他事物都大。」
5:37.160–5:44.360
Karen 的報導顯示,Altman 在試圖說服 Elon Musk 與他共同創立 OpenAI 的同時,寫下了這句特定的話。
5:44.360–5:51.400
而 Musk 在當時的同一時間,進行了著名的演講,包括在麻省理工學院的一次演講,
5:51.400–5:59.960
他將建構先進 AI 比作召喚惡魔。>> 所以,在那篇部落格文章中,如果你看看 Altman 使用的語言
5:59.960–6:08.440
與 Musk 當時使用的語言並排比較,它反映了 Musk 所說的一切。
6:08.440–6:15.000
>> 是的。>> 多年前,Musk 在播客、推文等場合表示,對人類存在最大的威脅是 AI。
6:15.000–6:21.640
>> Musk 最終共同創立了 OpenAI。他也最終離開了,據他自己說,感到被利用。兩人現在
6:21.640–6:29.360
互相起訴。但接下來這部分是 Karen 的報導感覺像是你需要在再次打開
6:29.360–6:36.800
聊天機器人應用程式之前真正知道的事情。>> 這就是我稱這些公司為 AI 帝國的原因。>> 你所謂的帝國議程是什麼?
6:36.800–6:43.000
聊天機器人應用程式。>> 這就是我稱這些公司為 AI 帝國的原因。>> 你所謂的帝國議程是什麼意思?
6:43.000–6:44.280
>> 帝國是我找到的唯一一個能完全概括這些公司所做的一切、它們運作的規模以及驅動它們這樣做的動機的隱喻。
6:44.280–6:51.840
你會看到我所謂的 AI 帝國與舊帝國之間有許多相似之處。>> Karen 的整本書基於一個大膽的主張:這些不只是公司。她
6:51.840–7:00.640
稱它們為 AI 帝國,並列出了四個與舊帝國的具體相似之處。第一,掠奪資源。
7:00.640–7:09.080
看看我所謂的 AI 帝國與舊帝國之間的區別。>> 凱倫的整本書都建立在一個大膽的主張上:這些不僅僅是公司。她
7:09.080–7:16.720
藝術家和作家的知識產權,以及用於建造大型計算設施的土地。第二,勞動力剝削。
7:16.720–7:22.360
全球數十萬名工人,包括在美國的工人,受僱協助建構這些系統,
7:22.360–7:28.720
通常是在工具部署後悄悄剝奪勞動保護的方式下進行的。第三,控制知識。
7:28.720–7:36.160
這些企業帝國投射出虛假的敘事,認為它們是唯一純潔或足夠聰明以
7:36.160–7:43.240
安全理解和引導 AI 的實體。它們積極地對公眾和政策制定者進行煤气燈效應(gaslighting)。更糟的是,它們花費數億美元來扼殺公共安全立法,並冷酷地審查任何可能威脅其企業利潤的內部研究人員。
7:43.240–7:48.640
以 Google 的 Dr. Timnit Gebru 為例。她被僱用來領導一個道德 AI 團隊。
7:48.640–7:56.000
但當她共同撰寫了一篇強調大型語言模型嚴重偏差和環境危害的出色論文時,Google 高管試圖壓制它,最終解僱了她和她的共同負責人 Margaret Mitchell。
7:56.000–8:01.880
第四,神話建構。這是讓我著迷的部分。Karen 說這些公司需要公眾同時相信兩件事。
8:01.880–8:09.040
那就是 AI 可能是災難性的,同時它也可能帶來烏托邦。為什麼兩者都要?因為
8:09.040–8:15.640
如果它危險,只有它們才應該被信任來控制它。而如果它是烏托邦,我們應該把我們的信任、
8:15.640–8:22.840
我們的數據和我們的錢交給它們,以便更快地到達那裡。這裡有一個她引用的 2017 年的真實引言。>> 播客記錄在其部落格
8:22.840–8:28.960
文章中,看看它如何隨時間演變。而 Dario,前 OpenAI 研究副總裁,現在已轉投
8:28.960–8:35.120
Anthropic,他們在開發 AI 方面採取了略有不同的方法,他在 2017 年還在
8:35.120–8:40.880
OpenAI 時曾表示,在最極端的情況下,是 Nick Bostrom 風格的恐懼,即 AGI 可能摧毀人類。我看不出原則上為什麼這不可能發生。我認為某事在人類文明規模上真正災難性地出錯的可能性可能在 10% 到 25% 之間。>> 10 到 25%。再讀一次那個數字。那不是邊緣部落客。那是一位
8:40.880–8:48.160
我們的數據和資金,以更快的速度到達那裡。這裡有一個她引用的 2017 年的真實引言。>> 播客和他們博客上的文章
8:48.160–8:54.240
文章,看看它如何隨時間演變。達里奧曾是 OpenAI 的研究副總裁,現在已轉投
8:54.240–9:00.360
Anthropic,他們在開發 AI 方面採取了略有不同的方法,他在 2017 年仍於
9:00.360–9:08.720
在OpenAI時,我認為在最極端的情況下,是尼克·波斯特羅姆式的恐懼,認為AGI可能會毀滅人類。我無法
9:08.720–9:16.160
從原則上看出為什麼這不可能發生。我認為在人類文明規模上出現非常災難性錯誤的可能性,
9:16.160–9:26.760
可能在10%到25%之間。>> 10%到25%。再讀一次這個數字。這不是一個邊緣部落客。這是一位
9:26.760–9:35.360
現在掌管地球上資金最豐厚的AI實驗室之一的男人。我原以為關於帝國論述的部分是這場對話中最令人震驚的。
9:35.360–9:42.120
我錯了。關於田納西州一個小社區的下一段,才是真正讓我過夜難眠的部分。
9:42.120–9:51.040
我們談論AI時,彷彿它住在雲端,彷彿它沒有重量。事實並非如此。>> OpenAI最大、其中一個最大的
9:51.040–9:56.800
數據中心項目,正在德克薩斯州阿比林建造,作為星際計劃的一部分,該計劃是在
9:56.800–10:03.840
特朗普第二任期開始時宣布的,旨在花費5000億美元用於AI計算基礎設施。
10:03.840–10:14.760
這個數據中心將擁有中央公園的大小。它將運行一百萬個計算芯片,並且需要超過紐約市
10:14.760–10:22.640
20%的電力。>> 然後,路易斯安那州正在建設一個元設施,據報導其規模是德克薩斯州那個的四倍,使用
10:22.640–10:31.000
大約相當於紐約市整體平均電力需求的一半。凱倫談到了田納西州的孟菲斯,埃隆·馬斯克在那裡建造了科索斯,
10:31.000–10:38.200
用於訓練Grok的超級計算機,部分由35台甲烷燃氣渦輪機供電。生活在該社區的居民,據報是工人階層,
10:38.200–10:44.880
並未正式被告知他們將 hosting 這個設施。根據凱倫的說法,居民們首先
10:44.880–10:51.920
意識到有些東西改變了,當他們在家裡聞到似乎像是氣體洩漏的味道時。她說,這個社區已經
10:51.920–10:58.440
擁有該地區最高比例的哮喘和呼吸道疾病率之一。現在加上工業規模的燃氣渦輪機
10:58.440–11:05.520
在隔壁運行,24小時不間斷,為一台大多數居民永遠無法直接受益的機器供電。
11:05.520–11:12.160
並且在一個諷刺的轉折中,這個設施幫助構建的自動化浪潮,也是最可能取代
11:12.160–11:22.160
他們中許多人工作的東西。我想在這裡保持公平,因為我認為這很重要。AI數據中心也創造了
11:22.160–11:29.160
建築工作、地方稅收,在某些情況下,確實為醫學和科學使用了有用的技術。這並不是一個簡單的邪惡故事。
11:29.160–11:34.240
凱倫說,這些公司選擇了一些最脆弱的社區來進行這些項目。當你查看
11:34.240–11:40.920
此類設施實際上落腳的地方時,很難稱之為巧合。現在,這裡是你可能會
11:40.920–11:48.760
期望我告訴你不要登入並永遠抵制AI的部分。這不是凱倫所說的。她的答案更有趣,也誠實地說更
11:48.760–11:57.160
有用。凱倫的建議不是刪除每個AI應用。這比那更聰明。她稱我們所有人為數據捐贈者,意味著每次
11:57.160–12:03.800
我們使用這些工具時,我們都在餵養這些公司需要繼續增長的系統。她的觀點是,你擁有比你想像中更多
12:03.800–12:09.440
的槓桿。她列出了幾件具體的事情,我想簡單地 walkthrough 它們。盡你所能 withhold(保留/不提供)。藝術家和作家起訴這些
12:09.440–12:15.600
公司關於他們的作品,在她看來,試圖 withhold(保留/不提供)這些公司最需要的資源,
12:15.600–12:23.640
原始的真人數據。在當地大聲疾呼。>> 這就是我們看到的所有這些人現在實際上
12:23.640–12:29.280
在街頭抗議數據中心,並且以某種方式產生了巨大的影響,實際上阻滯了數據
12:29.280–12:36.320
中心的項目,並且完全禁止在他們當地開發數據中心。>> 她指出,數據中心
12:36.320–12:42.760
正在普通城鎮湧現,在幾個地方,當地社區已經成功通過公眾壓力阻滯或
12:42.760–12:50.280
outright(完全)禁止新的數據中心項目。在你學校或工作場所施加壓力。如果你的公司或大學
12:50.280–12:56.720
正在起草AI採用政策,那場對話只是更大拼圖中的一小部分,而你在那個房間裡的聲音確實
12:56.720–13:05.200
有分量。支持替代方案。凱倫說,有研究表明,同樣有用的AI功能可以用
13:05.200–13:11.800
更高效的方法和遠少資源消耗來構建,她稱之為AI的自行車,而不是
13:11.800–13:20.200
資源飢渴的火箭。她還引用了一個引人注目的數字。>> 在最近的一次民意調查中,80%的美國人認為
13:20.200–13:25.790
AI行業需要監管。>> 是的,我看到了。>> 上一次80%的美國人在同一立場上是什么时候?
13:25.800–13:31.520
>> 她說,這種轉變很大程度上是因為像她書中那樣的故事情節被公眾所知。我想以公平的方式結束
13:31.520–13:38.200
這場對話,因為凱倫自己也這樣做。在對話後期,史蒂文稍微反駁說,作為一名
13:38.200–13:45.280
企業家和投資者,他真心認為AI令人驚嘆。更安全的駕駛,幫助他創造以前無法創造的東西的工具。
13:45.280–13:52.400
對於超過十億用戶的真實價值。而凱倫並沒有否定這一點。她的答案基本上是,你可以
13:52.400–14:01.320
同時接受這兩種事實。AI 確實能帶來實質幫助,而目前它被倉促建構並推向市場的方式,也可能造成真正的傷害。
14:01.320–14:07.120
這並非矛盾,而是對同一產業的兩個誠實觀察。這其實是整個兩小時對話中最具平衡感的一刻。
14:07.120–14:14.640
這也是我想帶出的結語。所以,這裡有個問題想問你,我是真心想了解。
14:14.640–14:20.280
現在你已經知道,你的聊天機器人訓練用水可能來自哪裡,這會改變你明天使用這些工具的方式嗎?
14:20.280–14:24.760
不是說永遠不再使用。我不是要你發誓遠離科技。
14:24.760–14:32.520
只是要你保持關注、提出問題,並支持那種不需要田納西州某個小鎮的居民在客廳裡聞到氣味,才能讓它運作的 AI 版本。
14:32.520–14:40.000
如果這讓你產生了不同的想法,哪怕只有一點點,請把它分享給一個每天使用 AI 卻從未問過這些問題的人。
14:40.000–14:46.200
與每天使用 AI 卻從未提出這些問題的一個人在一起。這確實是像這樣的改變
14:46.200–14:48.800
起頭。
0:00.000–0:04.080
Why are we trying to build AI systemsthat are duplicative of humans? They
為什麼我們要試圖建構與人類重複的AI系統?他們
0:04.080–0:09.400
said that we should be building AGI, sowe say that we should be building AGI,but I would like to ask like why are we
說我們應該建構AGI,所以我們說我們應該建構AGI,但我想要問的是,為什麼我們
0:09.400–0:12.240
doing that? Why is it that we are
在做這件事?為什麼我們
0:12.240–0:19.760
building a technology that is ultimatelydesigned to replace and automate peopleaway? We should be building technology
建構的技術最終是為了取代和自動化人們?我們應該建構技術
0:19.760–0:28.320
and the purpose of technology throughouthistory has been to improve humanflourishing, not to replace people. Is
而技術的目的在歷史上一直是為了提升人類繁榮,而不是取代人們。這是
0:28.320–0:30.280
this the right goal? What like why are
正確的目標嗎?為什麼我們
0:30.280–0:33.160
we doing this? Why can't we just build
在做這件事?為什麼我們不能只是建構
0:33.160–0:39.520
AI systems, which are systems that havenothing to do with the statisticalengines that they're trying to build to
人工智慧系統,這些系統與它們試圖建構的統計引擎毫無關係,
0:39.520–0:41.360
duplicate the human brain?
複製人腦?
0:41.360–0:48.640
>> That might be the most uncomfortablequestion I've heard about AI this year.Not can AI become smarter than humans,
>> 這可能是我今年聽過關於AI最不舒服的問題。不是AI會不會比人類更聰明,
0:48.640–0:56.440
not will AI destroy humanity, butsomething much simpler. Why are wetrying to build machines that replacepeople instead of helping people? Think
不僅是人工智慧會毀滅人類,而是更簡單的問題。為什麼我們試圖建造取代人類而非幫助人類的機器?想想
0:56.440–1:02.640
about it for a second. When cars wereinvented, they helped us travel faster.When smartphones became common, they
看。當汽車被發明時,它們幫助我們更快地旅行。當智慧型手機普及時,它們
1:02.640–1:08.800
connected us to the whole world.Technology has always been about givinghumans more power, but for the first
想一下。當汽車被發明時,它們幫助我們更快地旅行。當智慧型手機普及時,它們
1:08.800–1:15.040
ime, millions of people are asking acompletely different question. Whathappens when technology starts replacing
將我們與全世界連接起來。技術一直是關於賦予人類更多權力,但這是第一次
1:15.040–1:19.440
the person instead of just helping theperson? And that's exactly why I wanted
數百萬人問了一個完全不同的問題。當技術開始取代
1:19.440–1:26.560
to react to this podcast. A 2-hourconversation between Steven Bartlett andjournalist Karen Hao, a woman who spent
人們而不是僅僅幫助人們時,會發生什麼事?這正是我想要
1:26.560–1:35.320
8 years and over 300 interviewsinvestigating what's really happeninginside companies like OpenAI.And I need to be honest with you. Some
回應這個播客的原因。史蒂文·巴特利特(Steven Bartlett)和記者凱倫·豪(Karen Hao)之間兩小時的對話,這位女性花了
1:35.320–1:41.880
of what she says made me angry. Some ofit made me question things I use everyday, and one part about a small town in
8年和超過300次採訪,調查OpenAI等公司內部真正發生的事情。我必須對你誠實。她說的某些話讓我感到憤怒。有些話讓我質疑我每天使用的東西,還有一部分關於田納西州一個小鎮的內容讓我真正感到不安。讓我們一起探討它,並給出我對此的誠實意見。我想先給你一些背景
1:41.880–1:50.480
Tennessee genuinely disturbed me.Let's walk through it and give my honestopinion about it.I want to give you some background first
因為如果不了解領域,就無法理解這場爭鬥。>> 所以,這是在1956年,[音樂]有一群
1:50.480–1:59.080
because you can't understand the fightwithout understanding the field.>> So, this was back in 1956and [music] there were a group of
科學家聚集在達特茅斯學院,開始一個新的學科,一個科學學科,試圖追求
1:59.080–2:05.040
scientists that gathered at DartmouthUniversity to start a new discipline, ascientific discipline, to try and chase
一個抱負。具體來說,達特茅斯學院的一位助理教授約翰·麥卡錫(John McCarthy)決定
2:05.040–2:11.400
an ambition. And specifically, anassistant professor at DartmouthUniversity, John McCarthy, decided to
將這個學科命名為人工智慧。這不是他嘗試的第一個名字。前一年,他嘗試
2:11.400–2:18.320
name this discipline artificialintelligence.This was not the first name that hetried. The previous year, he tried to
將其命名為自動機研究。他的某些同事對這個名字感到擔憂的原因是,因為它
2:18.320–2:25.040
name it Automata Studies. And the reasonwhy some of his colleagues wereconcerned about this name was because it
命名為自動化研究。而他的一些同事對這個名稱感到擔憂的原因是,因為它
2:25.040–2:31.400
pegged the idea of this discipline torecreating human intelligence.>> His colleagues were deeply concerned
人工智慧這個詞深感擔憂,因為它將該領域與複製人類智慧的目標聯繫起來。無論當時
2:31.400–2:38.520
about the term artificial intelligencebecause it tied the field to the goal ofrecreating human intelligence. Both then
還是現在,對於人類智慧究竟是什麼,絕對沒有科學共識。沒有一個來自生物學、
2:38.520–2:46.000
and now, there is absolutely noscientific consensus on what humanintelligence actually is. There is nosingle definition from biology,
心理學或神經學的單一定義,這意味著,正如凱倫所指出的,當今天的AI公司談論建構AGI時,
2:46.000–2:54.040
psychology, or neurology, which means,as Karen points out, when today's AIcompanies talk about building AGI,
人工通用智慧,他們正在追逐一個根本沒有固定座標的終點線。
2:54.040–3:00.200
artificial general intelligence, they'rechasing a finish line that literallydoesn't have fixed coordinates.
人工通用智慧,他們追逐的終點線根本沒有固定的座標。
3:00.200–3:07.280
>> Well,there are no goalposts for this field,and there are no goalposts for theindustry when they say that they are
他們最終試圖建構與人類一樣聰明的人類智慧系統時,行業也沒有球門柱。我們甚至如何定義這意味著什麼?如果我們不知道如何定義目的地,
3:07.280–3:15.320
ultimately trying to recreate AI systemsthat would be as smart as humans. How dowe even define what that means? And when
最終試圖重建與人類一樣聰明的人工智慧系統。我們該如何定義這意味著什麼?並且當
3:15.320–3:20.600
are we going to get there if we don'tknow how to define the destination?
>> 凱倫展示了OpenAI如何根據房間裡的人重新定義AGI。當山姆·阿爾特曼(Sam Altman)對國會說話時,AGI是
3:20.600–3:28.840
>> Karen shows how OpenAI has redefined AGIdepending on who's in the room. When SamAltman talks to Congress, AGI is the
將治癒癌症、解決氣候變化、治癒貧窮的東西。當他向像我們這樣的普通客戶推銷時,
3:28.840–3:35.640
thing that will cure cancer, solveclimate change, cure poverty. When he'selling to regular customers like us,
它將是你擁有過的最驚人的數位助手。當微軟撰寫投資合約時,AGI在法律上被定義為
3:35.640–3:42.600
it's the most amazing digital assistantyou'll ever have. When Microsoft waswriting the investment contract, AGI was
產生1000億美元收入的系統。>> 他們可以隨心所欲地使用它,並且可以根據對他們方便的情況來定義和重新定義它。所以,在OpenAI的歷史中,它已經定義和重新定義了它很多次。當山姆
3:42.600–3:52.160
legally defined as a system thatgenerates $100 billion in revenue.>> They can use it however they want to,and they can define and redefine it
阿爾特曼與國會談話時,AGI是一個將治癒癌症、解決氣候變化、治癒貧窮的系統。當
3:52.160–3:58.200
based on what is convenient for them.So, in OpenAI's history, it has definedand redefined it many times. When Sam
基於對他們方便的事。因此,在 OpenAI 的歷史中,它定義並重新定義了許多次。當山姆
3:58.200–4:05.840
Altman is talking with Congress, AGI isa system that's going to cure cancer,solve climate change, cure poverty. When
將擁有過的最驚人的數位助手。當他與微軟談話時,你知道,在OpenAI和微軟的交易中
4:05.840–4:12.200
he's talking with consumers that he'strying to sell his products to, it's themost amazing digital assistant that
當他與消費者對話,也就是他試圖銷售產品的對象時,它是最令人驚嘆的數位助理
4:12.200–4:18.600
you're going to ever going to have. Whenhe was talking with Microsoft, you know,in the deal that OpenAI and Microsoft
當你即將擁有的。當他與微軟對話時,你知道,在 OpenAI 和微軟的交易中
4:18.600–4:25.480
struck where Microsoft invested in thecompany, it was defined as a system thatwill generate a hundred billion dollars
微軟投資該公司時,將其定義為將產生一千億美元營收的系統。
4:25.480–4:33.560
of revenue. And on OpenAI's own website,they define it as highly autonomousystems that outperform humans in most
而在 OpenAI 自己的網站上,他們將其定義為在大多數經濟上有價值的工作中表現優於人類的、高度自主的系統。
4:33.560–4:42.040
economically valuable work. This is likenot a coherent vision of one technology.These are very different definitions
這就像沒有關於一項技術的連貫願景。這些是非常不同的定義。
4:42.040–4:50.720
that are spoken out loud to the audiencethat needs to be mobilized to ward offregulation or get more consumer buy-in
這些定義是針對需要動員起來以抵禦監管或獲得更多消費者支持的受眾說出口的。
4:50.720–4:57.360
to the the industry's quest.>> Read that again. Same three letters,completely different meaning depending
>> 再讀一次。同樣的三個字母,根據需要說服對象的不同,意義完全相反。
4:57.360–5:04.160
on who needs convincing. I don't thinkthat's a coincidence. I think that's astrategy.Here's the clip that genuinely made me
我不認為這是巧合。我認為這是一種策略。這裡有一段讓我真正坐起來的影片。
5:04.160–5:13.440
sit up.>> In 2015,in a blog post that Sam Altman wrotebefore OpenAI was officially announced,he explicitly outlined the existential
>> 在 2015 年,Sam Altman 在 OpenAI 正式宣佈之前撰寫的一篇部落格文章中,他明確概述了存在性風險,他說:「開發出超人類的機器智能,可能是人類持續存在最大的威脅。」
5:13.440–5:21.200
risk by saying, "Development ofsuperhuman machine intelligence isprobably the greatest threat to thecontinued existence of humanity. There
「我認為其他威脅更有可能發生,例如製造的病毒,但 AI 可能是摧毀一切最可能的方式。」
5:21.200–5:28.560
are other threats that I think are morecertain to happen, for example, anengineered virus,but AI is probably the most likely way
>> 回到 2015 年,在 OpenAI 正式成立之前,Sam Altman 寫了一篇部落格文章,我說,這裡我謹慎地轉述,
5:28.560–5:37.160
to destroy everything.">> Back in 2015, before OpenAI officiallyexisted, Sam Altman wrote a blog posthat said, and I'm paraphrasing
「超人類的 AI 可能是人類生存的最大威脅,比幾乎任何其他事物都大。」
5:37.160–5:44.360
carefully here, that a superhuman AI wasprobably the single biggest threat tohumanity's survival, bigger than almost
Karen 的報導顯示,Altman 在試圖說服 Elon Musk 與他共同創立 OpenAI 的同時,寫下了這句特定的話。
5:44.360–5:51.400
anything else.Karen's reporting shows that Altmanwrote that specific line right aroundthe time he was trying to convince Elon
而 Musk 在當時的同一時間,進行了著名的演講,包括在麻省理工學院的一次演講,
5:51.400–5:59.960
Musk to co-found OpenAI with him. AndMusk, at that exact same time, was doingfamous talks, including one at MIT,
他將建構先進 AI 比作召喚惡魔。>> 所以,在那篇部落格文章中,如果你看看 Altman 使用的語言
5:59.960–6:08.440
where he compared building advanced AIto summoning the demon.>> And so, in that blog post, if you lookat the the language that Altman uses
與 Musk 當時使用的語言並排比較,它反映了 Musk 所說的一切。
6:08.440–6:15.000
ide by side with the language that Muskwas using at the time, it mirrors theall the things that Musk was saying.
>> 是的。>> 多年前,Musk 在播客、推文等場合表示,對人類存在最大的威脅是 AI。
6:15.000–6:21.640
>> Yeah.>> years ago Musk was going on podcastsaying, tweeting, whatever, that thegreatest existential risk to humanity
>> Musk 最終共同創立了 OpenAI。他也最終離開了,據他自己說,感到被利用。兩人現在
6:21.640–6:29.360
was AI.>> Musk did end up co-founding OpenAI. Healso ended up leaving, and by his ownaccount, feeling used. The two are now
互相起訴。但接下來這部分是 Karen 的報導感覺像是你需要在再次打開
6:29.360–6:36.800
suing each other.But this next part is where Karen'sreporting feels like something you needto actually know before you open your
聊天機器人應用程式之前真正知道的事情。>> 這就是我稱這些公司為 AI 帝國的原因。>> 你所謂的帝國議程是什麼?
6:36.800–6:43.000
chatbot app again.>> And that is why I call these companiesempires of AI.>> What do you mean by an imperial agenda?
聊天機器人應用程式。>> 這就是我稱這些公司為 AI 帝國的原因。>> 你所謂的帝國議程是什麼意思?
6:43.000–6:44.280
What What does that term mean?
>> 帝國是我找到的唯一一個能完全概括這些公司所做的一切、它們運作的規模以及驅動它們這樣做的動機的隱喻。
6:44.280–6:51.840
>> Empire is the only metaphor that I've[music] ever found to fully encapsulateall of the dimensions of what these
你會看到我所謂的 AI 帝國與舊帝國之間有許多相似之處。>> Karen 的整本書基於一個大膽的主張:這些不只是公司。她
6:51.840–7:00.640
companies do,and the scale that they operate, andwhat motivates them to do what they do.And there are many parallels that you
稱它們為 AI 帝國,並列出了四個與舊帝國的具體相似之處。第一,掠奪資源。
7:00.640–7:09.080
see between what I call the empires ofAI and the empires of old.>> Karen's whole book rests on one boldclaim: These are not just companies. She
看看我所謂的 AI 帝國與舊帝國之間的區別。>> 凱倫的整本書都建立在一個大膽的主張上:這些不僅僅是公司。她
7:09.080–7:16.720
calls them empires of AI, and she laysout four specific parallels to empiresof old. One, resource grabbing.
藝術家和作家的知識產權,以及用於建造大型計算設施的土地。第二,勞動力剝削。
7:16.720–7:22.360
She says these companies lay claim toresources that aren't theirs, thepersonal data of ordinary people, the
全球數十萬名工人,包括在美國的工人,受僱協助建構這些系統,
7:22.360–7:28.720
intellectual property of artists andwriters, and land for building massivecomputing facilities. Two, labor
通常是在工具部署後悄悄剝奪勞動保護的方式下進行的。第三,控制知識。
7:28.720–7:36.160
exploitation.Hundreds of thousands of workers aroundthe world, including in the US, arecontracted to help build these systems,
這些企業帝國投射出虛假的敘事,認為它們是唯一純潔或足夠聰明以
7:36.160–7:43.240
often in ways that quietly strip awaylabor protections once the tools aredeployed.Three, controlling knowledge.
安全理解和引導 AI 的實體。它們積極地對公眾和政策制定者進行煤气燈效應(gaslighting)。更糟的是,它們花費數億美元來扼殺公共安全立法,並冷酷地審查任何可能威脅其企業利潤的內部研究人員。
7:43.240–7:48.640
These corporate empires project thefalse narrative that they are the onlyones pure enough or smart enough to
以 Google 的 Dr. Timnit Gebru 為例。她被僱用來領導一個道德 AI 團隊。
7:48.640–7:56.000
understand and guide AI safely. Theyactively gaslight the public and policymakers. Even worse, they spend hundreds
但當她共同撰寫了一篇強調大型語言模型嚴重偏差和環境危害的出色論文時,Google 高管試圖壓制它,最終解僱了她和她的共同負責人 Margaret Mitchell。
7:56.000–8:01.880
of millions of dollars to kill publicsafety legislation and ruthlessly censorany internal researchers whose findings
第四,神話建構。這是讓我著迷的部分。Karen 說這些公司需要公眾同時相信兩件事。
8:01.880–8:09.040
threaten their corporate profits.Take the case of Dr. Timnit Gebru atGoogle.She was hired to lead an ethical AI
那就是 AI 可能是災難性的,同時它也可能帶來烏托邦。為什麼兩者都要?因為
8:09.040–8:15.640
team. But when she co-authored abrilliant paper highlighting the severebiases and environmental harms of large
如果它危險,只有它們才應該被信任來控制它。而如果它是烏托邦,我們應該把我們的信任、
8:15.640–8:22.840
language models, Google executives triedto suppress it and ultimately fired herand her co-lead, Margaret Mitchell.
我們的數據和我們的錢交給它們,以便更快地到達那裡。這裡有一個她引用的 2017 年的真實引言。>> 播客記錄在其部落格
8:22.840–8:28.960
Four, myth-making. This is the one thatgot me. Karen says these companies needthe public to believe two things at
文章中,看看它如何隨時間演變。而 Dario,前 OpenAI 研究副總裁,現在已轉投
8:28.960–8:35.120
once.That AI could be catastrophic and thatit could bring utopia. Why both? Because
Anthropic,他們在開發 AI 方面採取了略有不同的方法,他在 2017 年還在
8:35.120–8:40.880
if it's dangerous, only they should betrusted to control it. And if it'sutopian, we should hand them our trust,
OpenAI 時曾表示,在最極端的情況下,是 Nick Bostrom 風格的恐懼,即 AGI 可能摧毀人類。我看不出原則上為什麼這不可能發生。我認為某事在人類文明規模上真正災難性地出錯的可能性可能在 10% 到 25% 之間。>> 10 到 25%。再讀一次那個數字。那不是邊緣部落客。那是一位
8:40.880–8:48.160
our data, and our money to get therefaster. Here's a real quote she bringsup from 2017.>> the record on podcasts in their blog
我們的數據和資金,以更快的速度到達那裡。這裡有一個她引用的 2017 年的真實引言。>> 播客和他們博客上的文章
8:48.160–8:54.240
posts to see how it's evolved over time.And Dario, who was the former VP ofresearch at OpenAI, and has now moved on
文章,看看它如何隨時間演變。達里奧曾是 OpenAI 的研究副總裁,現在已轉投
8:54.240–9:00.360
to Anthropic, who are taking a slightlydifferent approach to developing AI,said back in 2017 while he was still at
Anthropic,他們在開發 AI 方面採取了略有不同的方法,他在 2017 年仍於
9:00.360–9:08.720
OpenAI, that I think at the extreme endis the Nick Bostrom style of fear thatan AGI could destroy humanity. I can't
在OpenAI時,我認為在最極端的情況下,是尼克·波斯特羅姆式的恐懼,認為AGI可能會毀滅人類。我無法
9:08.720–9:16.160
see any reason in principle why thatcouldn't happen. My chance thatsomething goes really quitecatastrophically wrong on the scale of
從原則上看出為什麼這不可能發生。我認為在人類文明規模上出現非常災難性錯誤的可能性,
9:16.160–9:26.760
human civilizationmight be somewhere between 10% and 25%.>> 10 to 25%. Read that number again.That's not a fringe blogger. That's a
可能在10%到25%之間。>> 10%到25%。再讀一次這個數字。這不是一個邊緣部落客。這是一位
9:26.760–9:35.360
man now running one of the most fundedAI labs on Earth.I thought the empire argument was themost shocking part of this conversation.
現在掌管地球上資金最豐厚的AI實驗室之一的男人。我原以為關於帝國論述的部分是這場對話中最令人震驚的。
9:35.360–9:42.120
I was wrong. The next section about asmall community in Tennessee is the parthat actually stayed with me overnight.
我錯了。關於田納西州一個小社區的下一段,才是真正讓我過夜難眠的部分。
9:42.120–9:51.040
We talk about AI like it lives in thecloud, like it's weightless. It is not.>> OpenAI's largest, one of its largest
我們談論AI時,彷彿它住在雲端,彷彿它沒有重量。事實並非如此。>> OpenAI最大、其中一個最大的
9:51.040–9:56.800
data center projects, is being built inAbilene, Texas as part of the Stargateinitiative, which was an effort
數據中心項目,正在德克薩斯州阿比林建造,作為星際計劃的一部分,該計劃是在
9:56.800–10:03.840
announced at the beginning of Trump'second administration to spend $500billion on AI computing infrastructure.
特朗普第二任期開始時宣布的,旨在花費5000億美元用於AI計算基礎設施。
10:03.840–10:14.760
That one would be the size of CentralPark.And it would run a million computerchips, and it wouldrequire the power of more than 20% of
這個數據中心將擁有中央公園的大小。它將運行一百萬個計算芯片,並且需要超過紐約市
10:14.760–10:22.640
New York City.>> Then there's a meta facility being builtin Louisiana that's reportedly fourtimes the size of that Texas one, using
20%的電力。>> 然後,路易斯安那州正在建設一個元設施,據報導其規模是德克薩斯州那個的四倍,使用
10:22.640–10:31.000
about half the average power demand ofthe entire city of New York.Karen talks about Memphis, Tennessee,where Elon Musk built Colossus, the
大約相當於紐約市整體平均電力需求的一半。凱倫談到了田納西州的孟菲斯,埃隆·馬斯克在那裡建造了科索斯,
10:31.000–10:38.200
supercomputer used to train Grok,powered partly by 35 methane gasturbines.The community living there, a working
用於訓練Grok的超級計算機,部分由35台甲烷燃氣渦輪機供電。生活在該社區的居民,據報是工人階層,
10:38.200–10:44.880
class, reportedly wasn't formally toldthey'd be hosting this facility.According to Karen, residents first
並未正式被告知他們將 hosting 這個設施。根據凱倫的說法,居民們首先
10:44.880–10:51.920
realized something had changed when theysmelled what seemed like a gas leakinside their own homes.That community, she says, already has
意識到有些東西改變了,當他們在家裡聞到似乎像是氣體洩漏的味道時。她說,這個社區已經
10:51.920–10:58.440
ome of the highest rates of asthma andrespiratory illness in the area.Now add industrial-scale gas turbines
擁有該地區最高比例的哮喘和呼吸道疾病率之一。現在加上工業規模的燃氣渦輪機
10:58.440–11:05.520
running next door, 24 hours a day, topower a machine most of those residentswill never directly benefit from.
在隔壁運行,24小時不間斷,為一台大多數居民永遠無法直接受益的機器供電。
11:05.520–11:12.160
And in a cruel twist, the same wave ofautomation this facility helps build isalso the thing most likely to replace
並且在一個諷刺的轉折中,這個設施幫助構建的自動化浪潮,也是最可能取代
11:12.160–11:22.160
the jobs many of them do.I want to be fair here, because I thinkit matters. AI data centers also createconstruction jobs, local tax revenue,
他們中許多人工作的東西。我想在這裡保持公平,因為我認為這很重要。AI數據中心也創造了
11:22.160–11:29.160
and in some cases genuinely usefultechnology for medicine and science.This isn't a simple villain story.
建築工作、地方稅收,在某些情況下,確實為醫學和科學使用了有用的技術。這並不是一個簡單的邪惡故事。
11:29.160–11:34.240
Karen says these companies pick some ofthe most vulnerable communities forthese projects. And when you look at
凱倫說,這些公司選擇了一些最脆弱的社區來進行這些項目。當你查看
11:34.240–11:40.920
where facilities like this actuallyland, it's hard to call that acoincidence, either.Now, here's the part where you might
此類設施實際上落腳的地方時,很難稱之為巧合。現在,這裡是你可能會
11:40.920–11:48.760
expect me to just tell you to log offand boycott AI forever.That's not what Karen says. Her answeris more interesting and honestly more
期望我告訴你不要登入並永遠抵制AI的部分。這不是凱倫所說的。她的答案更有趣,也誠實地說更
11:48.760–11:57.160
useful.Karen's suggestion isn't delete every AIapp. It's smarter than that. She callsall of us data donors, meaning every
有用。凱倫的建議不是刪除每個AI應用。這比那更聰明。她稱我們所有人為數據捐贈者,意味著每次
11:57.160–12:03.800
time we use these tools, we're feedingthe very systems these companies need tokeep growing. Her point, you have more
我們使用這些工具時,我們都在餵養這些公司需要繼續增長的系統。她的觀點是,你擁有比你想像中更多
12:03.800–12:09.440
leverage than you think. She lists a fewconcrete things, and I want to walkthrough them simply. Withhold what you
的槓桿。她列出了幾件具體的事情,我想簡單地 walkthrough 它們。盡你所能 withhold(保留/不提供)。藝術家和作家起訴這些
12:09.440–12:15.600
can. Artists and writers suing thesecompanies over their work are, in herwords, trying to withhold the resource
公司關於他們的作品,在她看來,試圖 withhold(保留/不提供)這些公司最需要的資源,
12:15.600–12:23.640
these companies need most, originalhuman data. Get loud locally.>> And that's what we've seen with all ofthese people that are now literally
原始的真人數據。在當地大聲疾呼。>> 這就是我們看到的所有這些人現在實際上
12:23.640–12:29.280
protesting the streets against datacenters and having an enormous effect,by the way, actually stalling data
在街頭抗議數據中心,並且以某種方式產生了巨大的影響,實際上阻滯了數據
12:29.280–12:36.320
center projects and also completelybanning data centers from beingdeveloped in their localities.>> She points out that data centers are
中心的項目,並且完全禁止在他們當地開發數據中心。>> 她指出,數據中心
12:36.320–12:42.760
popping up in ordinary towns right now,and in several places, local communitieshave already succeeded in stalling or
正在普通城鎮湧現,在幾個地方,當地社區已經成功通過公眾壓力阻滯或
12:42.760–12:50.280
outright banning new data centerprojects through public pressure. Pushat your school or workplace.If your company or university is
outright(完全)禁止新的數據中心項目。在你學校或工作場所施加壓力。如果你的公司或大學
12:50.280–12:56.720
drafting an AI adoption policy, thatconversation is one small piece of amuch bigger puzzle, and your voice in
正在起草AI採用政策,那場對話只是更大拼圖中的一小部分,而你在那個房間裡的聲音確實
12:56.720–13:05.200
that room genuinely counts. Supportalternatives.Karen says there's research showing thesame useful AI capabilities can be built
有分量。支持替代方案。凱倫說,有研究表明,同樣有用的AI功能可以用
13:05.200–13:11.800
with far more efficient methods and farless resource consumption, what shecalls the bicycles of AI as opposed to
更高效的方法和遠少資源消耗來構建,她稱之為AI的自行車,而不是
13:11.800–13:20.200
the resource-hungry rockets.She also cites a striking number.>> 80% of Americans in the most recent pollthink that the AI industry need to be
資源飢渴的火箭。她還引用了一個引人注目的數字。>> 在最近的一次民意調查中,80%的美國人認為
13:20.200–13:25.790
regulated.>> Yeah, I see.>> When was the last time that 80% ofAmericans were on the same side of anissue?
AI行業需要監管。>> 是的,我看到了。>> 上一次80%的美國人在同一立場上是什么时候?
13:25.800–13:31.520
>> A shift she says happened largelybecause of stories like the ones in herbook reaching the public. I want to end
>> 她說,這種轉變很大程度上是因為像她書中那樣的故事情節被公眾所知。我想以公平的方式結束
13:31.520–13:38.200
this fairly because Karen herself does.Late in the conversation, Steven pushesback a little saying that as an
這場對話,因為凱倫自己也這樣做。在對話後期,史蒂文稍微反駁說,作為一名
13:38.200–13:45.280
entrepreneur and investor, he genuinelyfinds AI incredible. Safer driving,tools that help him create things he
企業家和投資者,他真心認為AI令人驚嘆。更安全的駕駛,幫助他創造以前無法創造的東西的工具。
13:45.280–13:52.400
couldn't before. Real value for over abillion users. And Karen doesn't dismissthat. Her answer is basically, you can
對於超過十億用戶的真實價值。而凱倫並沒有否定這一點。她的答案基本上是,你可以
13:52.400–14:01.320
hold both truths at once.AI can be genuinely useful and thecurrent way it's being built and rushedto market can be genuinely harmful.
同時接受這兩種事實。AI 確實能帶來實質幫助,而目前它被倉促建構並推向市場的方式,也可能造成真正的傷害。
14:01.320–14:07.120
Those aren't contradictions. They're twohonest observations about the sameindustry. That's actually the most
這並非矛盾,而是對同一產業的兩個誠實觀察。這其實是整個兩小時對話中最具平衡感的一刻。
14:07.120–14:14.640
balanced moment in the whole two hours.And it's the note I want to end on, too.So, here's my question for you and I
這也是我想帶出的結語。所以,這裡有個問題想問你,我是真心想了解。
14:14.640–14:20.280
genuinely want to know. Knowing what younow know about where the water for yourchatbot's training run might be coming
現在你已經知道,你的聊天機器人訓練用水可能來自哪裡,這會改變你明天使用這些工具的方式嗎?
14:20.280–14:24.760
from, does it change how you'll usethese tools tomorrow? Not never. I'm not
不是說永遠不再使用。我不是要你發誓遠離科技。
14:24.760–14:32.520
asking you to swear off technology. Justpay attention. Ask questions.Support the version of AI that doesn't
只是要你保持關注、提出問題,並支持那種不需要田納西州某個小鎮的居民在客廳裡聞到氣味,才能讓它運作的 AI 版本。
14:32.520–14:40.000
need a small town in Tennessee to smellgas in their living room to make itwork. If this made you thinkdifferently, even a little, share it
如果這讓你產生了不同的想法,哪怕只有一點點,請把它分享給一個每天使用 AI 卻從未問過這些問題的人。
14:40.000–14:46.200
with one person who uses AI every daywithout ever asking these questions.That's genuinely how change like this
與每天使用 AI 卻從未提出這些問題的一個人在一起。這確實是像這樣的改變
14:46.200–14:48.800
tarts.
起頭。

影片筆記:AI Whistleblower WARNS: "We Are Not Prepared For 2027"

<!-- VIDEONOTE_PUBLICATION_WARNINGS_BEGIN -->

⚠️ 可先使用,仍有待修正項目

本筆記已完成可閱讀的逐字稿與時間跳轉;下列為待處理問題,建議播放原片確認。

  • 待修正:逐字稿全文快速校讀未完成(已保留自動字幕/原始逐字稿並繼續生成筆記。;未完成的全文速讀項目可在精進模式回聽確認。)

<!-- VIDEONOTE_PUBLICATION_WARNINGS_END -->

一句話總結

影片透過分析記者 Karen Hao 與 Steven Bartlett 的訪談,批判當前 AI 產業以「取代人類」為目標的開發模式,揭露大型科技公司透過掠奪資源、剝削勞工與控制知識建立「AI 帝國」,並呼籲用戶與社區從被動抵制轉向主動賦權,要求更透明、負責任且不干擾社區生活的 AI 發展。

核心重點

  • AI 發展目標的哲學質疑:技術應旨在提升人類福祉(human flourishing),而非單純取代人類或自動化人類工作。當前 AI 產業缺乏對「通用人工智慧」(AGI)的科學共識,其定義隨受眾(國會、消費者、投資者)隨意變動。
  • 「AI 帝國」的四個特徵:大型 AI 公司被比作舊時代帝國,具備掠奪資源(個人資料、智慧財產權、土地)、勞工剝削、控制知識(Gaslighting 公眾、審查內部研究人員)以及製造神話(宣稱只有它們能安全引導 AI)的特徵。
  • 實體社區的負擔:以田納西州 MEMPHIS 為例,數據中心(如訓練 Grok 的 Colossus)對當地工人階級社區造成環境負擔(空氣品質、哮喘率上升),且當地居民未獲充分告知,同時面臨工作被自動化的風險。
  • 用戶身份轉變:用戶應意識到自己作為「數據捐贈者」(data donors)的身份,每次使用工具都在餵養系統。用戶擁有比想像中更多的槓桿力,不應僅止於刪除應用程式,而應採取更聰明的應對策略。
  • 平衡觀點與監管需求:承認 AI 具有實用價值(如安全駕駛、創作輔助),但當前匆忙推向市場的方式對環境與社區有害。民調顯示 80% 美國人支持監管 AI 產業,公眾壓力已能影響數據中心的開發進程。

詳細大綱

1. AGI 定義的流動性與策略性

  • 歷史背景:1956 年達特茅斯學院(Dartmouth University)的 John McCarthy 命名該學科為「人工智慧」,旨在重現人類智慧。
  • 科學共識缺失:生物學、心理學或神經學中對「人類智慧」無單一定義,因此 AGI 沒有固定的終點線。
  • OpenAI 的定義隨對象變動:
  • 對國會:能治愈癌症、解決氣候變遷、消除貧窮。
  • 對消費者:最驚人的數位助手。
  • 對微軟(投資合約):產生 1000 億美元營收的系統。
  • 對官網:在大多數經濟價值工作中超越人類的系統。
  • 結論:這是一種策略,根據需要說服的不同受眾重新定義 AGI。

2. 存在性風險與早期警告

  • Sam Altman 於 2015 年(OpenAI 正式成立前)的部落格文章指出,超級人類機器智慧可能是人類持續存在最大的威脅。
  • 此言論與當時 Elon Musk 的說法(將先進 AI 比作召喚惡魔)相呼應。
  • Musk 最終聯合創立 OpenAI 後離開,並表示感覺被利用,雙方目前互相起訴。
  • Dario(前 OpenAI 研究副總裁,現 Anthropic 負責人)於 2017 年估計 AGI 導致文明級別災難的機率在 10% 到 25% 之間。

3. 「AI 帝國」理論(Karen Hao 的核心論點)

  • 將大型 AI 公司比作舊時代的帝國,具備四個特徵:
  1. 掠奪資源:佔有普通人個人資料、藝術家/作家的智慧財產權、用於建設大型計算設施的土地。
  2. 勞工剝削:全球數十萬名工人(包括美國)受僱建構系統,部署後常剝奪勞工保護。
  3. 控制知識:
  • 散播只有它們才純潔或聰明到能安全引導 AI 的敘事。
  • 對公眾和政策制定者進行「Gaslighting」(心理操縱)。
  • 花費數億美元扼殺公共安全立法。
  • 審查威脅公司利潤的內部研究人員(案例:Google 的 Dr. Timnit Gebru 與 Margaret Mitchell)。
  1. 製造神話:
  • 同時宣稱 AI 既可能是災難性的,也可能帶來烏托邦。
  • 若危險,則只有它們能控制;若烏托邦,則應將信任、數據和金錢交給它們以加速實現。

4. 實體影響:田納西州 MEMPHIS 社區案例

  • 背景:Elon Musk 在 MEMPHIS 建造 Colossus 超級電腦(用於訓練 Grok),由 35 部甲烷燃氣渦輪機供電。
  • 社區狀況:
  • 工人階級社區,未正式被告知將 hosting 該設施。
  • 居民首先透過家中聞到類似天然氣洩漏的味道察覺變化。
  • 該地區已有該區域最高的哮喘和呼吸道疾病率。
  • 矛盾:工業級燃氣渦輪機 24 小時運行,而自動化技術最可能取代當地居民的工作。
  • 公平性說明:AI 數據中心也創造建築工作、地方稅收,並有時用於醫學和科學。但選擇脆弱社區並非巧合。

5. 用戶應對策略:從抵制到賦權

  • Karen 的建議不是刪除所有 AI 應用程式。
  • 概念轉變:用戶是「數據捐贈者」(data donors)。
  • 每次使用工具都在餵養公司成長所需的系統。
  • 用戶擁有比想像中更多的槓桿力(leverage)。

6. 數據中心的社區抗爭與雙面性

  • 數據中心的社區抗爭:
  • 藝術家與作家透過訴訟試圖保留「原始人類數據」資源。
  • 當地社區透過街頭抗議,成功延遲或全面禁止當地數據中心的開發。
  • 數據中心正興起於普通城鎮,公眾壓力產生實質影響。
  • 個人與機構的參與:
  • 學校或工作場所若正在擬定 AI 採用政策,個人的聲音至關重要。
  • 支持更高效的 AI 替代方案。
  • AI 發展的雙面性(平衡觀點):
  • 企業家/投資者觀點:Steven 認為 AI 具備巨大價值(安全駕駛、創作工具、超過十億用戶的真實價值)。
  • Karen 的回應:可以同時持有兩個事實——AI 確實有用,但當前匆忙推向市場的方式確實有害。這不是矛盾,而是對同一產業的兩個誠實觀察。
  • 研究指出,同樣的 AI 能力可用資源消耗更少、效率更高的方法建構(稱為「AI 自行車」 vs. 「耗資巨大的火箭」)。
  • 公眾態度與監管:
  • 最新民調顯示 80% 美國人認為 AI 產業需要監管。
  • Karen 認為這種轉變源於書籍中的故事進入公眾視野。
  • 行動呼籲:
  • 不要求完全拒絕技術,但要求「關注」與「提問」。
  • 支持不需要犧牲小鎮(如田納西州的小鎮)居民生活品質的 AI 版本。
  • 鼓勵將此觀點分享給每天使用 AI 卻未提出這些問題的人。

工具 / 模型 / 名詞整理

  • 機構/公司名稱:
  • OpenAI
  • Microsoft
  • Google
  • Anthropic
  • Dartmouth University(達特茅斯學院)
  • MIT(麻省理工學院)
  • 產品/技術/設施名稱:
  • AGI(Artificial General Intelligence,通用人工智慧)
  • AI(Artificial Intelligence,人工智慧)
  • Stargate initiative(Stargate 計劃,特朗普第二任期初期宣布的 5000 億美元 AI 計算基礎設施努力)
  • Colossus(超級電腦,用於訓練 Grok)
  • Grok(由 Colossus 訓練的模型)
  • Large Language Models(大型語言模型)
  • Chatbot(聊天機器人)
  • Data Centers(數據中心)
  • AI Adoption Policy(AI 採用政策)
  • 人物名稱:
  • Steven Bartlett
  • Karen Hao(記者)
  • John McCarthy(達特茅斯學院助理教授,命名「人工智慧」學科)
  • Sam Altman(OpenAI 負責人)
  • Elon Musk(OpenAI 聯合創始人,現與 OpenAI 互相起訴)
  • Dario(前 OpenAI 研究副總裁,現 Anthropic 負責人)
  • Nick Bostrom(被引用其關於 AGI 毀滅人類的恐懼風格)
  • Dr. Timnit Gebru(Google 前負責道德 AI 團隊的主管)
  • Margaret Mitchell(Google 前共同主管)

操作流程整理

  • 識別數據捐贈者身份:用戶應意識到自己在使用 AI 工具時,實際上是在捐贈數據,餵養公司成長所需的系統。
  • 質疑 AGI 定義:面對不同受眾時,檢查 AI 公司對 AGI 的定義是否隨之改變,識別其策略性話術。
  • 關注社區影響:觀察當地數據中心建設對社區環境(如空氣品質、健康狀況)及勞工權益的影響。
  • 參與公眾抗爭:透過訴訟或街頭抗議,支持藝術家、作家及當地社區保留「原始人類數據」資源及反對未經告知的數據中心建設。
  • 推動監管與高效 AI:支持對 AI 產業進行監管,並倡導開發資源消耗更少、效率更高的「AI 自行車」替代方案,而非僅追求龐大算力。
  • 日常提問與關注:在日常使用 AI 工具時,持續提問其來源、影響及對社區的犧牲,並將此觀點分享給未提出這些問題的人。

值得注意的限制或風險

  • AGI 定義缺乏科學共識:由於對「人類智慧」無單一定義,AGI 沒有固定終點,導致 AI 公司可隨意定義目標以說服不同受眾。
  • 數據中心對脆弱社區的 disproportionately 影響:數據中心常選址於工人階級或呼吸道疾病率高達社區,且居民未獲充分告知,面臨環境與就業雙重風險。
  • 內部研究人員審查:公司可能審查威脅其利潤的內部研究人員(如 Google 案例),導致潛在風險被掩蓋。
  • 用戶槓桿力被低估:用戶可能低估自己作為「數據捐贈者」所擁有的影響力,僅採取刪除應用程式等被動抵制,而非主動賦權。
  • 當前開發模式的環境與社會成本:匆忙將 AI 推向市場的方式對環境與社區造成實質傷害,與 AI 的實用價值形成矛盾。

逐字稿辨識疑點

  • ime:逐字稿中出現 "for the first ime",疑為 "time" 之聽寫錯誤。
  • ystems:逐字稿中出現 "highly autonomousystems",疑為 "autonomous systems" 之聽寫錯誤。
  • hat:逐字稿中出現 "blog post hat said",疑為 "that" 之聽寫錯誤。
  • ide by side:逐字稿中出現 "ide by side",疑為 "side by side" 之聽寫錯誤。
  • ome:逐字稿中出現 "already has ome of the highest rates",疑為 "some" 之聽寫錯誤。
  • usefultechnology:逐字稿中出現 "usefultechnology",疑為 "useful technology" 之連字錯誤。
  • the the industry's quest:逐字稿中出現 "the the industry's quest",疑為重複字詞。
  • you're going to ever going to have:逐字稿中出現 "you're going to ever going to have",疑為語法重複或口誤。
  • Read that again. Same three letters:講者引用 "Same three letters" 指代 AGI,但 AGI 為三個字母,此處語意需查證是否指代特定縮寫或口誤。
  • Trump's second administration:逐字稿提及 "Trump's second administration" 與 "beginning of" 時間點,需查證具體時間背景是否與當前現實或影片發布時間相符。
  • Dario:逐字稿僅稱 "Dario",未提及姓氏,需查證是否指 Anthropic 的 Dario Amodei。
  • 35 methane gas turbines:逐字稿提及 "35 methane gas turbines",需查證 Colossus 設施確切的燃氣渦輪機數量與類型。
  • Stargate initiative:逐字稿稱其為 "effort announced at the beginning of Trump's second administration",需查證 Stargate 計劃的實際宣布時間與背景。
  • "can.":開頭單字,語意不明,疑為語音辨識截斷或誤聽。
  • "Get loud locally.":語意稍顯突兀,疑為聽寫錯誤或特定口號。
  • "Pushat your school or workplace.":疑為 "Push at" 或 "Push at" 的連寫,或應為 "Push at" 相關動詞。
  • "80% of Americans in the most recent poll think that the AI industry need to be regulated.":語法上 "need" 應為 "needs",但依規則保留原樣。
  • "Safer driving,tools that help him create things he couldn't before.":標點符號後缺少空格,依規則保留原樣。
  • "Real value for over abillion users.":"abillion" 疑為 "a billion" 的連寫或辨識錯誤。
  • "That's actually the most balanced moment in the whole two hours.":提及 "two hours",指代對話時長。
  • "does it change how you'll usethese tools tomorrow?":"usethese" 疑為 "use these" 的連寫。
  • "That's genuinely how change like this tarts.":"tarts" 疑為 "starts" 的聽寫錯誤,但依規則標為疑點,不更名為 "starts"。

可延伸追問

  • Karen Hao 的書中具體提出了哪些「AI 自行車」的技術範例或研究數據?
  • 田納西州 MEMPHIS 社區目前是否有具體的法律訴訟或政策改變來應對數據中心的影響?
  • 藝術家與作家提起的訴訟,在保留「原始人類數據」方面取得了哪些具體的法律進展?
  • 80% 美國人支持監管 AI 產業的民調,具體是由哪個機構進行,且監管建議包含哪些具體條款?
  • Dario Amodei 在 Anthropic 期間,對於 AGI 風險的評估是否有更新或具體的技術對策?

生字列表

生字讀音類型中文
duplicative/duː.plɪ.kə.tɪv/重複的;複製的
flourishing/ˈflɜː.rɪ.ʃɪŋ/繁榮;興旺
uncomfortable/ʌnˈkʌm.fər.tə.bəl/令人不舒服的;令人不安的
consensus/kənˈsen.səs/共識
coherent/koʊˈhɪr.ənt/連貫的;一致的
catastrophic/ˌkæt.əˈstrɒf.ɪk/災難性的
utopia/juːˈtoʊ.pi.ə/烏托邦;理想國
vulnerable/ˈvʌl.nər.ə.bəl/脆弱的;易受攻擊的
leverage/ˈliː.vər.ɪdʒ/槓桿作用;影響力
withhold/wɪðˈhoʊld/保留;不提供
autonomous/ɔːˈtɒn.ə.məs/自主的;獨立的
mobilized/ˈmoʊ.bəl.aɪzd/動員起來
paraphrasing/ˈpær.ə.fræ.zɪŋ/改述;意譯
encapsulate/ɪnˈkæp.sjə.leɪt/概括; encapsulate
ruthlessly/ˈruːθ.ləs.li/殘酷地;毫不留情地
coincidence/ˌkoʊ.ɪnˈsɪd.əns/巧合

生字解說

duplicative /duː.plɪ.kə.tɪv/

· C1

意思:重複的;複製的

解說:形容詞,指與某事物相同或重複的內容。在原文中用來形容 AI 系統試圖複製人類的特性,帶有負面語感,暗示缺乏創新或過度模仿。

影片原句
Why are we trying to build AI systems that are duplicative of humans?
為什麼我們要試圖建構與人類重複的 AI 系統?
延伸例句
The report contained duplicative entries that needed to be removed.
這份報告包含了需要刪除的重複條目。

flourishing /ˈflɜː.rɪ.ʃɪŋ/

· C1

意思:繁榮;興旺

解說:名詞,指蓬勃發展或處於最佳狀態。原文用「human flourishing」對比「replace people」,強調技術應促進人類福祉而非取代人類。

影片原句
The purpose of technology throughout history has been to improve human flourishing, not to replace people.
技術的目的在歷史上一直是為了提升人類繁榮,而不是取代人們。
延伸例句
The city's flourishing arts scene attracts tourists from all over.
該城市繁榮的藝術場景吸引了來自各地的遊客。

uncomfortable /ʌnˈkʌm.fər.tə.bəl/

· B2

意思:令人不舒服的;令人不安的

解說:形容詞,指引起心理上的不適或焦慮。原文用來形容那個關於「為什麼要建造取代人類的機器」的問題,因為它挑戰了主流敘事。

影片原句
That might be the most uncomfortable question I've heard about AI this year.
這可能是我今年聽過關於 AI 最不舒服的問題。
延伸例句
The silence in the room was uncomfortable for everyone.
房間裡的沉默讓每個人都不自在。

consensus /kənˈsen.səs/

· C1

意思:共識

解說:名詞,指多數人同意或一致的意見。原文指出科學界對「人類智慧」沒有共識,因此「AGI」的目標是模糊的。

影片原句
There is absolutely no scientific consensus on what human intelligence actually is.
對於人類智慧究竟是什麼,絕對沒有科學共識。
延伸例句
The committee reached a consensus on the new policy.
委員會就新政策達成了共識。

coherent /koʊˈhɪr.ənt/

· C1

意思:連貫的;一致的

解說:形容詞,指邏輯清晰、前後一致。原文批評 AI 公司對 AGI 的定義隨場合變化,缺乏一個連貫的願景。

影片原句
This is like not a coherent vision of one technology.
這就像沒有關於一項技術的連貫願景。
延伸例句
She presented a coherent argument that convinced the jury.
她提出了一個有說服力的連貫論點,說服了陪審團。

catastrophic /ˌkæt.əˈstrɒf.ɪk/

· C1

意思:災難性的

解說:形容詞,指造成巨大破壞或災難的。原文引用 Altman 的話,指出超級 AI 可能帶來災難性後果。

影片原句
That AI could be catastrophic and that it could bring utopia.
那就是 AI 可能是災難性的,同時它也可能帶來烏托邦。
延伸例句
The hurricane caused catastrophic damage to the coastal towns.
颶風對沿海城鎮造成了災難性的破壞。

utopia /juːˈtoʊ.pi.ə/

· C1

意思:烏托邦;理想國

解說:名詞,指一個完美的社會或理想狀態。原文指出大公司同時宣傳 AI 的「災難性」和「烏托邦」潛力,以獲取控制權和資源。

影片原句
That AI could be catastrophic and that it could bring utopia.
那就是 AI 可能是災難性的,同時它也可能帶來烏托邦。
延伸例句
The community tried to build a utopia based on equality.
這個社區試圖建立一個基於平等的烏托邦。

vulnerable /ˈvʌl.nər.ə.bəl/

· B2

意思:脆弱的;易受攻擊的

解說:形容詞,指容易受到傷害或影響的。原文用來形容那些被選中建設數據中心的社區,它們往往缺乏政治或經濟權力。

影片原句
Karen says these companies pick some of the most vulnerable communities for these projects.
凱倫說,這些公司選擇了一些最脆弱的社區來進行這些項目。
延伸例句
Elderly people are often more vulnerable to scams.
老年人通常更容易受到詐騙的侵害。

leverage /ˈliː.vər.ɪdʒ/

· C1

意思:槓桿作用;影響力

解說:名詞,指利用某種優勢或資源來達成目標的能力。原文鼓勵用戶意識到自己作為「數據捐贈者」所擁有的影響力。

影片原句
Her point, you have more leverage than you think.
她的觀點是,你擁有比你想像中更多的槓桿(影響力)。
延伸例句
Negotiators used their market position as leverage.
談判者利用他們的市场地位作為槓桿。

withhold /wɪðˈhoʊld/

· C1

意思:保留;不提供

解說:動詞,指故意不給予或分享某物。原文建議用戶和創作者「withhold」他們的數據和作品,以減少對 AI 公司的供養。

影片原句
Withhold what you can.
盡你所能 withhold(保留/不提供)。
延伸例句
The witness decided to withhold information from the police.
證人決定向警方保留信息。

autonomous /ɔːˈtɒn.ə.məs/

· C1

意思:自主的;獨立的

解說:形容詞,指能夠獨立運作而不受外部控制的。原文引用 OpenAI 網站定義,指 AI 系統在經濟價值工作中表現優於人類。

影片原句
They define it as highly autonomous systems that outperform humans in most economically valuable work.
他們將其定義為在大多數經濟上有價值的工作中表現優於人類的、高度自主的系統。
延伸例句
The company is moving towards more autonomous decision-making.
公司正轉向更自主的決策方式。

mobilized /ˈmoʊ.bəl.aɪzd/

· C1

意思:動員起來

解說:動詞過去分詞,指組織或動員人群以達成某個目標。原文指出這些定義是為了動員特定受眾(如消費者或監管者)以支持行業利益。

影片原句
These are very different definitions that are spoken out loud to the audience that needs to be mobilized to ward off regulation or get more consumer buy-in to the industry's quest.
這些定義是針對需要動員起來以抵禦監管或獲得更多消費者支持的受眾說出口的。
延伸例句
The community was quickly mobilized to help with the relief efforts.
社區迅速動員起來協助救援工作。

paraphrasing /ˈpær.ə.fræ.zɪŋ/

· C1

意思:改述;意譯

解說:名詞,指用自己的話重述別人的意思,通常為了簡化或澄清。原文作者表示他在轉述 Altman 的話時進行了謹慎的改述。

影片原句
I'm paraphrasing carefully here, that a superhuman AI was probably the single biggest threat to humanity's survival.
這裡我謹慎地轉述,「超人類的 AI 可能是人類生存的最大威脅」。
延伸例句
His paraphrase of the poem lost some of its original emotion.
他對詩歌的改述失去了一些原本的情感。

encapsulate /ɪnˈkæp.sjə.leɪt/

· C1

意思:概括; encapsulate

解說:動詞,指將某事物的本質或特徵包含在一個概念或比喻中。原文作者認為「帝國」這個隱喻完全概括了這些公司的行為和規模。

影片原句
Empire is the only metaphor that I've ever found to fully encapsulate all of the dimensions of what these companies do.
「帝國」是我找到的唯一一個能完全概括這些公司所做的一切的隱喻。
延伸例句
The book encapsulates the history of the 20th century.
這本書概括了 20 世紀的歷史。

ruthlessly /ˈruːθ.ləs.li/

· C1

意思:殘酷地;毫不留情地

解說:副詞,指沒有同情心或猶豫地。原文用來形容大公司對內部研究人員的審查和打壓。

影片原句
They spend hundreds of millions of dollars to kill public safety legislation and ruthlessly censor any internal researchers whose findings threaten their corporate profits.
它們花費數億美元來扼殺公共安全立法,並冷酷地審查任何可能威脅其企業利潤的內部研究人員。
延伸例句
The dictator ruthlessly eliminated his political opponents.
獨裁者殘酷地消除了他的政治對手。

coincidence /ˌkoʊ.ɪnˈsɪd.əns/

· B2

意思:巧合

解說:名詞,指兩件或多件事同時發生但無因果關係。原文作者認為 AGI 定義隨場合變化並非巧合,而是一種策略。

影片原句
I don't think that's a coincidence. I think that's a strategy.
我不認為這是巧合。我認為這是一種策略。
延伸例句
It was just a coincidence that we met at the same cafe.
我們在同一間咖啡館相遇只是巧合。

句型解說(含實例)

Why are we trying to build [X] that is [Y]?

意思:為什麼我們要試圖建構 [X],而它是 [Y]?

接續:Why + are + we + trying to build + [Object] + that is + [Adjective/Participle]?

解說:用於質疑某種技術或行動的根本目的,特別是當該技術或行動具有爭議性或負面後果時。強調對「目的」與「手段」之間關係的反思。

影片原句
Why are we trying to build AI systems that are duplicative of humans?
為什麼我們要試圖建構與人類重複的 AI 系統?
實例
  1. Why are we trying to build cities that are isolating us from nature?
    為什麼我們要試圖建構將我們與自然隔離的城市?
  2. Why are we trying to build algorithms that are biased against minorities?
    為什麼我們要試圖建構對少數族裔有偏見的算法?

Not [A], not [B], but [C]

意思:不是 [A],不是 [B],而是 [C]

接續:Not + [Noun Phrase A], not + [Noun Phrase B], but + [Noun Phrase C]

解說:用於排除常見的誤解或次要問題,並強調真正核心或更簡單的問題。這種結構具有強烈的對比和聚焦效果。

影片原句
Not can AI become smarter than humans, not will AI destroy humanity, but something much simpler.
不是 AI 會不會比人類更聰明,不僅是人工智慧會毀滅人類,而是更簡單的問題。
實例
  1. Not the price, not the brand, but the quality matters most.
    不是價格,不是品牌,而是品質最重要。
  2. Not the speed, not the size, but the efficiency is what we should focus on.
    不是速度,不是大小,而是效率是我們應該關注的。

It is [Adjective] that [Clause]

意思:正是 [Clause] 使得 [Adjective]

接續:It + is + [Adjective] + that + [Subject] + [Verb]

解說:強調句型,用於突出某個特定因素或情況的重要性或影響。原文用來強調田納西州小鎮的案例給作者帶來的深層不安。

影片原句
And that's exactly why I wanted to react to this podcast.
這正是我想要回應這個播客的原因。
實例
  1. It is the lack of transparency that makes the process confusing.
    正是缺乏透明度讓這個過程令人困惑。
  2. It is the sudden change in policy that caused the market crash.
    正是政策的突然變化導致了市場崩盤。

There is no [Noun] on [Topic]

意思:在 [Topic] 上沒有 [Noun]

接續:There + is + no + [Noun] + on + [Topic]

解說:用於指出某個領域缺乏共識、定義或標準。原文用來強調「人類智慧」沒有科學定義,因此「AGI」的目標是模糊的。

影片原句
There is absolutely no scientific consensus on what human intelligence actually is.
對於人類智慧究竟是什麼,絕對沒有科學共識。
實例
  1. There is no clear definition of success in the modern workplace.
    在現代職場中沒有對成功的明確定義。
  2. There is no universal standard for measuring happiness.
    沒有衡量幸福的普遍標準。

If [Condition], [Result]

意思:如果 [Condition],[Result]

接續:If + [Condition Clause], + [Result Clause]

解說:條件句,用於闡述在特定條件下會發生的結果。原文用來探討如果我們不知道如何定義目的地,就無法到達那裡。

影片原句
And when are we going to get there if we don't know how to define the destination?
並且當我們不知道如何定義目的地時,我們何時才能到達那裡?
實例
  1. If we don't invest in education, we will face a skilled labor shortage.
    如果我們不投資教育,我們將面臨熟練勞動力短缺。
  2. If the company doesn't adapt to market changes, it will lose its competitive edge.
    如果公司不適應市場變化,它將失去競爭優勢。

Not [A], but [B]

意思:不是 [A],而是 [B]

接續:Not + [Noun/Phrase A], but + [Noun/Phrase B]

解說:用於否定一個觀點並提出另一個對立的觀點。原文用來區分 AI 的「有用性」與「建構方式的危害性」。

影片原句
Those aren't contradictions. They're two honest observations about the same industry.
這並非矛盾,而是對同一產業的兩個誠實觀察。
實例
  1. It's not a bug, but a feature.
    這不是錯誤,而是一個功能。
  2. It's not about winning, but about participating.
    這無關勝負,而在於參與。

The [Noun] that [Relative Clause]

意思:那個 [Relative Clause] 的 [Noun]

接續:The + [Noun] + that + [Relative Clause]

解說:定語從句結構,用於精確描述某個名詞的特徵。原文用來定義 AGI 在不同場合下的具體含義。

影片原句
When Microsoft was writing the investment contract, AGI was legally defined as a system that generates $100 billion in revenue.
當微軟撰寫投資合約時,AGI 在法律上被定義為產生 1000 億美元收入的系統。
實例
  1. The tool that helps us organize our data is essential.
    幫助我們組織數據的工具是必不可少的。
  2. The policy that affects our daily lives needs to be revised.
    影響我們日常生活的政策需要修訂。

It is [Adjective] to [Verb]

意思:做 [Verb] 是 [Adjective] 的

接續:It + is + [Adjective] + to + [Verb]

解說:形式主語結構,用於表達做某事的性質或評價。原文用來表達在田納西州小鎮的居民聞到氣味才能讓 AI 運作是不公平的。

影片原句
Just pay attention. Ask questions. Support the version of AI that doesn't need a small town in Tennessee to smell gas in their living room to make it work.
只是要你保持關注、提出問題,並支持那種不需要田納西州某個小鎮的居民在客廳裡聞到氣味,才能讓它運作的 AI 版本。
實例
  1. It is crucial to verify the source of information.
    驗證信息來源至關重要。
  2. It is important to maintain a healthy work-life balance.
    保持健康的工作與生活平衡很重要。