“Empire of AI”: Karen Hao on How AI Is Threatening Democracy & Creating a New Colonial World

Democracy Now! 首頁 每日列表 原始影片
實際影片長度：56:15.000。原文、繁中、雙語可點擊句子跳轉影片。
0:03.200–0:12.320
This is democracyow [music]democracynow.orgthe War and Peace Report. I'm AmyGoodman. The Empire of AI, that's the
0:12.320–0:20.320
name of a new book by journalist KarenHow, who's been closely reporting on therise of the artificial intelligence
0:20.320–0:31.359
industry with a focus on Sam Alman'sOpen AI. That's the company behind ChatGPT. Karen how compares the actions of
0:31.359–0:40.320
the AI industry to those of colonialpowers of the past. She writes, quote,"The empires of AI are not engaged in
0:40.320–0:48.399
the same overt violence and brutalitythat marked this history, but they tooseize and extract precious resources to
0:48.399–0:56.879
feed their vision of artificialintelligence. the work of artists andwriters, the data of countlessindividuals posting about their
0:56.879–1:05.920
experiences and observations online, theland, energy, and water required tohouse and run massive data centers and
1:05.920–1:15.520
supercomputers. She writes, "Over thepast year, the Trump administration hasincreasingly embraced the AI industry.
1:15.520–1:22.880
In December, Trump signed an executiveorder to bar states and localgovernments from enacting their own AI
1:22.880–1:32.000
regulations. Soon after he signed theorder, his family's company, Trump Mediand Technology, announced a $6 billion
1:32.000–1:40.880
merger with a firm aiming to build theworld's first viable nuclear fusionplant to power AI projects. Karen How is
1:40.880–1:49.040
a former reporter at the Wall StreetJournal and MIT Technology Review whereshe became the first journalist to
1:49.040–1:58.159
profile open AI. Democracy Now's JuanGonzalez and I spoke to her in May. TheNational Book Critic Circle recently
1:58.159–2:08.000
named her book The Empire of AI: Dreamsand Nightmares and Sam Alman's Open AIas a finalist for best non-fiction book
2:08.000–2:20.160
of 2025.I began by asking Karen How to explainjust what artificial intelligence is. SoAI is a collection of many different
2:20.160–2:27.680
echnologies but most people wereintroduced to it through chatbt and whatI argue in the book and what the title
2:27.680–2:34.160
refers to empire of AI it's actually acritique of the specific trajectory ofAI development that led us to chatbt and
2:34.160–2:42.720
has continued since chatbt and that ispecifically Silicon Valley's scale atall costs approach to AI development AI
2:42.720–2:50.480
models in modern day they are trained ondata they need computers to train themon that data. But what Silicon Valley
2:50.480–2:57.760
did and what OpenAI did in the last fewyears is they started blowing up theamount of data and the the size of the
2:57.760–3:03.840
computers that need to do this training.So we are talking about the full Englishlanguage internet being fed into these
3:03.840–3:10.720
models, books, scientific articles, allof the intellectual property that isbeing created and also massive
3:10.720–3:17.040
supercomputers that run tens ofthousands even hundreds of thousands ofcomputer chips that are the size of
3:17.040–3:25.360
dozens maybe hundreds of football fieldsand use practically the entire energydemands of cities now. So this is an
3:25.360–3:31.840
extraordinary um type of AI developmenthat is causing a lot of social, laborand environmental harms and that is
3:31.840–3:40.480
ultimately why I evoke this analogy toempire.>> And Karen, could you talk some moreabout not only the energy requirements
3:40.480–3:50.000
but the water requirements of these hugedata centers that are essence in essencethe backbone of of this uh widening uh
3:50.000–3:51.030
industry?
3:51.040–3:56.560
>> Absolutely. I'll give you two stats onboth the energy and the water. Whentalking about the energy demand,
3:56.560–4:03.920
McKenzie recently came out with a reporthat said in the next 5 years based onthe current pace of AI computational
4:03.920–4:13.040
infrastructure expansion, we would needto put as much energy on the global gridas what is consumed by two to six times
4:13.040–4:22.479
the energy consumed annually by thestate of California. And that willmostly be serviced by fossil fuels.We're already seeing reporting of uh
4:22.479–4:29.120
coal plants with their lives beingextended. They were supposed to retirebut now they cannot to support this data
4:29.120–4:36.800
center development. We are seeingmethane gas turbines, unlicensed onesbeing popped up to uh service these data
4:36.800–4:44.080
centers as well. From a freshwaterperspective, these data centers need tobe trained on freshwater. They cannot be
4:44.080–4:50.000
trained on any other type of waterbecause it can corrode the equipment. Itcan lead to bacterial growth and most of
4:50.000–4:58.080
the time it actually taps directly intoa public drinking water supply becausethat is the infrastructure that has been
4:58.080–5:05.919
laid to deliver this clean fresh waterto different businesses to differenthomes. And Bloomberg recently had an
5:05.919–5:13.680
alysis where they looked at thexpansion of these data centers aroundthe world and 2thirds of them are being
5:13.680–5:21.840
placed in water scarce areas. So they'rebeing placed in communities that do nothave access to fresh water. So it's not
5:21.840–5:27.520
just the total amount of fresh waterthat we need to be concerned about, butactually the distribution of this
5:27.520–5:37.840
infrastructure around the world.And most people are familiar with chatGPT the consumer aspect of AI but what
5:37.840–5:45.600
about the military uh aspect of AI wherein essence uh we're finding SiliconValley companies becoming the next
5:45.600–5:53.199
generation of defense contractors.>> One of the reasons why OpenAI and manyother companies are turning to the
5:53.199–5:59.600
defense industry is because they havespent an extraordinary amount of moneyin developing these technologies.
5:59.600–6:07.520
They're spending hundreds of billions totrain these models and they need torecoup those costs and there are only so
6:07.520–6:15.360
many industries and so many places thathave that size of a paycheck to pay. Andso that's why we're seeing a cozying up
6:15.360–6:22.720
to the defense industry. We're alsoseeing Silicon Valley use the USgovernment in their empire buildingambitions. You could argue that the US
6:22.720–6:29.039
government is also trying to use SiliconValley vice versa in their empirebuilding ambitions. Um but certainly
6:29.039–6:39.039
these technologies are not they are notdesigned to be used in a sensitivemilitary context. And so the theaggressive push of these companies to
6:39.039–6:44.800
try and get those defense contracts andintegrate their technologies more andmore to the into the infrastructure of
6:44.800–6:52.560
the military is really alarming. Iwanted to go to the countries you wento or the stories you covered because I
6:52.560–7:00.240
mean this is amazing the depth of youreporting from Kenya to Uruguay toChile. Um you were talking about the use
7:00.240–7:08.240
of water and I also want to ask youabout nuclear power. Uh but in Chile umwhat is happening there around these
7:08.240–7:15.520
data centers and the water they woulduse and the resistance to that.>> Yeah. So Chile has an interestinghistory in that it's been under it was
7:15.520–7:25.440
under a dictatorship for a very longtime. And so during that time, mostpublic resources were privatized,including water. But because of an
7:25.440–7:31.440
omaly, there's one community in thegreater Santiago metropolitan regionthat actually still has access to a
7:31.440–7:38.960
public freshwater resource that servicesboth that community as well as the restof the country in emergency situations.
7:38.960–7:48.479
That is the exact community that Googlechose to try to put a data center in>> and it would be free>> and it, you know, I have no idea. That
7:48.479–7:56.319
is a great question. But what thecommunity told me was they weren't evenpaying taxes for this because they they
7:56.319–8:01.520
believed based on reading thedocumentation that the taxes that Googlewas paying was in fact to where they had
8:01.520–8:08.479
registered their offices, theiradministrative offices, not where theywere putting down the data center. So
8:08.479–8:14.240
they were not seeing any benefit fromthis data center directly to thatcommunity and they were seeing no checks
8:14.240–8:20.960
placed on the fresh water that this datacenter would have been allowed toextract. And so these activists said,
8:20.960–8:27.599
"Wait a minute, absolutely not. We'renot going to allow this data center tocome in unless they give us a legitimate
8:27.599–8:35.440
reason for why it benefits us." And sothey started doing boots on the groundactivism, pushing back, knocking on
8:35.440–8:40.399
every single one of their neighborsdoors, handing out flyers to thecommunity, telling them this company is
8:40.399–8:46.800
taking our freshwater resources withoutgiving us anything in return. And sothey escalated so dramatically that it
8:46.800–8:51.839
escalated to Google Chile. It escalatedto Google Mountain View, which by theway then sent representatives to Chile
8:51.839–9:01.680
that only spoke English.But then it eventually escalated to theChilean government. And the Chileangovernment now has roundts where they
9:01.680–9:06.160
ask these community residents and thecompany representatives andrepresentatives from the government to
9:06.160–9:12.640
come together to actually discuss how tomake data center development morebeneficial to the community. The
9:12.640–9:19.920
activists say it is the fight is notover. Just because they've been invitedto the table doesn't mean thateverything is suddenly better. They need
9:19.920–9:25.519
to stay vigilant. They need to continuescrutinizing these projects. But thusfar they've been able to block this
9:25.519–9:36.640
project for four to five years and havegained that seat at the table.>> And how is it that these uh westerncompanies in essence are exploiting
9:36.640–9:40.240
labor in the global south? You go into
9:40.240–9:44.640
something called data annotation firms.What what are those?
9:44.640–9:51.279
>> Yeah. So because AI modern day AIsystems are trained on massive amountsof data and they're sc that's scraped
9:51.279–9:59.279
from the internet, you can't actuallypump that data directly into your AImodel because there are a lot of things
9:59.279–10:05.279
within within that data. It's heavilypolluted. It needs to be cleaned. Itneeds to be annotated. So this is where
10:05.279–10:14.160
data annotation firms come in. These aremiddleman firms that hire contract laborto provide to these AI companies to do
10:14.160–10:21.600
that kind of data preparation. And openAI when it was starting to think aboutcommercializing its products and
10:21.600–10:28.880
thinking about let's put text generationmachines that can spew any kind of textinto the hands of millions of users.
10:28.880–10:34.880
They realized they needed to have somekind of content moderation. They neededto develop a filter that would wrap
10:34.880–10:42.320
around these models and prevent thesemodels from actually spewing racist,hateful, and harmful speech to users
10:42.320–10:48.640
that would not make a very goodcommercially viable product. And so theycontracted these middleman firms in
10:48.640–10:56.320
Kenya where the Kenyan workers had toread through reams of the worst text onthe internet as well as AI generated
10:56.320–11:03.600
text where open AI was prompting its ownAI models to imagine the worst text onthe internet and then telling these
11:03.600–11:11.839
Kenyan workers to detail to categorizethem in detailed taxonomies of is thisexual content, is this violent content,
11:11.839–11:18.480
how graphic is that violent content inorder to teach its filter all thedifferent categories of content it had
11:18.480–11:25.120
to block. And this is incrediblyuncommon form of labor. There are lotsof other different types of contract
11:25.120–11:32.480
labor that they use. But these workers,they're paid a few bucks an hour, if atall. And just like the era of social
11:32.480–11:40.079
media, these content moderators are leftvery deeply psychologically traumatized.And ultimately, there is no real
11:40.079–11:46.640
philosophy behind why these workers arepaid a couple bucks an hour and havetheir lives destroyed. And why AI
11:46.640–11:52.240
researchers who also contribute to thesemodels are paid million-dollcompensation packages simply because
11:52.240–12:00.079
they sit in Silicon Valley in OpenAI'soffices. That is the logic of Empire.And that hearkens back to my title,
12:00.079–12:10.320
Empire of AI.>> So, let's go back to your title, Empireof AI. the subtitle dreams andnightmares in Sam Alman's Open AI. So
12:10.320–12:17.839
tell us the story of Sam Alman and whatOpen AI is all about right through tothe deal he just made in the Gulf when
12:17.839–12:27.040
President Trump uh Sam Alman and ElonMusk were there. Alman is very much aproduct of Silicon Valley. His career
12:27.040–12:32.880
was first as a founder of a startup andthen as the president of Y Combinator,which is one of the most famous startup
12:32.880–12:40.800
accelerators in Silicon Valley, and thenthe CEO of OpenAI. And there's nocoincidence that OpenAI ended up
12:40.800–12:47.519
introducing the world to the scale atall cost approach to AI developmentbecause that is the way that Silicon
12:47.519–12:59.839
Valley has operated in the entire timethat Altman came up in it. And so he isa very strategic person. He isincredibly good at telling stories about
12:59.839–13:06.639
he future and painting these sweepingvisions that investors and employeeswant to be a part of. And so early on at
13:06.639–13:14.959
YC, he identified that AI would be oneof the trends that could take off. Andhe was trying to build a portfolio of
13:14.959–13:20.639
different investments and differentinitiatives to place himself in thecenter of various different trends
13:20.639–13:26.160
depending on which one took off. He wasinvesting in quantum computing. He wasinvesting in nuclear fusion. He was
13:26.160–13:33.360
investing in self-driving cars. And hewas developing a fundamental AI researchlab. Ultimately, the AI research lab was
13:33.360–13:43.760
the ones that started acceleratingreally quickly. So he makes himself theCEO of that company. Um and andoriginally he started it as a nonprofit
13:43.760–13:52.800
o try and position it as a counter toforprofit driven incentives in SiliconValley. But within one and a half years,
13:52.800–14:00.240
OpenAI's executives identified that ifthey wanted to be the lead in thispace, they had to go for this scale at
14:00.240–14:05.519
all cost approach and had to should bein quotes. They thought that they had todo this. There are actually many other
14:05.519–14:11.920
ways to develop AI and to have progressin AI that does not take this approach.But once they decided that, they
14:11.920–14:19.440
realized the bottleneck was capital. Itjust so happens Sam Alman is a once in ageneration fundraising talent. He
14:19.440–14:26.880
created this new structure nesting afor-profit arm within the nonprofit tobecome this fundraising vehicle for the
14:26.880–14:33.440
tens of billions and ultimately hundredsof billions that they needed to pursuethe approach that they decided on. And
14:33.440–14:41.600
that is how we ultimately get to presentday OpenAI, which is one of the mostcapitalistic companies in the history of
14:41.600–14:49.600
Silicon Valley, continuing to raisehundreds of billions and and Altman hasjoked even trillions to produce a
14:49.600–14:58.399
technology that ultimately has amiddling economic impact thus far. We'llreturn to our conversation in a minute
14:58.399–15:10.639
with Karen How, author of the new bookEmpire of AI: Dreams and Nightmares inSam Alman's Open AI. Stay with us.
15:10.639–15:14.160
This is Democracy Now! democracynow.org.
15:14.160–15:22.000
I'm Amy Goodman. In this holidayspecial, we continue with the journalistKaren How, author of the new book Empire
15:22.000–15:32.000
of AI: Dreams and Nightmares, and SamAlman's Open AI. Karen came into ourstudio in May when she discussed how AI
15:32.000–15:41.519
will impact workers.One of the things that we have seen isthis technology is already having a hugeimpact on jobs.
15:41.519–15:48.480
Not necessarily because the technologyitself is really capable of replacingjobs, but it is perceived as capable
15:48.480–15:59.600
nough that executives are laying offworkers. And we need more some kind ofmore guard rails to actually prevent
15:59.600–16:10.399
hese companies from continuing to tryand develop labor automatingtechnologies and try to shift them toproducing labor assistive technologies.
16:10.399–16:11.519
What do you mean?
16:11.519–16:20.720
>> So, open AI, their definition of whathey call artificial generalintelligence is highly autonomousystems that outperform humans in most
16:20.720–16:28.160
economically valuable work. So, theyexplicitly state that they are trying toautomate jobs away. I mean, what are
16:28.160–16:30.079
what is economically valuable work? But
16:30.079–16:39.279
he things that people do to get paid.Um but there's this really great bookcalled Power in Progress by MITeconomists Jeron Austin and Simon
16:39.279–16:46.399
Johnson who mention that technologydevelopment all technology revolutionsthey take a labor automating approach
16:46.399–16:54.480
not because of inevitability but becausethe people at the top choose to automatethose jobs away. They choose to design
16:54.480–17:03.839
the technology so that they can sell ito executives and say you can shrinkyour costs by laying off all theseworkers and using our AI services
17:03.839–17:12.400
instead. But in the past, we've seenstudies that for example suggest that ifyou develop an AI tool that a doctor
17:12.400–17:19.280
uses rather than replacing the doctor,you will actually get better health carefor patients. you will get better di
17:19.280–17:25.600
cancer diagnoses. If you develop an AItool that teachers can use rather thanjust an AI tutor that replaces the
17:25.600–17:32.559
teacher, your kids will get bettereducational outcomes. And so that's whatI mean by labor assistive than labor
17:32.559–17:38.720
>> and explain uh what you mean because Ithink a lot of people don't evenunderstand artificial intelligence. And
17:38.720–17:42.720
when you say replace the doctor, whatare you talking about?
17:42.720–17:49.600
>> Right. So these companies they try todevelop a technology that they positionas an everything machine that can do
17:49.600–17:59.919
anything. Um and so they will try to sayou can use this you can talk to chatbtfor therapy. No, you cannot. Chat GBT is
17:59.919–18:05.919
not a licensed therapist. And in fact,these models actually spew lots ofmedical misinformation. And there have
18:05.919–18:14.559
been lots of um examples of actuallyusers being psychologically harmed bythe model because the model will
18:14.559–18:22.480
continue to reinforce um selfharmingbehaviors. And we've even had caseswhere uh children who speak to chatbots
18:22.480–18:29.120
and develop huge emotional relationshipswith these chatbots have actually killedthemselves after using these chatbot
18:29.120–18:34.559
systems. Um but that's what I mean whenthese companies are trying to developlabor automating tools. They're
18:34.559–18:42.480
positioning it as you can now hire thistool instead of hire a worker. I mean,most recently, Sam Alman was speaking at
18:42.480–18:50.080
a conference and said, "We originallysaid that these models were junior levelpartners at a law firm, and now we think
18:50.080–18:59.280
that they can really be more seniorcolleagues at a law firm." What he'saying is don't hire the junior levelpartners, don't hire the senior
18:59.280–19:08.160
colleagues, and just use our AI models.And we are already seeing the careerladder breaking because many different
19:08.160–19:15.280
white collar job uh white collar serviceindustries as well as other industriesare becoming convinced that they do not
19:15.280–19:21.360
need to hire interns. They do not needto hire entry-level positions that theyjust need these AI models and new
19:21.360–19:28.559
college graduates are struggling now tofind job opportunities to help them geta foothold into these industries. So,
19:28.559–19:36.640
you've talked about Sam Alman, and inpart one, we touched on uh who he is,but I'd like you to go more deeply into
19:36.640–19:46.559
what uh who Sam Alman is, how hexploded onto the um US scene,testifying before Congress, actuallywarning about the dangers of AI. So,
19:46.559–19:56.960
that really protected him in a way.>> Um people seeing him as a prophet.That's a P O P. But now, we can talkabout the other kind of prophet, P O Fit
19:56.960–20:05.360
T.um and how open AI was formed. How isopen AI different from AI?
20:05.360–20:15.440
OpenAI is a com I mean it was originallyfounded as a nonprofit as I mentionedand Alman specifically when he was
20:15.440–20:25.120
thinking about how do I make afundamental AI research lab that isgoing to make a big splash he chose tomake it a nonprofit because he
20:25.120–20:33.200
identified that if he could not competeon capital and he was relatively late tothe game Google already had a monopoly
20:33.200–20:41.520
on a lot of top AI research talent athe time if he could not compete oncapital and he could not compete um interms of being a first mover he
20:41.520–20:49.600
needed some other kind of ingredienthere to really recruit talent recruitum public goodwill and establish a name
20:49.600–20:58.159
for open AAI so he identified a missionhe identified let me make this anonprofit and let me give it a reallycompelling mission so the mission of
20:58.159–21:06.880
openai is to ensure artificial generalintelligence benefits all of humanityAnd one of the quotes that I open my
21:06.880–21:16.000
book with is this quote that Sam Almancited himself in 2013 um in his blog. Hewas an avid blogger back in the day
21:16.000–21:23.600
talking about his learnings on businessand strategy and Silicon Valley startuplife. And the quote is successful people
21:23.600–21:31.200
built companies, more successful peoplebuild countries. The most successfulpeople build religions. And then he
21:31.200–21:37.520
reflects on that quote in his blogsaying, "It appears to me that the bestway to build a religion is actually to
21:37.520–21:46.240
build a company.">> And so talk about how Alman was thenforced out of the company and then cameback. And also I just found it so
21:46.240–21:52.480
fascinating that you were able to speakwith so many Open AI workers. Youthought there was a kind of total ban on
21:52.480–21:59.360
you.>> Yes. Yeah. Exactly. So I was the firstjournalist to profile OpenAI. Um, Iembedded within the company for 3 days
21:59.360–22:06.880
in 2019 and then my profile published in2020 for MIT technology review and athe time I identified in the profile
22:06.880–22:13.760
this tension that I was seeing where itwas a nonprofit by name but behind thescenes a lot of the public values that
22:13.760–22:18.960
hey exposed were actually the oppositeof how they operated. So they espousedtransparency but they were highly
22:18.960–22:26.720
secretive. They espousedcollaboriveness. They were highlycompetitive. and they espoused that theyhad no commercial intent, but in fact it
22:26.720–22:32.159
seemed like they had just gotten a $1billion investment from Microsoft. Itseems like they were rapidly going to
22:32.159–22:38.320
develop commercial intent. And so Iwrote that into the profile and OpenAIwas deeply unhappy about it and they
22:38.320–22:44.640
would not refuse to talk to me for 3years. But when I started working on thebook, when I started reaching out to
22:44.640–22:50.880
employees, current and former, Idiscovered that many employees actuallyreally liked the profile and they
22:50.880–22:59.200
specifically wanted to talk to mebecause they thought that I would dojustice to the truth of what hadactually happened within the company and
22:59.200–23:07.520
be able to discover behind what thexecutives mythologized and narrativizedabout this technology [clears throat]
23:07.520–23:13.679
and about the course of this company. Iwould be able to actually get beneathat to the real heart of the matter.
23:13.679–23:21.760
And so um one of the things that youreally have to understand about AIdevelopment today is that there are what
23:21.760–23:29.200
I call quasi religious movements thathave developed within Silicon Valley.The concept of artificial general
23:29.200–23:36.480
intelligence is not one that'scientifically grounded. It is this ideathat we can fundamentally recreate human
23:36.480–23:42.400
intelligence in computers. And this ideahas been around for actually a reallylong time. The field of AI was founded
23:42.400–23:48.320
all the way back in the 1950s and thatwas the original intent of the field.How do we recreate intelligence in
23:48.320–23:50.080
computers? Can machines think? That was
23:50.080–23:58.159
the famous question that Britishmathematician Alan Turing asked. But weto this day do not have scientific
23:58.159–24:05.840
onsensus around even what humanintelligence is. And so to peg an entiresearch field and a technology to the
24:05.840–24:12.640
basis of human intelligence is a verytricky endeavor because there are nogood metrics to assess have we actually
24:12.640–24:22.159
gotten there yet and there's noblueprint to say what should AI looklike and how should it work andultimately who should it serve. And so
24:22.159–24:29.440
when OpenAI took up this mission ofartificial general intelligence, theywere able to essentially shape and mold
24:29.440–24:36.640
what they wanted this technology to bebased on what is most convenient forthem. But when they identified it, it
24:36.640–24:45.279
was at a time when scientists reallylooked down on this term even AGI. Andso they absorbed just a small group of
24:45.279–24:52.159
self-identified AGI believers. This iswhy I call it quasi religious becausethere's no scientific evidence that we
24:52.159–24:59.279
can actually develop AGI. The people whoare strongly con have this strongconviction that they will do it and that
24:59.279–25:05.279
it's going to happen soon. It is justpurely based on belief and they talkabout it as a belief too. But there are
25:05.279–25:11.919
two factions within this belief systemof the AGI religion. There are peoplewho think AGI is going to bring us to
25:11.919–25:20.559
utopia and there are people who thinkAGI is going to destroy all of humanity.Both of them believe that it ispossible. It's coming soon. And
25:20.559–25:27.520
therefore they conclude that they needto be the ones to control the technologyand not democratize it. And this is
25:27.520–25:33.760
ultimately what leads to your questionof what happened when Sam Alman wasfired and rehired through the history of
25:33.760–25:39.840
OpenAI. There's been a lot of clashingbetween the boomers and doomers aboutwho should actually>> the boomers and doomers.
25:39.840–25:47.120
>> The boomers and the doomers. [laughter]>> Those that say it'll bring us theapocalypse. topia boomers and those that
25:47.120–25:54.960
say it'll destroy humanity. The doomersand they have clashed relentlessly andaggressively about how quickly to build
25:54.960–26:04.799
the technology, how quickly to releasethe technology and ultimately Altman isone that he is really good at saying to
26:04.799–26:10.960
people what they need to hear and hewill say different things to differentpeople if he thinks they need to hear
26:10.960–26:14.320
different things. So when I askedboomers, is Altman a boomer? They said
26:14.320–26:19.279
yes. When I asked doomers, is Altman adoomer? They said yes. And I want to
26:19.279–26:28.799
take this up until today to um inJanuary, the Trump administrationannouncing the Stargate project, a $500
26:28.799–26:39.520
billion project to boost AIinfrastructure in the United States.This is Open AI Sam Alman speakingalongside President Trump.
26:39.520–26:45.919
I think this will be the most importantproject of this era and as Masa said forAGI to get built here to create hundreds
26:45.919–26:51.830
of thousands of jobs to create a newindustry centered here. Uh we wouldn'tbe able to do this without you Mr.President.
26:51.840–27:01.679
>> He also there referred to AGI um uhartificial general intelligence. Explainwhat happened here and what this is and
27:01.679–27:14.080
has it actually happened. So Altmanbefore Trump was electedum he already was sensing throughobservation that it was possible that
27:14.080–27:21.279
he administration would shift and thathe would need to start politicking quiteheavily to ingruiate himself to a new
27:21.279–27:30.480
administration.Alman is very strategic. Um he was undera lot of pressure at the time as wellbecause his original co-founder Elon
27:30.480–27:38.240
Musk now has great beef with him. UhMusk feels like Alman used his name andhis money to set up OpenAI and then he
27:38.240–27:46.080
got nothing in return. So Musk had beensuing him, still suing him and suddenlybecame first buddy of the Trump
27:46.080–27:56.559
administration. So Altman basicallycleverly orchestrateda um this announcement where by the waythe the announcement is quite strange
27:56.559–28:03.600
because the Trump President Trump is notit's not the US government giving $500billion. It's private investment coming
28:03.600–28:13.840
into the US um from places like SoftBank>> which is>> uh which is one of the largestinvestment funds um run by Masay Yoshi
28:13.840–28:22.240
Son a Japanese businessman who made alot of his wealth from the previous techera. So, so it's not even the USgovernment that's that's providing this
28:22.240–28:33.520
money.>> And take that right through to now thatGulf trip that um Elon Musk was on, butso was Sam Alman to the fury of Elon
28:33.520–28:45.200
Musk and then a deal was sealed in AbuDhabi.>> Yeah.>> It didn't include Elon Musk but wasabout open AI.>> Exactly. So Altman has continued to try
28:45.200–28:59.919
and use the US government as a way to toget access to more places and uh morepowerful spaces to build out thisempire. And one of the one of the things
28:59.919–29:06.720
because OpenAI's computationalinfrastructure needs are so aggressive.You know, I had an OpenAI employee tell
29:06.720–29:13.360
me we're running out of land and power.So they are running out of resources inthe US which is why they're trying to
29:13.360–29:19.679
get access to land and energy in otherplaces. The Middle East has a lot ofland and has a lot of energy and they're
29:19.679–29:26.399
willing to strike deals and that is whyAltman was part of that trip looking tostrike a deal and what they the deal
29:26.399–29:35.360
that they struck was to build a massivedata center or multiple data centers inthe Middle East using their land and
29:35.360–29:42.559
their energy. But one of the things thatOpenAI has recently rolled out, theycall it the OpenAI for countries program
29:42.559–29:53.440
and it is this idea that they want toinstall OpenAI hardware and software inplaces around the world and explicitly
29:53.440–30:05.039
says we want to build democratic AIrails.We want to install our hardware andsoftware as a foundation of democratic
30:05.039–30:15.919
AI globally so that we can stop Chinafrom installing authoritarian AIglobally. But the thing that he does not
30:15.919–30:23.360
acknowledge is that there is nothingdemocratic about what he's doing. Youknow, the Atlantic executive editor says
30:23.360–30:31.360
we need to call these companies for whathey are. They are technoauthoritarians. They do not ask thepublic for any perspective on how they
30:31.360–30:36.559
develop the technology, what data theytrain the technology on, where theydevelop these data centers. In fact,
30:36.559–30:44.159
these data centers are often developedin the cover of night um under shellcompanies like Meta recently entered New
30:44.159–30:54.720
Mexico under the Shell company namedGreater Kudu LLC.>> Greater Kudu.LLC. And once the deal wasactually closed and the residents
30:54.720–30:58.960
couldn't do anything about anymore,that's when it was revealed, surprise,we're Meta and you're going to get a
30:58.960–31:04.320
data center that drinks all of yourfresh water.>> And then there was this wholecontroversy in Memphis around a data
31:04.320–31:12.159
center.>> Yes. So that is the data center thatElon Musk is building. So meanwhile,Musk is saying Alman is terrible.
31:12.159–31:19.120
Everyone should use my AI. And ofcourse, his AI is also being developedusing the same environmental and public
31:19.120–31:25.039
health costs. So he built this massivesupercomputer called Colossus inMemphis, Tennessee that's training
31:25.039–31:36.880
Grock, the chatbot that people canaccess through X and that is beingpowered byaround 35 unlicensed methane gas
31:36.880–31:45.440
turbines that are pumping thousands oftons of toxic air pollutants into thegreater Memphis community. And that
31:45.440–31:54.640
community has long suffered a lack ofaccess to clean air, a fundamental humanright.>> So I want to go to interestingly Sam
31:54.640–32:03.279
Alman testifying in front of Congressabout solutions to the high energyconsumption of artificial intelligence.
32:03.279–32:09.360
>> In the short term, I think this probablylooks like more natural gas. Um althoughthere are some applications where I
32:09.360–32:16.000
think solar can really help. In themedium-term, I hope it's advancednuclear uh fish and fusion. More energy
32:16.000–32:24.320
is important well beyond AI.>> So that's open AI's Sam Alman. This istestifying before the Senate and talking
32:24.320–32:31.600
about everything from uh solar tonuclear power. Something that was foughtin the United States by environmental
32:31.600–32:38.480
activists for decades. So you have thesehuge old uh nuclear power plants, butmany say you can't make them safe no
32:38.480–32:45.279
matter how small. and smart you makethem.>> This is one of the things of the manythings that I'm concerned about with the
32:45.279–32:52.000
current trajectory of AI development.This is a second order tertiary ordereffect is that because these companies
32:52.000–32:59.679
are trying to claim that the AIdevelopment approach they took doesn'thave climate harms. They are explicitly
32:59.679–33:05.360
evoking nuclear again and again andagain as nuclear will solve the problem.And it has been effective. I have talked
33:05.360–33:13.519
with certain AI researchers who thoughthe problem was solved because ofnuclear and in order to try and actually
33:13.519–33:24.000
build more and more nuclear plants, theyare lobbying governments to try andunwind the regulatory structure around
33:24.000–33:32.159
nuclear power plant building. I meanthis is this is like crazy on so manylevels that they're not just trying to
33:32.159–33:42.320
develop these the AI technologyrecklessly. They are also trying to laydown infrastructure and nuclearinfrastructure in this move fast break
33:42.320–33:50.480
things ideology. But for those who umare environmentalists and have longopposed nuclear will they be sucked in
33:50.480–33:54.320
by the solar alternative? But that exact
33:54.320–34:02.799
so data centers have to run 247. So theycannot actually run on just renewables.That is why the companies keep trying to
34:02.799–34:11.599
evoke nuclear as the solve all but solardoes not actually work when we do nothave sufficient enough energy storage
34:11.599–34:18.879
solutions for that 24/7 operation. We'llreturn to our conversation in a minutewith Karen How, author of the new book
34:18.879–34:32.800
Empire of AI: Dreams and Nightmares inSam Alman's Open AI. Stay with us.This is Democracy Now! Democracynow.org.
34:32.800–34:42.480
I'm Amy Goodman. In this holidayspecial, we're speaking with thejournalist Karen How, author of the newbook Empire of AI: Dreams and Nightmares
34:42.480–34:53.200
in Sam Alman's Open AI. She came intour studio in May. She lives in HongKong. I asked her to talk about what'shappening in China around artificial
34:53.200–35:02.079
intelligence.>> China and the US are the largest hubsfor AI research. They are the largestconcentration of AI research talent
35:02.079–35:09.440
globally. Um, China other than SiliconValley, China really is the only otherival in terms of talent density and the
35:09.440–35:14.800
amount of capital investment and theamount of infrastructure that is goinginto AI development. In the last few
35:14.800–35:25.440
years, what we have seen is the USgovernment has been aggressively tryingto stay number one and one of themechanisms that they have used is export
35:25.440–35:34.480
controls. A key input into these AImodels is the computationalinfrastructure and the computer chipsfor installing into the data centers for
35:34.480–35:43.040
training these models. And thesecomputer chips are the in order todevelop the AI models. Companies areusing the most bleeding edge computer
35:43.040–35:48.640
chip technology. It's like the every twoyears a new chip comes out and theyimmediately start using that to train
35:48.640–35:55.760
the next generation of AI models. Thosecomputer chips are designed by Americancompanies, the most prominent one being
35:55.760–36:03.920
Nvidia in California. And so the USgovernment has been trying to use exportcontrols to prevent Chinese companies
36:03.920–36:12.320
from getting access to the most cuttingedge computer chips. That has all beenunder the recommendation of Silicon
36:12.320–36:23.119
Valley saying this is the way to preventChina from being number one. and likeput export controls on them and don't
36:23.119–36:32.800
regulate us at all so we can stay numberone and they will fall behind. What hashappened insteadis because there is a strong base of
36:32.800–36:41.920
talent of AI research talent in Chinaunder the constraints of fewercomputational resources, Chinesecompanies have actually been able to
36:41.920–36:50.079
innovate and develop the same level ofAI model capabilities as Americancompanies with two orders of magnitude
36:50.079–37:00.240
less computational resources, lessenergy, less data. So, I'm talkingspecifically about um the Chinesecompany Highfire, which developed this
37:00.240–37:09.040
model called Deep Seek earlier this yearthat briefly tanked the global economybecause the company said that their
37:09.040–37:21.680
um training this one AI model costaround $6 million when OpenAI wastraining models that cost hundreds ofmillions if not over tens of billions of
37:21.680–37:30.640
dollars. And that delta demonstrated topeople that this what Silicon Valley hastried to convince everyone for the last
37:30.640–37:39.040
few years that this is the only path togetting more AI capabilities is totallyfalse and actually the techniques that
37:39.040–37:46.560
chi the Chinese company was using wereones that existed in the literature andjust had to be assembled. They used a
37:46.560–37:52.800
lot of engineering sophistication to dothat, but they weren't actually usingfundamentally new techniques. They were
37:52.800–37:58.720
ones that actually already existed.>> So explain it further because I think alot of people just can't get their minds
37:58.720–38:02.550
around this. How do you do thistraining?
38:02.560–38:11.280
>> So there's software called neuralnetworks which is essentially a massivestatistical engine. is doing lots and
38:11.280–38:19.599
lots of sophisticated statisticalcomputation to try and ascertain whatkinds of patterns exist in data sets. So
38:19.599–38:26.000
typically in in the past before we goto large language models it would bedoing something like um looking at MRI
38:26.000–38:34.560
scans and checking the patterns of whatdoes cancer look like in an MRIscan. Um now with GBT what it's looking
38:34.560–38:38.320
at is what are the patterns of theEnglish language? what is the syntax,
38:38.320–38:48.640
the structure, figures of speech thatare typically used and then it usesthose patterns to construct newsentences. That's how generative AI
38:48.640–38:55.280
works. And the reason why it's socomputationally expensive is becauseit's crunching the numbers for those
38:55.280–39:03.680
patterns. And the more data you feed in,the more it has to crunch. And so it weused to train these AI models on you
39:03.680–39:12.640
know a powerful laptop like maybe onecomputer chip maybe the richest labsacademic labs like MIT they would be uh
39:12.640–39:20.240
training on a couple or a dozen computerchips and companies like Google theywould be training maybe on a couple
39:20.240–39:27.680
hundred computer chips. We are nowtalking about hundreds of thousands ofcomputer chips training a single model.
39:27.680–39:37.119
Um and that is the you know that is whatopen AI says is necessary to build thesetechnologies and that is what deepseek
39:37.119–39:45.119
proved wrong. So, let me ask yousomething, Karen. uh the latest news umas you're traveling in the United States
39:45.119–39:55.760
before you go back to Hong Kong uh ofTrump's attack on academia, how thisfits in. Um how could Trump's attack on
39:55.760–40:04.240
international students specificallytargeting the what more than 250,000 aquarter of a million Chinese students
40:04.240–40:11.839
and revoking their visas impact thefuture of the AI industry. But not justChinese students because what's going on
40:11.839–40:21.280
here now is terrifying students aroundthe world and because labs are shuttingdown in all kinds of ways here uh US
40:21.280–40:29.920
students as well uh deciding to goabroad.This is just the latest action that theUS government has taken over the last
40:29.920–40:42.480
few years to really alienate a keytalent pool for US innovation.Originally, there were more Chineseresearchers working in the US
40:42.480–40:52.160
contributing to US AI than there were inChina because [snorts] just a few yearsago, Chinese researchers aspired to work
40:52.160–41:02.960
for American companies. They wanted tomove to the US. They wanted tocontribute to the US economy. Theydidn't want to go back to their home
41:02.960–41:11.760
country. But because of what was calledthe China Initiative, which was the afirst Trump era initiative to try and
41:11.760–41:18.960
criminalize Chinese academics orethnically Chinese academics, some ofwhom were actually Americans um based on
41:18.960–41:26.480
just paperwork errors. They would accusethem of being spies. That was one of thefirst actions. Then of course the
41:26.480–41:34.720
pandemic happened and the USChina tradescalations started amplifyinganti-Chinese rhetoric. All of these led
41:34.720–41:44.800
and now with the potential ban oninternational students. All of thesehave led more and more Chineseresearchers to just opt for staying at
41:44.800–41:54.319
home and contributing to the Chinese AIecosystem. And this was a prerequisiteto High-fly pulling off Deepseek. If
41:54.319–42:03.280
there had not been that concentrationand buildup of AI talent in China, theyprobably would have had a much harder
42:03.280–42:11.040
time innovating around circumventingthese export controls that the USgovernment was imposing on them. But
42:11.040–42:21.680
because they now have a highconcentration of top talent, some of thetop talent globally,when those restrictions were imposed,
42:21.680–42:29.839
they were able to innovate around them.So Deepseek is literally a product ofthis continuation of that alienation and
42:29.839–42:38.160
with the US continuing to take thistance, it is just going to get worse.And as you mentioned, it's not justChinese researchers. I literally just
42:38.160–42:45.680
alked to a friend in academia that saidshe's considering going to Europe nowbecause she just cannot survive without
42:45.680–42:52.400
hat public funding. And Europe Europeancountries are seeing a criticalopportunity offering milliondoll
42:52.400–42:59.599
packages. Come here, we'll give you alab. We'll give you millions of dollarsof funding. I mean this is the fastest
42:59.599–43:09.839
way to brain drain this country.>> I mean what many are saying US's braindrain is their brain gain. Yes.>> And this also reminds us of history. You
43:09.839–43:20.560
have the Chinese rocket scientist ChenShuen who in the 1950s was inexplicablyheld under house arrest for years and
43:20.560–43:28.800
then Eisenhower has him deported toChina. He becomes the father of rocketscience and uh China's entry into space.
43:28.800–43:34.400
And he said he would never again stepfoot into the United States even thoughoriginally that was the only place he
43:34.400–43:41.920
wanted to live.>> Yes. And there was a I believe agovernment official, a US governmentofficial who said that was the dumbest
43:41.920–43:53.040
mistake the US ever made.>> Um you we talk about the brain drain andthe brain gain. Okay. Again, uh somemore rhyming, the doomers and the
43:53.040–44:04.560
boomers. Um, I want to talk about whatan AI apocalypse looks like, meaning howit brings us to apocalypse, but also um
44:04.560–44:13.349
how uh people say it could lead us to autopia. What are the two trackstrajectories?
44:13.359–44:19.680
>> It's a great question and I ask boomersand doomers this all the time. Can youarticulate to me exactly how we get
44:19.680–44:21.200
here? And the issue is that they
44:21.200–44:30.079
cannot. And this is why I call it quasireligious. It really is based on belief.I mean, I was talking with oneresearcher who identified as a boomer.
44:30.079–44:39.359
And I said, you know, he his his eyeswere wide and he he really lit upsaying, you know, once we get to AGI,game over. Everything becomes perfect.
44:39.359–44:49.760
And I asked him, I was like, can youexplain to me how does AGI feed peoplethat haven't don't have food on thetable right now? And he was like, "Oh,
44:49.760–44:59.680
you're talking about like the floorand how to elevate their qualityof life." And I was like, "Yes, becausethey are also part of all of humanity."
44:59.680–45:04.640
And he was like, "I'm not really surehow that would happen, but I think itcould it could help the middle class get
45:04.640–45:09.920
more economic opportunity." And I waslike, "Okay, but how does that happen aswell?" And he was like, "Well, once
45:09.920–45:15.839
these come once we have AGI and it canjust create trillions of dollars ofeconomic value, we can just give them
45:15.839–45:19.119
cash payouts." And I was like, who'sgiving them cash payouts? What
45:19.119–45:20.560
institutions are giving them? You know,
45:20.560–45:28.319
like it it doesn't when you actuallytest their logic, it doesn't reallyhold. And with the doomers, I mean, it's
45:28.319–45:37.920
the same thing. like their belief isultimatelywhat I realized when reporting on thebook is they believe AGI is possible
45:37.920–45:46.480
because of their belief of how the humanbrain works. They believe humanintelligence is inherently fullycomputational. So if you have enough
45:46.480–45:53.040
data and you have enough computationalresources, you will inevitably be ableto recreate human intelligence. It's
45:53.040–46:01.200
just a matter of time. And to them, thereason why there would that would leadto an apocalyptic scenario is humans, we
46:01.200–46:07.920
learn and improve our intelligencethrough communication. And communicationis inefficient. We miscommunicate all
46:07.920–46:19.280
the time. And so for AI intelligences,they would be able to rapidly getsmarter and smarter and smarter byhaving perfect communication with one
46:19.280–46:26.800
another as digital intelligences. And somany of these people who selfidentify asdoomers say there has never been in the
46:26.800–46:36.640
history of the the universe a speciesthat was superior to another species a aspecies that was able to rule over um a
46:36.640–46:45.280
more superior species. So they thinkthat ultimately AI will evolve into ahigher species and then start ruling us
46:45.280–46:54.319
and then maybe decide to get rid of usaltogether. I'm wondering if you cantalk about any model of a country, not a
46:54.319–47:05.599
company,>> that is pioneering a way ofdemocratically controlled artificialintelligence.>> I don't think it's actively happening
47:05.599–47:15.599
right now.The EU has had the EU AI act, which istheir major piece of legislation tryingto develop a riskbased, rightsbased
47:15.599–47:27.760
framework for governing AI umdeployment.Buto me, one of the keys of democratic AIgovernance is also democratically
47:27.760–47:38.480
developing AI. And I don't think anycountry is really doing that. And what Imean by that is there are AI has asupply chain. It needs data. It needs
47:38.480–47:45.599
land. It needs energy. It needs water.And it also needs spaces in which thesecompanies need access to to then deploy
47:45.599–47:52.240
their technology. Schools, hospitals,government agencies. Silicon Valley hasdone a really good job over the last
47:52.240–47:59.599
decade of making people feel that theircollectively owned resources are SiliconValleys. You know, I have I talk with
47:59.599–48:05.920
friends all the time who say, "We don'thave data privacy anymore." So, like,what's more what's what is more data to
48:05.920–48:07.359
these companies? Like, I'm fine just
48:07.359–48:16.560
giving them all of my data. But thatdata is yours. You know, thatintellectual property is the writers andartists intellectual property. That land
48:16.560–48:24.480
is a community's land. Those schools arethe students and teachers schools. Thehospitals are the doctors and nurses and
48:24.480–48:32.640
patients hospitals. These are all sitesof democratic contestation in thedeployment in the development and the
48:32.640–48:38.720
deployment of AI. And just like thoseChilean water activists that we talkedabout who aggressively understood that
48:38.720–48:47.680
fresh water was theirs and theywere not willing to give it up unlessthey got some kind of mutuallybeneficial agreement for it. We need to
48:47.680–49:00.559
have that spirit in protectingour data, our land, our water, and ourschools so that companies inevitablywill have to adjust their approach
49:00.559–49:07.200
because they will no longer get accessto the resources they need or the spacesthat they need to deploy in. In 2022,
49:07.200–49:14.079
Karen, you wrote a piece for MITechnology Review headlined a new visionof artificial intelligence for the
49:14.079–49:22.160
people. In a remote rural town in NewZealand, an indigenous couple ischallenging what AI could be and who it
49:22.160–49:24.640
should serve. Who are they? This was a
49:24.640–49:36.400
wonderful story that I did where thecouple um they run to media. It's anonprofit MAI radio station in NewZealand. And the Maui people have
49:36.400–49:43.440
suffered a lot of the same um challengesas many indigenous peoples around theworld. The history of colonization led
49:43.440–49:49.200
them to rapidly lose their language andthere are very few Mauy speakers in theworld anymore. And so in the last few
49:49.200–49:54.960
years there's been an attempt to revivethe language and the New Zealandgovernment has tried to repent by by
49:54.960–50:02.079
trying to encourage that the revival ofthat language. But this nonprofit radiostation, they had all this wonderful
50:02.079–50:10.079
archival material, archival audio oftheir ancestors speaking the Mailanguage that they wanted to provide to
50:10.079–50:16.319
Maui speakers, ma Mai learners aroundthe world as an educational resource.The problem is in order to do that they
50:16.319–50:22.800
needed to transcribe the audio so thatMai learners could actually listen, seewhat was being said, click on the words,
50:22.800–50:29.760
understand the translation and actuallyturn it into an active learning tool.But there were so few Maui speakers that
50:29.760–50:39.200
can speak at that advanced level thathey realized they had to turn to AI.And this is a key part of my book'sargument is I'm not critiquing all AI
50:39.200–50:44.800
development. I'm specifically critiquingthe scale at all cost approach thatSilicon Valley has taken. But there are
50:44.800–50:50.880
many different kinds of beneficial AImodels, including what they ended updoing. So they took a fundamentally
50:50.880–50:57.760
different approach. First and foremost,they asked their community, do we wanthis AI tool? Once the community said
50:57.760–51:06.480
yes, then they moved to the next step ofasking people to fully consent todonating data for the training of this
51:06.480–51:12.559
tool. They explained to the communitywhat this data was for, how it would beused, how they would then guard that
51:12.559–51:20.640
data and make sure that it wasn't usedfor other purposes.They collected around a couple hundredhours of audio data in just a few days
51:20.640–51:27.119
because the community rallied supportaround this project and only a couplehundred hours was enough to create a
51:27.119–51:33.040
performant speech recognition modelwhich is crazy when you think about thescales of data that these Silicon Valley
51:33.040–51:39.200
companies require. And that is onceagain a lesson that can be learned isactually there's plenty of research that
51:39.200–51:46.800
shows when you have highly curated smalldata sets, you can actually create verypowerful AI models and then once they
51:46.800–51:55.040
had that tool, they were able to doexactly what they wanted to open sourceand resour uh open source thiseducational resource to their community.
51:55.040–52:08.800
And so my vision for AI development inthe future is to have more smalltaskspecific AI models that are notrained on vast polluted data sets but
52:08.800–52:18.240
small curated data sets and thereforeonly need small amounts of computationalpower and can be deployed in challenges
52:18.240–52:27.920
that we actually need to tackle forhumanity.mitigating climate change by integratingmore renewable energy into the grid,
52:27.920–52:37.599
improving health care, by doing moredrug discovery.>> So, as we finally do wrap up, what youwere most shocked by, you've been doing
52:37.599–52:49.599
uh this journalism, this research foryears, what you were most shocked by inwriting Empire of AI.I originally thought that I was going to
52:49.599–52:56.800
write a book focused on vertical harmsof the AI supply chain. Here's how laborexploitation happens in the AI industry.
52:56.800–53:04.559
Here's how the environmental harms arearising out of the AI industry. And athe end of my reporting, I realized that
53:04.559–53:10.319
here is a horizontal harm that'shappening here. Every single communitythat I spoke to, whether it was artists
53:10.319–53:16.800
having their intellectual property takenor Chilean water water activists havingtheir fresh water taken, they all said
53:16.800–53:24.640
that when they encountered the empire,they initially felt exactly the sameway. A complete loss of agency to
53:24.640–53:33.839
self-determine their future. And that iswhen I realized the horizontal harm hereis AI is threatening democracy. If the
53:33.839–53:46.240
majority of the world is going to feelthis loss of agency overself-determining their future, democracycannot survive and again specifically
53:46.240–53:55.599
Silicon Valley's approach scale at allcosts AI development.>> But you also chronicle the resistance.You talk about how the Chilean water
53:55.599–54:05.680
actors felt at first, how the artistsfeel at first. So talk about thestrategies that these people havemployed and if they've been effective.
54:05.680–54:12.400
>> So the amazing thing is that there hasince been so much push back. Theartists have then said wait a minute we
54:12.400–54:18.640
can sue these companies. The Chileanwater activists said wait a minute wecan fight back and protect these water
54:18.640–54:25.200
esources. The Kenyan workers that Ispoke to who are contracted by OpenAIthey said we can unionize and escalate
54:25.200–54:35.680
our story to international mediattention. And so even in these evenwhen I thought that these communitiesyou could argue are the most vulnerable
54:35.680–54:45.040
in the world have the least amount ofagency, they were the ones thatremembered that they do have agency andthat they can seize that agency and
54:45.040–54:54.960
fight back. And I think it it was it wasremarkably heartening to encounter thosepeople to remind me that actually the
54:54.960–55:02.960
first step to reclaiming democracy isremembering that no one can take youragency away.>> Karen How, author of the new book Empire
55:02.960–55:11.680
of AI: Dreams and Nightmares and SamAlman's Open AI. Karen How is a formereporter at the Wall Street Journal and
55:11.680–55:19.760
MIT Technology Review. And that does itfor this special broadcast. [music]Democracy Now is produced with Mike
55:19.760–55:25.040
Burke, Renee Fel, Dina Guster, MessiahRhodess, Nurmine Shake, Maria Terasena,Nicole Salazar, [music] Sarin Nasser,
55:25.040–55:31.599
Trina Nadura, Sam Alov, T Maria Joe,John Hamilton, Robbie [music] Karen,Honey Massud, and Safwet Naz. Our
55:31.599–55:37.280
executive directors, Julie Crosby.Special thanks to Becca Stelli, JohnRandolph, Paul Powell, [music] Mike
55:37.280–55:44.160
DeFippo, Miguel Nggera, Hugh Grant, CarlMarxer, Dennis Moahan, David Prud,Dennis McCormack, Matt Elely, Anna
55:44.160–55:52.799
Osbeck, Emily Anderson, Dante Toriieri,Buffy [music] St. Marie Hernandez withJuan Gonzalez. I'm Amy Goodman. Happy
55:52.799–56:01.920
New Year.Thanks for watching Democracy Now onYouTube. Subscribe to the channel andturn on notifications to make sure you
56:01.920–56:08.559
never miss a video. And for more of ouraudience supported journalism, go todemocracynow.org or where you can
56:08.559–56:15.000
download our news app, sign up for ournewsletter, subscribe to the dailypodcast, and so much
0:03.200–0:12.320
這是民主現在 [音樂] 民主現在.org 的戰爭與和平報導。我是艾美·古德曼。AI帝國，那是
0:12.320–0:20.320
記者Karen How新書的名稱，她一直密切報導人工智慧產業的崛起，並聚焦於Sam Altman的Open AI。
0:20.320–0:31.359
那就是ChatGPT背後的企業。Karen How將這些行動與標誌著這段歷史的同樣公開暴力與殘酷相比較，但他們也掠奪並提取珍貴資源以
0:31.359–0:40.320
開放AI。那就是ChatGPT背後的企業。卡倫·豪將
0:40.320–0:48.399
餵養他們對人工智慧的願景。藝術家與作家的作品、無數個人發布關於他們
0:48.399–0:56.879
超級電腦的數據。她寫道：「過去一年來，川普政府日益擁抱人工智慧產業。
0:56.879–1:05.920
同樣顯著的暴力與殘酷行為，這些行為標誌著這段歷史，但它們同樣
1:05.920–1:15.520
超級電腦。她寫道：「過去一年來，川普政府越來越擁抱人工智慧產業。
1:15.520–1:22.880
餵養它們對人工智慧的願景。藝術家與作家的作品、無數個人
1:22.880–1:32.000
監管規定。簽署行政命令後不久，其家族企業Trump Media and Technology便宣布以60億美元
1:32.000–1:40.880
與一家旨在建造全球首座可行核融合發電廠以供應人工智慧專案電力的公司合併。Karen How是
1:40.880–1:49.040
超級電腦所需的土地、能源與水資源。她寫道：「過去一年來，特朗普政府日益擁抱AI產業。
1:49.040–1:58.159
十二月，特朗普簽署行政命令，禁止各州與地方政府制定自己的AI
1:58.159–2:08.000
監管規定。簽署命令後不久，其家族企業特朗普媒體與科技宣布與一家旨在建造
2:08.000–2:20.160
全球首座可行核融合發電廠以供應AI專案的企業進行60億美元
2:20.160–2:27.680
技術，但大多數人是透過ChatGPT接觸到它的，而我在書中提出的論點以及書名
2:27.680–2:34.160
《華爾街日報》與《麻省理工科技評論》的前記者，在那裡她成為首位
2:34.160–2:42.720
為開放AI撰寫人物專訪的記者。民主現在新聞的胡安·岡薩雷斯與我在五月採訪了她。國家書評人協會最近
2:42.720–2:50.480
將她的書《AI帝國：山姆·阿爾特曼開放AI的夢想與夢魘》評為
2:50.480–2:57.760
2025年最佳非小說類書籍的決選作品。我首先請卡倫·豪解釋
2:57.760–3:03.840
需要進行這些訓練的電腦。所以我們談的是將整個英文互聯網的內容輸入到這些
3:03.840–3:10.720
技術的集合，但大多數人是透過ChatGPT接觸到它的，而我在書中論述的以及書名
3:10.720–3:17.040
所指的AI帝國，實際上是对導致我們走向ChatGPT的特定AI發展軌跡的批判，並且
3:17.040–3:25.360
自ChatGPT以來持續至今，這特別是指矽谷在AI發展上不惜代價追求規模的
3:25.360–3:31.840
非凡的、呃，某種類型的 AI 開發，這造成了許多社會、勞動和環境上的危害，而且這正是
3:31.840–3:40.480
模型是在數據上訓練的，它們需要電腦來在這些數據上進行訓練。但矽谷
3:40.480–3:50.000
在過去幾年裡所做的事，以及開放AI所做的事，是他們開始大幅增加數據量以及進行此類訓練所需電腦的
3:50.000–3:51.030
規模。所以我們談論的是將整個英文互聯網的內容餵入這些
3:51.040–3:56.560
模型，包括書籍、科學文章、所有正在創作的知識產權，以及運行數萬甚至數十萬顆
3:56.560–4:03.920
晶片的大型超級電腦，這些晶片的面積可能相當於數十個甚至數百個足球場，並且現在實際上佔用了城市的全部能源需求。所以這是
4:03.920–4:13.040
一種非凡的AI發展類型，造成了許多社會、勞工和環境危害，而這正是
4:13.040–4:22.479
我借用帝國這個類比的根本原因。>> 卡倫，你能否再多談談這些龐大資料中心不僅是能源需求
4:22.479–4:29.120
，還有水資源需求，這些資料中心本質上正是這個不斷擴張產業的骨幹？
4:29.120–4:36.800
>> 絕對可以。我提供兩項關於能源與水資源的數據。談到能源需求時，
4:36.800–4:44.080
麥肯錫最近發布一份報告指出，根據目前AI運算基礎設施擴張的速度，在未來五年內，
4:44.080–4:50.000
我們需要在全球電網上架設的能源量，將相當於加州年度能源消耗量的兩到六倍。而且這主要將由化石燃料供應。我們已經看到報導指出，
4:50.000–4:58.080
煤電廠的壽命被延長。它們本應退役，但現在為了支援這些資料
4:58.080–5:05.919
中心的發展，它們無法退役。我們也看到未經許可的天然氣燃氣輪機被安裝，以服務這些資料
5:05.919–5:13.680
中心。從淡水角度來看，這些資料中心需要淡水進行訓練。它們不能
5:13.680–5:21.840
使用其他類型的水，因為它會腐蝕設備。它可能導致細菌生長，而且大多數時候它實際上直接接入公共飲用水供應，因為這是已經鋪設好的基礎設施，用於向不同企業和不同家庭提供乾淨的淡水。而彭博最近進行了一項
5:21.840–5:27.520
不僅是我們需要關注的淡水總量問題，更是這些淡水在全球的分配問題。
5:27.520–5:37.840
基礎設施在全球的分布情況。大多數人熟悉的是 ChatGPT 這種面向消費者的 AI，但
5:37.840–5:45.600
AI 的軍事層面呢？基本上，我們發現矽谷公司正成為下一代
5:45.600–5:53.199
國防承包商。>> OpenAI 和許多其他公司轉向國防產業的原因之一是，
5:53.199–5:59.600
國防工業的原因，因為他們在開發這些技術上花費了巨額資金。
5:59.600–6:07.520
它們花費數千億美元來訓練這些模型，並且需要收回這些成本，而能夠支付如此龐大金額的行業和地區有限。因此，
6:07.520–6:15.360
我們看到它們與國防產業的關係日益緊密。我們也看到矽谷利用美國政府來實現其帝國擴張的野心。你可以說美國
6:15.360–6:22.720
轉向國防工業。我們也看到矽谷利用美國政府來實現他們建立帝國的野心。你可以說美國
6:22.720–6:29.039
政府也在試圖利用矽谷，反過來實現他們建立帝國的野心。呃，但顯然
6:29.039–6:39.039
試圖獲取國防合約，並將其技術越來越深入地整合到軍事基礎設施中，這確實令人擔憂。我想談談你去過的國家或你報導的故事，因為
6:39.039–6:44.800
這太驚人了，你從肯亞到烏拉圭再到智利的報導深度。嗯，你談到了水的使用，我也想問你關於核能的問題。嗯，但在智利，
6:44.800–6:52.560
這些數據中心及其使用的水資源以及對此的抵抗情況如何？>> 是的。智利有一段有趣的歷史，因為它曾
6:52.560–7:00.240
長期處於獨裁統治之下。在那段時期，大多數公共資源被私有化，包括水資源。但由於一個
7:00.240–7:08.240
例外情況，大聖地亞哥都會區有一個社區實際上仍然擁有公共淡水資源，該資源在緊急情況下既服務該社區也服務全國其餘地區。
7:08.240–7:15.520
這正是 Google 選擇嘗試建立數據中心的社區。>> 並且它是免費的>> 而且，你知道，我完全不知道。這
7:15.520–7:25.440
在獨裁統治下很長一段時間。所以在這段時間內，大多數公共資源被私有化，包括水資源。但由於
7:25.440–7:31.440
根據閱讀文件後的信念，認為 Google 繳納的稅款是繳給他們註冊辦公室、行政辦公室所在地的，而不是他們建立數據中心的地方。所以
7:31.440–7:38.960
他們沒有從這個數據中心獲得任何直接利益，也沒有看到對該數據中心允許提取的淡水有任何限制。因此，這些活動人士說，
7:38.960–7:48.479
「等等，絕對不行。我們不會允許這個數據中心進入，除非他們給我們一個合理的理由，說明它如何使我們受益。」於是，
7:48.479–7:56.319
這是一個很好的問題。但社區告訴我的是，他們甚至沒有為此繳納稅款，因為他們
7:56.319–8:01.520
每一個鄰居的門，向社區散發傳單，告訴他們這家公司正在拿走我們的淡水資源，卻沒有給予任何回報。因此，他們的行動如此激烈地升級，
8:01.520–8:08.479
甚至升級到了 Google 智利分公司。然後升級到了 Google 山景城總部，順便說一句，後者隨後派出了只會說英語的代表前往智利。
8:08.479–8:14.240
但最終，這升級到了智利政府。現在，智利政府舉行圓桌會議，邀請
8:14.240–8:20.960
對新鮮水資源施加的限制，這個數據中心被允許抽取的水量。所以這些活動人士說，
8:20.960–8:27.599
共同討論如何使數據中心的發展對社區更加有益。活動人士表示，戰鬥尚未結束。僅僅因為他們被邀請到談判桌旁，並不意味著一切都突然變好。他們需要
8:27.599–8:35.440
保持警惕。他們需要繼續嚴格審查這些項目。但到目前為止，他們已經能夠阻止這個
8:35.440–8:40.399
項目長達四到五年，並獲得了坐在談判桌旁的席位。>> 那麼，這些西方公司如何
8:40.399–8:46.800
正在未經回報的情況下奪取我們的淡水資源。因此他們大規模升級行動，
8:46.800–8:51.839
升級到了智利谷歌。接著升級到了加州山景城的谷歌總部，順便一提，谷歌總部隨後派遣了
8:51.839–9:01.680
只會說英文的代表前往智利。但最終還是升級到了智利政府。而智利政府現在定期舉行會議，
9:01.680–9:06.160
互聯網上抓取來的，你不能直接將這些數據輸入到你的 AI 模型中，因為數據中包含很多
9:06.160–9:12.640
東西。數據嚴重受污染。需要進行清理。需要進行標註。所以這就是
9:12.640–9:19.920
活動人士表示，這場戰鬥尚未結束。僅僅被邀請到談判桌上，並不意味著一切都突然變好了。他們需要
9:19.920–9:25.519
保持警惕。他們需要繼續嚴格審查這些專案。但到目前為止，他們已成功阻止該
9:25.519–9:36.640
專案長達四到五年，並獲得了談判桌上的席位。>> 那麼，這些西方公司本質上是如何剝削
9:36.640–9:40.240
全球南方的勞工呢？你進入
9:40.240–9:44.640
一種稱為資料標註公司的地方。那是什麼？
9:44.640–9:51.279
是的。因為現代的人工智慧系統是透過大量數據進行訓練，這些數據是從網路抓取而來的，
9:51.279–9:59.279
你無法直接將這些數據輸入到人工智慧模型中，因為數據中包含許多
9:59.279–10:05.279
在那些數據中，它受到嚴重污染，需要進行清理和標註，這就是
10:05.279–10:14.160
資料標註公司就此介入。這些是中間商公司，雇用合約勞工來為這些 AI 公司提供服務，以進行
10:14.160–10:21.600
這類型的資料準備工作。而當 OpenAI 開始考慮將其產品商業化，
10:21.600–10:28.880
並思考讓我們將能生成各種類型文字的文本生成機器交到數百萬用戶手中時。
10:28.880–10:34.880
他們意識到需要某種內容審核機制。他們需要開發一種過濾器，將這些模型包裹起來，
10:34.880–10:42.320
以防止這些模型實際上向用戶噴出種族歧視、仇恨和有害的言論，
10:42.320–10:48.640
否則這將無法成為一個非常具有商業可行性的產品。因此，他們與這些中間商公司簽約，
10:48.640–10:56.320
在肯亞，肯亞工人必須閱讀大量網際網路上的最糟文字，以及 AI 生成的
10:56.320–11:03.600
文字，其中 OpenAI 提示其自己的 AI 模型去想像網際網路上的最糟文字，然後告訴這些
11:03.600–11:11.839
肯亞工人詳細地將它們分類到詳細的分類法中，例如這是色情內容嗎，這是暴力內容嗎，
11:11.839–11:18.480
該暴力內容有多麼具視覺衝擊力，以便教導其過濾器所有它必須
11:18.480–11:25.120
封鎖的不同類型內容。這是一種極其不尋常的勞動形式。他們使用了許多其他類型的合約
11:25.120–11:32.480
勞工。但這些工人，如果有的話，每小時只賺幾美元。就像社交媒體時代一樣，
11:32.480–11:40.079
媒體、這些內容審核員所面臨的情況，他們在心理上遭受了極深的創傷。而最終，並沒有真正的
11:40.079–11:46.640
哲學來解釋為什麼這些工人每小時只賺幾美元，卻讓他們的生活毀滅。以及為什麼
11:46.640–11:52.240
研究人員，他們為這些模型貢獻心力，卻僅獲得百萬美元的薪酬，僅僅因為
11:52.240–12:00.079
研究人員，僅僅因為他們坐在矽谷 OpenAI 的辦公室裡，就能獲得百萬美元的薪酬包。這就是帝國的邏輯。這也呼應了我的標題，
12:00.079–12:10.320
AI 帝國。>> 所以，讓我們回到你的標題，AI 帝國。副標題是薩姆·阿爾特曼的 OpenAI 中的夢想與噩夢。所以
12:10.320–12:17.839
告訴我們關於薩姆·阿爾特曼的故事，以及 OpenAI 到底是怎麼回事，一直到他剛剛在中東
12:17.839–12:27.040
與特朗普總統、薩姆·阿爾特曼和伊隆·馬斯克在場時達成的協議。阿爾特曼非常具有矽谷特色。他的職業生涯
12:27.040–12:32.880
首先是作為一家初創公司的創始人，然後是 Y Combinator 的總裁，這是矽谷最著名的初創公司
12:32.880–12:40.800
加速器之一，之後是 OpenAI 的首席執行官。OpenAI 最終向世界介紹了
12:40.800–12:47.519
不惜一切代價的 AI 發展規模，這並非巧合，因為這就是矽谷
12:47.519–12:59.839
在阿爾特曼成長期間一直運作的模式。所以他是一個非常有策略性的人。他非常擅長講述關於
12:59.839–13:06.639
未來的故事，並描繪這些宏大的願景，讓投資者和員工希望成為其中的一部分。所以在早期
13:06.639–13:14.959
於 YC 時，他識別出 AI 將是可能爆發的趨勢之一。而他試圖建立一個
13:14.959–13:20.639
不同的投資組合和不同的計劃，以便根據哪一個爆發，將自己置於各種不同趨勢的
13:20.639–13:26.160
取決於哪一項技術獲得了成功。他投資於量子計算、核融合，以及
13:26.160–13:33.360
自駕車。他還建立了一個基礎人工智慧研究實驗室。最終，該人工智慧研究實驗室
13:33.360–13:43.760
開始迅速發展。因此，他成為該公司的執行長。最初，他將其設立為非營利組織
13:43.760–13:52.800
開始加速發展。所以他讓自己成為該公司的 CEO。嗯，最初他將其作為非營利
13:52.800–14:00.240
組織，試圖將其定位為對矽谷以利潤為驅動力的激勵措施的制衡。但在一年半內，
14:00.240–14:05.519
OpenAI 的高管們識別出，如果他們想成為這個領域的領先者，他們必須採取這種不惜一切代價的規模化方法，並且應該加引號。他們認為他們必須
14:05.519–14:11.920
這樣做。實際上還有其他許多方法來開發 AI 並在 AI 方面取得進展，而不採取這種方法。但一旦他們決定這樣做，他們
14:11.920–14:19.440
就意識到瓶頸是資本。巧合的是，薩姆·阿爾特曼是一代難得的募資人才。他
14:19.440–14:26.880
創建了這種新結構，在非營利組織內部嵌套一個營利部門，以成為這個募資工具，
14:26.880–14:33.440
為數百億乃至最終數千億美元提供資金，以追求他們決定採取的方法。而
14:33.440–14:41.600
這就是我們最終到達當前的 OpenAI，它是矽谷歷史上最資本主義的公司之一，繼續募集數百億，甚至阿爾特曼開玩笑說數萬億，以生產一種
14:41.600–14:49.600
矽谷，持續籌措數千億美元，甚至艾爾曼開玩笑說要籌措數兆美元，以生產出
14:49.600–14:58.399
與《AI 帝國：薩姆·阿爾特曼的 OpenAI 中的夢想與噩夢》一書的作者卡倫·霍（Karen Ho）繼續我們的對話。請繼續收看。
14:58.399–15:10.639
與《AI帝國：艾爾曼的OpenAI中的夢想與夢魘》新書作者Karen How一起，請繼續收看。
15:10.639–15:14.160
這是《Democracy Now!》，網站是 democracynow.org。
15:14.160–15:22.000
我是艾美·古德曼。在這次假日特別節目中，我們繼續採訪記者卡倫·豪（Karen Ho），她是新書《AI帝國：山姆·阿爾曼的Open AI中的夢想與夢魘》的作者。
15:22.000–15:32.000
卡倫於五月進入我們的攝影棚，當時她討論了AI將如何影響勞工。我們看到的一項事實是，這項技術已經對就業產生了巨大影響。
15:32.000–15:41.519
這不一定是因為技術本身真的有能力取代工作，而是人們認為它有能力做到這一點，因此高層主管正在裁員。
15:41.519–15:48.480
並非因為技術本身真的有能力取代工作，而是人們認為它有能力
15:48.480–15:59.600
這些公司繼續嘗試開發自動化勞工的技術，並試圖將它們轉向生產輔助勞工的技術。
15:59.600–16:10.399
這些公司繼續嘗試開發自動化勞動力的技術，並試圖將它們轉向生產輔助勞動力的技術。
16:10.399–16:11.519
>> 所以，Open AI對他們所稱的「人工通用智能」的定義是高度自主的系統，在大多數
16:11.519–16:20.720
經濟上有價值的工作中表現優於人類。所以，他們明確表示他們正試圖自動化地取代工作。我的意思是，什麼是
16:20.720–16:28.160
經濟上有價值的工。所以，他們明確表示他們正試圖自動化掉工作。我的意思是，
16:28.160–16:30.079
人們為了獲得報酬所做的事情。嗯，但麻省理工學院經濟學家傑羅姆·奧斯汀（Jaron Austin）和西蒙·
16:30.079–16:39.279
約翰遜（Simon Johnson）撰寫了一本非常棒的書《進步中的權力》，他們提到技術發展、所有技術革命都採取自動化勞工的方法
16:39.279–16:46.399
這不是因為不可避免，而是因為頂層的人選擇自動化這些工作。他們選擇設計
16:46.399–16:54.480
技術，以便他們能將其銷售給高層主管，並說你可以通過裁掉所有這些員工並使用我們的AI服務
16:54.480–17:03.839
來縮減成本。但在過去，我們看過一些研究，例如建議如果你開發一個醫生使用的AI工具
17:03.839–17:12.400
而不是取代醫生，你實際上會為病人提供更好的醫療保健。你會得到更好的癌症
17:12.400–17:19.280
診斷。如果你開發一個教師可以使用的AI工具，而不僅僅是取代教師的AI導師，你的孩子將獲得更好的
17:19.280–17:25.600
教育成果。這就是我所說的輔助勞工而非自動化勞工的意思。
17:25.600–17:32.559
>> 請解釋一下你的意思，因為我認為很多人甚至不理解人工智慧。而且
17:32.559–17:38.720
>> 並解釋一下你的意思，因為我認為很多人甚至不了解人工智慧。而且
17:38.720–17:42.720
>> 沒錯。所以這些公司試圖開發一種技術，他們將其定位為一台什麼都能做的
17:42.720–17:49.600
萬能機器。嗯，所以他們會嘗試說你可以使用這個，你可以跟ChatGPT進行心理治療。不，你不行。ChatGPT
17:49.600–17:59.919
不是持照的心理治療師。事實上，這些模型實際上會噴出大量醫療誤導資訊。而且已經有
17:59.919–18:05.919
許多例子顯示，用戶實際上因模型而受到心理傷害，因為模型會
18:05.919–18:14.559
繼續強化自我傷害行為。我們甚至見過一些案例，兒童與聊天機器人對話
18:14.559–18:22.480
並與這些聊天機器人發展出深厚的情感關係，在使用這些聊天機器人系統後實際上自殺了。
18:22.480–18:29.120
嗯，但這就是我的意思，當這些公司試圖開發自動化勞工的工具時。他們
18:29.120–18:34.559
將其定位為你可以現在僱用這個工具來取代僱用勞工。我的意思是，最近，山姆·阿爾曼在
18:34.559–18:42.480
一次會議上說：「我們最初說這些模型是律師行的初級合夥人，現在我們認為
18:42.480–18:50.080
他們實際上可以成為律師行的高級同事。」他的意思是不要僱用初級合夥人，不要僱用高級
18:50.080–18:59.280
同事，而只是使用我們的AI模型。我們已經看到職業階梯正在崩壞，因為許多不同的
18:59.280–19:08.160
白領工作、白領服務業以及其他行業都逐漸相信他們不需要
19:08.160–19:15.280
僱用實習生。他們不需要僱用入門級職位，他們只需要這些AI模型，而新
19:15.280–19:21.360
的大學畢業生現在正努力尋找工作機會，以幫助他們在這些行業中立足。所以，
19:21.360–19:28.559
你談到了山姆·阿爾曼，在第一部分中，我們觸及了他是誰，但我希望你能更深入地探討
19:28.559–19:36.640
山姆·阿爾曼是誰，他是如何突然出現在美國舞台上，在國會作證，實際上
19:36.640–19:46.559
警告AI的危險。所以，這在某種程度上保護了他。>> 嗯，人們把他視為先知。那是P-O-P。但現在，我們可以談談另一種先知，P-O-F-I-T。
19:46.559–19:56.960
並探討Open AI是如何成立的。Open AI與AI有何不同？
19:56.960–20:05.360
Open AI是一家公司，我的意思是它最初是作為非營利組織成立的，正如我所提到的，阿爾曼特別是在他
20:05.360–20:15.440
思考如何建立一個將引起轟動的根本AI研究實驗室時，他選擇將其設為非營利組織，因為他
20:15.440–20:25.120
思考如何建立一個能引起巨大轟動的根本性 AI 研究實驗室時，他選擇將其設為非營利組織，因為他
20:25.120–20:33.200
認清了自己無法在資金上競爭，而且他入局較晚，當時 Google 已經壟斷了
20:33.200–20:41.520
許多頂尖 AI 研究人才。如果他在資金上無法競爭，也無法在成為先驅者方面競爭，他
20:41.520–20:49.600
就需要其他類型的要素來真正招募人才、爭取公眾善意並建立名聲
20:49.600–20:58.159
以打造 OpenAI。因此他確立了一項使命，他確立了讓我將其設為非營利組織，並賦予它一項極具吸引力的使命，所以 OpenAI 的使命是
20:58.159–21:06.880
確保通用人工智慧造福全人類。我在書中引用的其中一句話是 Sam Altman
21:06.880–21:16.000
在 2013 年他自己部落格中引用的這句話。他當時是一位熱衷於部落格寫作的
21:16.000–21:23.600
談論他在商業、策略和矽谷創業生活方面的學習心得。這句話是：成功的人
21:23.600–21:31.200
建立公司，更成功的人建立國家。最成功的人建立宗教。然後他
21:31.200–21:37.520
在他的部落格中反思這句話，說：「對我來說，建立宗教的最佳方式其實是
21:37.520–21:46.240
建立一家公司。」>> 接著談論 Altman 如何被公司驅逐，然後又回歸。還有，我覺得非常
21:46.240–21:52.480
迷人的是，你能與這麼多 OpenAI 的員工交談。你認為有一道全面的禁令
21:52.480–21:59.360
阻止你。>> 是的。沒錯。正是如此。所以我是第一位為 OpenAI 撰寫專訪的記者。我在
21:59.360–22:06.880
2019 年於該公司內部進行了 3 天的嵌入式採訪，隨後我的專訪於
22:06.880–22:13.760
2020 年發表於 MIT Technology Review。當時我在專訪中指出
22:13.760–22:18.960
我所觀察到的這種張力：該組織名義上是非營利組織，但背地裡所展現的許多公共價值
22:18.960–22:26.720
秘密。他們宣揚合作精神。他們高度競爭。他們宣稱沒有商業意圖，但實際上它
22:26.720–22:32.159
保密。他們宣稱協作，卻高度競爭。他們宣稱沒有商業意圖，但事實上看來
22:32.159–22:38.320
他們剛剛獲得了微軟的 10 億美元投資。看來他們正迅速發展
22:38.320–22:44.640
商業意圖。因此我將這些寫入專訪中，OpenAI 對此深感不滿，他們
22:44.640–22:50.880
拒絕與我交談長達 3 年。但當我開始撰寫這本書，開始聯絡
22:50.880–22:59.200
特別想和我談話，因為他們認為我會公正地呈現公司內部實際發生的真相，並
22:59.200–23:07.520
特別希望與我交談，因為他們認為我會公正地呈現公司內部實際發生的真相，
23:07.520–23:13.679
並了解這家公司的發展歷程。我能夠真正深入探究，觸及問題的核心。
23:13.679–23:21.760
和這家公司發展歷程神話化、敘事化的背後，發現真相。我能夠真正深入核心，
23:21.760–23:29.200
我稱之為在矽谷內部發展出來的準宗教運動。人工通用
23:29.200–23:36.480
智能的概念並非基於科學基礎。這是一個我們可以根本性地重現人類
23:36.480–23:42.400
智能於電腦中的想法。這個想法其實存在了很長一段時間。人工智慧領域成立於
23:42.400–23:48.320
早在 1950 年代，這也是該領域最初的意圖。我們如何在
23:48.320–23:50.080
智能於電腦中的想法。這個想法其實存在了很長一段時間。AI 領域早在
23:50.080–23:58.159
1950 年代就已創立，那也是該領域最初的意圖。我們如何在
23:58.159–24:05.840
甚至對於人類智能是什麼都無法達成共識。因此，將整個搜尋領域和技術與
24:05.840–24:12.640
英國數學家艾倫·圖靈提出的著名問題。但直到今天，我們甚至對人類
24:12.640–24:22.159
智能是什麼都還沒有科學共識。因此，將整個研究領域和技術建立在
24:22.159–24:29.440
人類智能的基礎上是一項非常棘手的任務，因為沒有好的指標來評估我們是否真的
24:29.440–24:36.640
達成了目標，也沒有藍圖來說明 AI 應該是什麼樣子、如何運作，以及最終應該為誰服務。因此
24:36.640–24:45.279
當 OpenAI 承擔起通用人工智慧的使命時，他們能夠根據對他們最方便的方式， essentially 塑造和塑造
24:45.279–24:52.159
他們希望這項技術成為的樣子。但當他們確立這一概念時，
24:52.159–24:59.279
科學家們真的輕視這個詞，甚至 AGI。因此他們只吸收了一小群
24:59.279–25:05.279
自我認同的 AGI 信仰者。這就是我稱之為準宗教的原因，因為沒有科學證據表明我們
25:05.279–25:11.919
這個信仰體系內的兩個派系，即 AGI 宗教的兩派。有些人認為 AGI 將為我們帶來
25:11.919–25:20.559
烏托邦，而有些人則認為 AGI 將毀滅全人類。兩派都相信這有可能發生。它即將到來。而且
25:20.559–25:27.520
因此他們認為必須由他們來掌控這項技術，而非將其民主化。而這正是
25:27.520–25:33.760
這最終引出了你的問題：在 OpenAI 的歷史中，當山姆·阿爾特曼被解僱又復聘時，究竟發生了什麼事。
25:33.760–25:39.840
在 OpenAI 內部，嬰兒潮世代與末日論者之間就誰才應該主導大局產生了許多衝突。
25:39.840–25:47.120
>> 嬰兒潮世代和末日論者。[笑聲] >> 那些認為它會帶來世界末日的人。烏托邦的嬰兒潮世代和那些
25:47.120–25:54.960
說它會毀滅人類。末日論者與他們 relentless 且具攻擊性地爭論著技術建構的速度
25:54.960–26:04.799
這項技術、多快釋出這項技術，以及最終奧爾特曼非常擅長對人們說他們想聽的話
26:04.799–26:10.960
對人們說他們想聽的話，並且如果他认为某些人需要聽不同的內容，他會對不同的人說不同的話
26:10.960–26:14.320
不同的說法。所以當我問嬰兒潮世代，奧特曼是嬰兒潮世代嗎？他們說
26:14.320–26:19.279
是的。當我問千禧世代，奧特曼是千禧世代嗎？他們說是的。而我想
26:19.279–26:28.799
這件事一直持續到今天，直到今年一月，特朗普政府宣布星門計畫，一個價值五千億美元的
26:28.799–26:39.520
千億美元計畫，旨在提升美國的人工智慧基礎建設。這是 OpenAI 的山姆·奧爾特曼與特朗普總統共同發表的講話。
26:39.520–26:45.919
我認為這將是本時代最重要的計畫，正如馬薩所說，為了在這裡建構通用人工智慧，創造數十萬個工作機會
26:45.919–26:51.830
創造數以萬計的就業機會，建立以這裡為中心的新產業。總統先生，沒有您我們無法做到這一點。
26:51.840–27:01.679
>> 他也在那裡提到了人工通用智能，也就是AGI。請解釋這裡發生了什麼，這究竟是什麼，以及它是否真的實現了。
27:01.679–27:14.080
事實上有沒有發生呢？所以在特朗普當選之前，奧爾特曼就已經透過觀察察覺到這種可能性
27:14.080–27:21.279
政府會轉向，而他必須開始大力進行政治遊說，以贏得新政府的青睞
27:21.279–27:30.480
政府。奧爾特曼非常具有策略性。當時他也承受著巨大壓力，因為他的共同創始人伊隆·馬斯克
27:30.480–27:38.240
現在馬斯克跟他有嚴重的糾紛。馬斯克覺得奧爾特曼利用他的名聲和資金來創立 OpenAI，然後他
27:38.240–27:46.080
一無所獲，所以 Musk 一直在對提告，現在卻突然成為特朗普新政府的親密夥伴
27:46.080–27:56.559
行政團隊。所以奧特曼基本上巧妙地策劃了這項公告，順便一提，這項公告相當奇怪
27:56.559–28:03.600
因為特朗普總統並非美國政府提供五千億美元，而是來自軟銀等地區的私人投資
28:03.600–28:13.840
進入美國，嗯，來自軟銀等地方，也就是，呃，也就是由孫正義經營的其中一個最大投資基金
28:13.840–28:22.240
孫正義是一位日本商人，他的財富主要來自上一個科技時代，所以提供這些資金的甚至不是美國政府
28:22.240–28:33.520
錢>>然後一路延續到現在的沙特阿拉伯之行，當時伊隆·馬斯克去了，山姆·阿爾特曼也去了，這讓伊隆·馬斯克非常憤怒
28:33.520–28:45.200
然後在阿布達比敲定了交易。對，雖然沒有伊隆·馬斯克參與，但談的是開放人工智能。沒錯，所以阿爾特曼繼續嘗試
28:45.200–28:59.919
並利用美國政府作為管道，以取得更多地點與更具權力的空間，來擴張這個帝國。而其中一項
28:59.919–29:06.720
因為 OpenAI 的運算基礎設施需求非常龐大。你知道，我有一位 OpenAI 的員工告訴我
29:06.720–29:13.360
我有一位 OpenAI 的員工告訴我，我們即將耗盡土地和電力資源，因此他們在美國面臨資源匱乏，這也是他們試圖
29:13.360–29:19.679
在其他地方取得土地和能源的存取權。中東擁有大量的土地和能源，並且他們願意進行交易，這就是為什麼奧特曼會參與那次行程，以尋求達成協議，而他們所達成的協議是，利用當地的土地和能源，在中東建造一座或數座大型資料中心。但最近開放人工智能推出的一項新計畫，他們稱之為「面向國家的開放人工智能」計畫
29:19.679–29:26.399
願意簽訂協議，這也是為什麼奧特曼參與了那次行程，旨在敲定協議，而他們所達成的協議
29:26.399–29:35.360
他們所簽署的協議，是要利用當地的土地和能源，在中東建設大型或多個數據中心
29:35.360–29:42.559
但 OpenAI 最近推出的一項計畫，他們稱之為 OpenAI for countries 計畫
29:42.559–29:53.440
而他們正是基於這個理念，希望在全球各地安裝 OpenAI 的硬體與軟體，並且明確地
29:53.440–30:05.039
表示我們希望建立民主的 AI 規範。我們希望在全球安裝我們的硬體和軟體，作為民主的基礎
30:05.039–30:15.919
全球部署AI，以便阻止中國在全球安裝專制AI。但他沒有提到的是
30:15.919–30:23.360
他沒有承認的是，他所做的毫無民主可言。你知道，大西洋雜誌執行編輯說
30:23.360–30:31.360
我們必須揭穿這些公司的真面目，它們是技術專制主義者，它們不會就技術開發、訓練數據以及數據中心的建設地點向公眾徵求任何意見
30:31.360–30:36.559
開發技術、使用哪些數據來訓練技術，以及這些數據中心設在哪裡。事實上
30:36.559–30:44.159
這些數據中心通常是在夜幕掩護下開發的，例如 Meta 最近透過殼公司進入新墨西哥州。
30:44.159–30:54.720
殼公司名為 Greater Kudu LLC。>> Greater Kudu LLC。一旦交易實際完成，居民
30:54.720–30:58.960
再也無能為力時，才揭露真相，驚喜的是，我們是 Meta，而你將獲得一座
30:58.960–31:04.320
耗盡你所有新鮮水源的數據中心。>> 然後在孟菲斯發生了一場關於數據中心
31:04.320–31:12.159
的爭議。>> 是的。這就是伊隆·馬斯克正在建造的數據中心。與此同時，
31:12.159–31:19.120
每個人都應該使用我的 AI。當然，他的 AI 也是同樣的環境和公共
31:19.120–31:25.039
健康成本下開發的。所以他在田納西州孟菲斯建造了一個名為 Colossus 的大型
31:25.039–31:36.880
超級電腦，用於訓練 Grock，人們可以透過 X 存取這個聊天機器人，而它是由
31:36.880–31:45.440
大約 35 部未獲許可的天然氣渦輪機驅動，向大孟菲斯社區排放數千噸有毒空氣污染物。而該
31:45.440–31:54.640
社區長期以來缺乏獲得清潔空氣的權利，這是一項基本人權。>> 所以我想轉向一個有趣的議題，Sam
31:54.640–32:03.279
Alman 在國會作證，討論人工智慧高能源消耗的解決方案。
32:03.279–32:09.360
>> 短期內，我認為這可能看起來像是更多的天然氣。嗯，雖然在某些應用場景中，我認為
32:09.360–32:16.000
太陽能確實能提供幫助。在中長期，我希望是先進核能、氫能和核融合。更多能源
32:16.000–32:24.320
的重要性遠超 AI 範疇。>> 所以這是 Open AI 的 Sam Alman。他在參議院作證，談論
32:24.320–32:31.600
從太陽能到核能等各種議題，這是在美國被環保人士數十年來爭論的話題
32:31.600–32:38.480
活動人士數十年來反對的議題。所以你有這些巨大的舊核電廠，但
32:38.480–32:45.279
許多人說，無論你把它們做得多小、多聰明，都無法保證它們的安全。>> 這是讓我擔憂的許多事情之一，關於
32:45.279–32:52.000
AI發展的當前軌跡，這是一種二階或三階效應，因為這些公司
32:52.000–32:59.679
試圖宣稱他們採取的 AI 開發方法沒有氣候危害。他們明確地
32:59.679–33:05.360
一次又一次地提及核能，因為核能將解決問題。而且這很有效。我與某些 AI 研究人員談過，他們認為
33:05.360–33:13.519
與某些認為核能已解決該問題的人工智慧研究人員進行交流，並且為了實際推動更多核電廠的建設，他們正在遊說政府試圖鬆綁核電廠建設的監管架構。
33:13.519–33:24.000
建造越來越多的核電廠，他們正在遊說政府試圖鬆綁核電廠建設的監管架構
33:24.000–33:32.159
核電廠建設的監管結構。我的意思是，這在許多層面上都瘋狂，他們不僅試圖
33:32.159–33:42.320
魯莽地開發這些 AI 技術。他們還試圖在「快速行動，打破常規」的意識形態下，
33:42.320–33:50.480
鋪設基礎設施和核基礎設施。但對於那些長期反對核能的環境主義者來說，他們會被
33:50.480–33:54.320
太陽能替代方案所吸引嗎？但事實正是如此
33:54.320–34:02.799
數據中心必須 24/7 運行。所以它們實際上不能僅靠再生能源運行。這就是為什麼公司不斷試圖
34:02.799–34:11.599
將核能塑造成萬靈丹，但當我們缺乏足夠的能源儲存解決方案時，太陽能實際上並不可行
34:11.599–34:18.879
解決方案來支持 24/7 運行時，太陽能實際上並不起作用。我們將在一分鐘後回到與 Karen Ho 的對話，她是新書
34:18.879–34:32.800
《AI 帝國：Sam Alman 的 Open AI 中的夢想與噩夢》的作者。請繼續收看。這是《Democracy Now!》。Democracynow.org。
34:32.800–34:42.480
我是 Amy Goodman。在這個假日特別節目中，我們正在採訪記者 Karen Ho，新書《AI 帝國：Sam Alman 的 Open AI 中的夢想與噩夢》的作者。
34:42.480–34:53.200
在薩姆·阿爾曼的開放人工智能中。她於五月進入我們的攝影棚。她住在香港。我請她談談中國在人工智慧領域正在發生的事情
34:53.200–35:02.079
人工智慧的發展情況。>> 中國和美國是最大的人工智慧研究樞紐。它們是全球最大
35:02.079–35:09.440
的人工智慧研究人才聚集地。嗯，除了矽谷之外，中國在人才密度方面確實是唯一的競爭對手，以及
35:09.440–35:14.800
資本投資的數量和進入 AI 開發的基礎設施數量。在最近幾
35:14.800–35:25.440
年裡，我們看到美國政府積極試圖保持第一，其中一種機制是出口管制。這些 AI
35:25.440–35:34.480
管制。這些 AI 模型的一個關鍵輸入是運算基礎設施，以及安裝在資料中心用於訓練這些模型的電腦晶片。為了開發 AI 模型，企業正在使用最先進的電腦晶片技術。這就像每兩年就會推出新款晶片，他們便立即開始使用該晶片來訓練下一代 AI 模型。這些電腦晶片由美國公司設計，其中最著名的是
35:34.480–35:43.040
訓練這些模型的電腦晶片。而這些電腦晶片是開發 AI 模型所必需的。公司正在使用最先進的電腦
35:43.040–35:48.640
晶片技術。就像每兩年出一款新晶片，他們立即開始使用它來訓練
35:48.640–35:55.760
下一代 AI 模型。這些電腦晶片由美國公司設計，其中最著名的是
35:55.760–36:03.920
加州的輝達（Nvidia）。因此，美國政府一直試圖利用出口管制，阻止中國公司
36:03.920–36:12.320
取得最尖端的電腦晶片。這一切都是基於矽谷的建議，說這是防止
36:12.320–36:23.119
中國成為第一名的方法。並對他們實施出口管制，完全不監管我們，讓我們保持第一名，而他們則落後。
36:23.119–36:32.800
實際發生的情況是，由於中國擁有強大的AI研究人才基礎，在計算資源較少的限制下，中國公司實際上已經能夠
36:32.800–36:41.920
創新並開發出與美國公司相同水平的AI模型能力，但計算資源少了兩個數量級
36:41.920–36:50.079
，更少的能源、更少的數據。所以，我特別指的是中國公司Highfly，它今年早些時候開發了這個
36:50.079–37:00.240
名為Deep Seek的模型，該模型曾短暫導致全球經濟震盪，因為該公司表示他們
37:00.240–37:09.040
訓練這一個AI模型的費用大約為600萬美元，而OpenAI訓練模型的費用數以億計，甚至超過數百億
37:09.040–37:21.680
美元。這種差異向人們證明，矽谷過去幾年試圖說服大家的觀點——即獲得更多AI能力的唯一途徑——是完全錯誤的，實際上，中國公司使用的技術
37:21.680–37:30.640
是文獻中已經存在的技術，只需進行組合。他們使用了大量的工程技巧來完成這項工作，但他們並沒有使用
37:30.640–37:39.040
根本性的新技術。他們使用的是實際上已經存在的技術。>> 所以請進一步解釋，因為我認為很多人無法理解這一點。
37:39.040–37:46.560
奇虎（Chi）使用的技術是文獻中已存在的，只需進行組合。他們投入了
37:46.560–37:52.800
>> 有一種稱為神經網絡的軟件，本質上是一個巨大的統計引擎，它進行大量複雜的統計計算，試圖確定數據集中存在哪些類型的模式。所以
37:52.800–37:58.720
通常，在我們轉向大型語言模型之前，它會做一些事情，比如查看MRI
37:58.720–38:02.550
掃描，並檢查癌症在MRI掃描中的模式。現在，對於GPT，它查看的是
38:02.560–38:11.280
>> 因此有一種稱為神經網路的軟體，它本質上是一個龐大的統計引擎，正在進行大量
38:11.280–38:19.599
結構、通常使用的修辭手法，然後它使用這些模式來構建新句子。這就是生成式AI
38:19.599–38:26.000
的工作原理。之所以計算成本如此高昂，是因為它在處理這些模式的數字。你輸入的數據越多，它需要處理的數字就越多。因此，我們過去訓練這些AI模型時，使用的是
38:26.000–38:34.560
你知道的強大筆記型電腦，也許只有一個電腦晶片，也許是像麻省理工學院這樣的富裕實驗室，他們會在幾個或幾十個電腦晶片上進行訓練，而像Google這樣的公司
38:34.560–38:38.320
可能會在幾百個電腦晶片上進行訓練。我們現在談論的是數十萬個電腦晶片訓練單一模型。
38:38.320–38:48.640
這就是OpenAI所說的構建這些技術所必需的，而Deep Seek證明了這是錯誤的。所以，讓我問你一個問題，Karen。你最近在美國旅行期間，
38:48.640–38:55.280
在返回香港之前，關於特朗普對學術界的攻擊，這如何契合。特朗普對
38:55.280–39:03.680
模式。你輸入的資料越多，它需要運算的數量就越大。因此，我們過去在訓練這些 AI 模型時，使用的是
39:03.680–39:12.640
，並撤銷他們的簽證，將如何影響AI行業的未來？但不僅是中國學生，因為現在發生的事情讓世界各地的學生感到恐懼，而且實驗室以各種方式關閉，美國
39:12.640–39:20.240
學生也決定出國。這只是美國政府過去幾年採取的最新行動，旨在真正疏遠美國創新的一個關鍵人才庫。最初，在美國工作的中國研究人員
39:20.240–39:27.680
對美國AI的貢獻比在中國的多，因為[噴鼻聲]就在幾年前，中國研究人員渴望在
39:27.680–39:37.119
美國公司工作。他們想搬到美國。他們想為美國經濟做出貢獻。他們不想回到自己的
39:37.119–39:45.119
祖國。但由於所謂的「中國行動計劃」，這是第一屆特朗普政府試圖將中國學者或華裔學者刑事化的行動，其中一些人是美國人，僅因
39:45.119–39:55.760
文件錯誤。他們指控他們是間諜。這是其中一項最初的行動。然後當然
39:55.760–40:04.240
疫情爆發，美中貿易升級開始加劇反華言論。所有這些都導致
40:04.240–40:11.839
以及撤銷他們的簽證，將如何影響人工智慧產業的未來，但受影響的不僅是中國學生，因為目前發生的事
40:11.839–40:21.280
現在的情況讓全球學生感到恐懼，因為實驗室正在以各種方式關閉，美國
40:21.280–40:29.920
以及決定出國留學的學生。這是美國政府過去幾年來為疏遠美國創新關鍵人才庫所採取的最新行動
40:29.920–40:42.480
近年來美國政府採取的這項最新舉措，旨在疏遠對美國創新至關重要的關鍵人才庫。最初，在美國工作的中國研究人員比在中國的多
40:42.480–40:52.160
對美國 AI 的貢獻比在中國還多，因為 [噴鼻息聲] 就在幾年前，中國研究人員還渴望在美國工作
40:52.160–41:02.960
為美國企業工作。他們希望移居美國，希望為美國經濟做出貢獻。他們不想回到祖國
41:02.960–41:11.760
國家。但由於所謂的中國行動計劃，這是特朗普第一任期內試圖
41:11.760–41:18.960
將中國學者或具有中國族裔背景的學者刑事化，其中一些學者實際上已是美國人，基於
41:18.960–41:26.480
只是文件上的疏失。他們會指控這些人是間諜。這是其中一項最早的行動。然後當然
41:26.480–41:34.720
疫情爆發後，美中貿易升級開始加劇反華言論。這些因素共同導致
41:34.720–41:44.800
而現在加上可能禁止國際學生入境，所有這些因素導致越來越多的中國研究人員選擇留在
41:44.800–41:54.319
國內，為中國的人工智慧生態系統做出貢獻。而這是High-fly成功推出Deepseek的前提條件。如果
41:54.319–42:03.280
中國沒有這樣集中和積累的人工智慧人才，他們在創新以繞過美國政府對他們實施的出口管制時，
42:03.280–42:11.040
在創新以繞過美國政府對他們實施的出口管制方面所花費的時間。但
42:11.040–42:21.680
因為他們現在擁有高度集中的人才，其中一些是全球頂尖人才，當這些限制措施實施時，
42:21.680–42:29.839
他們能夠在這些限制下進行創新。因此，Deepseek literally 是這種疏離感延續的產物，
42:29.839–42:38.160
隨著美國繼續採取這種態度，情況只會越來越糟。正如你所提到的，這不僅僅是中國研究人員的問題。我剛剛
42:38.160–42:45.680
和一位學術界的朋友聊過，她說她現在考慮去歐洲，因為她無法在沒有
42:45.680–42:52.400
公共資金的情況下生存。歐洲國家正看到一個關鍵機會，提供百萬美元
42:52.400–42:59.599
的套餐。來這裡，我們會給你實驗室，我們會給你數百萬美元的資金。我的意思是，這是這個國家人才外流最快的方式。>> 我的意思是，許多人說美國的人才外流就是他們的人才增益。是的。>> 這也讓我們想起歷史。你
42:59.599–43:09.839
這將是這個國家人才外流最快的方式。>> 我的意思是，許多人說美國的人才外流就是他們的人才流入。是的。>> 這也讓我們想起歷史。你
43:09.839–43:20.560
然後艾森豪威爾將他驅逐回中國。他成為火箭科學之父，以及中國進入太空領域的關鍵人物。
43:20.560–43:28.800
他說他永遠不會再踏足美國，儘管最初那裡是他唯一想居住的地方。>> 是的。還有一位我認為是
43:28.800–43:34.400
政府官員，一位美國政府官員說，這是美國犯過最愚蠢的錯誤。>> 嗯，我們談論人才外流和人才增益。好的。再次，嗯，一些更多的押韻，末日論者和
43:34.400–43:41.920
嬰兒潮一代。嗯，我想談談人工智慧末日看起來是什麼樣子，意味著它如何將我們帶向末日，但也
43:41.920–43:53.040
美國犯過最愚蠢的錯誤。我們談到人才外流和人才回流，好吧，再來一些押韻詞，末日派和嬰兒潮世代。
43:53.040–44:04.560
>> 這是一個很好的問題，我經常問嬰兒潮一代和末日論者這個問題。你能向我精確地闡述我們如何到達
44:04.560–44:13.349
那麼人們說它可能會帶我們走向烏托邦，這兩條軌跡是什麼呢
44:13.359–44:19.680
做不到。這就是為什麼我稱之為類宗教。它確實基於信念。我的意思是，我和一位自認為是嬰兒潮一代的研究人員談話。
44:19.680–44:21.200
我說，你知道，他的眼睛睜得大大的，他真的興奮起來說，你知道，一旦我們達到通用人工智慧（AGI），遊戲結束。一切都變得完美。
44:21.200–44:30.079
我問他，我說，你能解釋一下通用人工智慧如何餵養那些現在桌上沒有食物的人嗎？而他說，「哦，
44:30.079–44:39.359
你是在談論地板以及如何提升他們的生活質量。」我說，「是的，因為他們也是全人類的一部分。」
44:39.359–44:49.760
然後我問他，我說，你能不能解釋一下，AGI 怎麼餵飽那些現在連飯都吃不飽的人？然後他說，「喔，
44:49.760–44:59.680
會幫助中產階級獲得更多的經濟機會。」我說，「好的，但這也是如何發生的呢？」他說，「嗯，一旦
44:59.680–45:04.640
這些到來，一旦我們擁有通用人工智慧，它僅僅創造數萬億美元的經濟價值，我們就可以直接給他們
45:04.640–45:09.920
更多經濟機會。」然後我說：「好吧，但這是怎麼發生的呢？」他說：「嗯，一旦
45:09.920–45:15.839
這些情況會在我們擁有通用人工智慧且它能創造數兆美元經濟價值時出現，我們可以就此給予他們
45:15.839–45:19.119
就像當你實際測試他們的邏輯時，它並不完全成立。對於末日論者來說，我的意思是，這
45:19.119–45:20.560
是一樣的，他們的信念最終是我在報導這本書時意識到的，他們相信通用人工智慧是可能的
45:20.560–45:28.319
因為他們對人類大腦運作方式的信念。他們相信人類智能本質上是完全計算性的。所以如果你擁有足夠的
45:28.319–45:37.920
數據和足夠的計算資源，你最終將能夠重建人類智能。這只是時間問題。對他們來說，
45:37.920–45:46.480
導致末日場景的原因是人類，我們通過溝通來學習和提高我們的智能。而溝通是低效的。我們經常誤解彼此。
45:46.480–45:53.040
因此對於人工智慧智能來說，它們將能夠通過作為數字智能彼此之間完美的溝通，迅速變得更聰明、更聰明、更聰明。所以
45:53.040–46:01.200
許多自認為是末日論者的人說，在宇宙歷史上，從未有一個物種
46:01.200–46:07.920
優於另一個物種，一個能夠統治更優越物種的物種。所以他們認為最終人工智慧將進化為更高級的物種，然後開始統治我們
46:07.920–46:19.280
然後也許決定徹底消滅我們。我在想你是否能談論任何國家的模型，而不是
46:19.280–46:26.800
作為數位智能彼此之間進行完美溝通。因此，許多自稱為末日論者的人表示，在宇宙歷史中，從未出現過一個優越於其他物種、能夠統治更優越物種的物種。所以他們認為，人工智能最終會進化成為一個更高級的物種，然後開始統治我們，接著可能決定將我們徹底消滅。我在想，你是否能談論任何國家的模型，而非
46:26.800–46:36.640
公司，>> 正在開創民主控制人工智慧的方式。>> 我不認為這正在積極發生
46:36.640–46:45.280
現在。歐盟已經有了歐盟人工智慧法案，這是他們主要的立法，試圖開發基於風險、基於權利
46:45.280–46:54.319
的人工智慧部署治理框架。但對我來說，民主人工智慧治理的關鍵之一也是民主地
46:54.319–47:05.599
公司，>> 正在開創一種民主控制人工智慧的途徑。>> 我不認為這正在積極發生
47:05.599–47:15.599
現在。歐盟已有歐盟人工智能法案，這是他們主要的立法，試圖開發基於風險、基於權利
47:15.599–47:27.760
的框架來監管人工智慧部署。但對我來說，民主人工智慧治理的關鍵之一也是民主地
47:27.760–47:38.480
開發人工智慧。我不認為任何國家真正在做這件事。我的意思是，人工智慧有供應鏈。它需要數據。它需要
47:38.480–47:45.599
土地。它需要能源。它需要水。它也需要空間，讓這些公司能夠進入並部署
47:45.599–47:52.240
他們的技術。學校、醫院、政府機構。矽谷過去十年做得很好，讓人們覺得他們集體擁有的資源是矽谷的。你知道，我經常與
47:52.240–47:59.599
朋友談話，他們說：「我們不再有數據隱私了。」所以，像什麼，什麼是更多數據給
47:59.599–48:05.920
我時常和朋友聊天，他們說：「我們已經沒有資料隱私了。」所以，也就是說，對這些公司來說，更多的資料意味著什麼？
48:05.920–48:07.359
把他們所有的數據都給他們。但這些數據是你的。你知道，那些知識產權是作家和
48:07.359–48:16.560
藝術家的知識產權。那些土地是社區的土地。那些學校是學生和老師的學校。那些醫院是醫生、護士和
48:16.560–48:24.480
是社區的土地。那些學校是學生和老師的學校。醫院是醫生、護士和
48:24.480–48:32.640
醫院是病患的醫院。這些都是民主爭奪的場域，在人工智慧的部署、開發與部署中。
48:32.640–48:38.720
部署中。就像我們談到的那些智利水資源活動人士，他們強烈地理解
48:38.720–48:47.680
淡水是他們的，他們不願意放棄，除非他們獲得某種互惠互利的協議。我們需要
48:47.680–49:00.559
在保護我們的數據、土地、水和學校時擁有這種精神，這樣公司最終將不得不調整他們的方法
49:00.559–49:07.200
因為他們將不再獲得他們需要的資源或他們需要部署的空間。在2022年，
49:07.200–49:14.079
卡倫，你為《麻省理工技術評論》寫了一篇標題為「為人民的人工智慧新願景」的文章。在新
49:14.079–49:22.160
西蘭偏遠的鄉村小鎮，一對原住民夫婦正在挑戰人工智慧可能成為什麼以及它應該服務誰。他們是誰？這是一篇
49:22.160–49:24.640
精彩的報導，我做的，這對夫婦經營著媒體。它是新西蘭的一家非營利性毛利語廣播電台。而毛利人已經
49:24.640–49:36.400
這是一個很棒的報導，我做了這個報導，那對夫婦他們經營一家非營利的毛利語廣播電台，位於紐西蘭。而毛利人經歷了許多與全球其他原住民族相同的挑戰。
49:36.400–49:43.440
他們迅速失去語言，現在世界上說毛利語的人非常少。所以在過去幾年裡
49:43.440–49:49.200
年，有人試圖復興這種語言，新西蘭政府試圖通過嘗試鼓勵這種語言的復興來贖罪。但這家非營利性廣播電台，他們擁有所有這些精彩的
49:49.200–49:54.960
近年來已有人嘗試復興該語言，而紐西蘭政府也試圖透過努力促進該語言的復興來贖罪
49:54.960–50:02.079
世界各地的毛利語說者、毛利語學習者作為教育資源。問題是，為了做到這一點，他們
50:02.079–50:10.079
檔案資料，祖先說馬語的音檔，他們希望提供給
50:10.079–50:16.319
全球毛利語說客和學習者作為教育資源。問題在於，為了做到這一點，他們
50:16.319–50:22.800
理解翻譯，並實際上將其變成一個積極的學習工具。但說這種高級別毛利語的人太少了，以至於
50:22.800–50:29.760
他們意識到自己必須轉向人工智慧。這是我書中論點的一個關鍵部分，我並非批評所有人工智慧
50:29.760–50:39.200
能達到那個進階程度的茂宜語使用者少之又少，以致於他們不得不轉向人工智慧，而這也是我書中論點的核心：我並非批判所有的人工智慧
50:39.200–50:44.800
許多不同類型的有益人工智慧模型，包括他們最終所做的。所以他們採取了根本不同的方法。首先，他們詢問社區，我們想要這個
50:44.800–50:50.880
許多不同類型的有益人工智慧模型，包括他們最終所採取的行動。因此，他們採取了根本不同的方法。
50:50.880–50:57.760
是，然後他們進入下一步，要求人們完全同意捐贈數據以訓練這個
50:57.760–51:06.480
工具。他們向社區解釋了這些數據的用途、如何使用，以及他們將如何保護該
51:06.480–51:12.559
數據，確保它不會被用於其他目的。他們在短短幾天內收集了大約幾百小時的音頻數據
51:12.559–51:20.640
並確保資料受到保護，且未被用於其他目的。他們在短短幾天內就收集了約數百小時的音訊資料
51:20.640–51:27.119
因為社群為這個專案凝聚了支持，僅需幾百小時的數據就足以建立一個
51:27.119–51:33.040
所需的數據規模時是瘋狂的。而這再次是一個可以吸取的教訓，實際上有很多研究
51:33.040–51:39.200
表明，當你擁有高度策劃的小數據集時，你實際上可以創建非常強大的人工智慧模型，然後一旦他們
51:39.200–51:46.800
有了這個工具，他們就能夠做他們想做的：開源並向他們的社區提供這個教育資源。
51:46.800–51:55.040
有了這個工具後，他們得以完全按照自己的意願，將這項教育資源以開源方式分享給社群。
51:55.040–52:08.800
因此，我對未來人工智慧發展的願景是擁有更多針對特定小型任務的人工智慧模型，它們並非在龐大且受污染的數據集上訓練而成，而是
52:08.800–52:18.240
小型精心策劃的資料集，因此只需少量運算能力，且能部署於我們人類亟需克服的挑戰中
52:18.240–52:27.920
我們人類真正需要解決的問題，包括透過將更多再生能源整合進電網來緩解氣候變遷，
52:27.920–52:37.599
改善醫療保健，透過進行更多藥物研發。>> 所以，當我們終於要結束時，你感到最震驚的是什麼，你一直在
52:37.599–52:49.599
這些年來你從事新聞與研究工作，在撰寫《AI帝國》時，最讓你震驚的是什麼？我原本以為我會寫一本聚焦於AI供應鏈垂直危害的書籍，這裡展示了AI產業中的勞動力剝削如何發生，這裡展示了AI產業如何產生環境危害，而在我的報導結束時，我意識到這裡存在著一種橫向危害，我採訪過的每個社區，無論是藝術家
52:49.599–52:56.800
撰寫一本聚焦於人工智慧供應鏈垂直危害的書籍。以下是人工智慧產業中勞動力剝削的運作方式。
52:56.800–53:04.559
這就是環境危害如何從人工智慧產業中產生的方式。而在我的報導結束時，我意識到
53:04.559–53:10.319
這裡存在一種橫向的傷害，在我採訪過的每個社區中，無論是藝術家
53:10.319–53:16.800
知識產權被剝奪的藝術家，或是新鮮水源被奪取的智利水資源活動人士，他們都說
53:16.800–53:24.640
當他們遇到這個帝國時，最初感受到的都是同樣的情緒，一種完全喪失自主權的無力感
53:24.640–53:33.839
自主決定自己的未來。就在這時我意識到這裡存在的橫向傷害是人工智慧正在威脅民主。如果
53:33.839–53:46.240
世界大多數的人都會感受到這種喪失自主決定未來能力的無力感，而民主將無法存续，特別是針對
53:46.240–53:55.599
矽谷的開發策略是無所不用其極地追求規模。>> 但你同時也記錄了抵抗的行動。你談到智利的水資源
53:55.599–54:05.680
他們最初有何感受，藝術家們最初又有何感受，請談談這些人採用的策略，以及這些策略是否有效。
54:05.680–54:12.400
>> 所以令人驚訝的是，此後出現了如此強烈的反彈。藝術家們接著說，等等，我們
54:12.400–54:18.640
可以起訴這些公司。智利的水資源活動人士表示，我們可以反擊並保護這些水資源。
54:18.640–54:25.200
資源。我採訪過的肯亞工人由 OpenAI 僱用，他們表示我們可以組織工會並升級
54:25.200–54:35.680
將我們的故事推向國際媒體關注。因此，即使在我認為這些社區可能是最脆弱的時候，
54:35.680–54:45.040
在全球擁有最少自主權的人群中，正是他們記起自己確實擁有自主權，並且能夠掌握這份自主權
54:45.040–54:54.960
反擊。我認為，遇到這些人並讓我想起其實這是非常令人振奮的
54:54.960–55:02.960
奪回民主的第一步，就是記住沒有人能奪走你的自主權。>> 卡倫·霍，新書《帝國：人工智能的夢想與夢魘》作者
55:02.960–55:11.680
人工智慧的夢與夢魘，以及山姆·阿爾特曼的開放人工智慧。凱倫·豪曾是華爾街日報的記者，並
55:11.680–55:19.760
麻省理工科技評論。本期特別節目到此結束。[音樂]民主新聞由麥克·伯克、蕾妮·費爾、迪娜·古斯特、梅西婭·羅德斯、努爾明·沙克、瑪麗亞·特雷塞納、妮可·薩拉薩、[音樂]薩林·納賽爾、特里娜·納杜拉、薩姆·阿洛夫、T·瑪麗亞·喬、約翰·漢密爾頓、羅比、[音樂]卡倫、霍尼·馬蘇德和薩夫韋特·納茲共同製作。我們的執行總監是茱莉·克羅斯比。特別感謝貝卡·斯蒂利、約翰·蘭道夫、保羅·鮑威爾、[音樂]麥克
55:19.760–55:25.040
伯克、蕾妮·費爾、迪娜·古斯特、梅西婭·羅德斯、努爾明·沙克、瑪麗亞·特拉塞納、妮可·薩拉薩、[音樂] 薩林·納賽爾、
55:25.040–55:31.599
崔娜·納杜拉、薩姆·阿洛夫、T·瑪麗亞·喬、約翰·漢密爾頓、羅比 [音樂] 凱倫、哈尼·馬蘇德和薩夫韋特·納茲。我們的
55:31.599–55:37.280
執行董事茱莉·克羅斯比。特別感謝貝卡·斯特利、約翰·蘭道夫、保羅·鮑威爾、[音樂] 麥克
55:37.280–55:44.160
德菲波、米格爾·恩格拉、休·葛蘭、卡爾·馬克瑟、丹尼斯·莫漢、大衛·普魯德、丹尼斯·麥考馬克、馬特·埃萊利、安娜
55:44.160–55:52.799
奧斯貝克、艾米麗·安德森、但丁·托里耶里、巴菲 [音樂] 聖·瑪麗·埃爾南德斯與胡安·岡薩雷斯。我是艾米·古德曼。新年快樂
55:52.799–56:01.920
新年快樂。感謝收看 Democracy Now 的 YouTube 頻道。訂閱頻道並開啟通知，確保您
56:01.920–56:08.559
絕不錯過任何影片。若想支持更多由觀眾資助的我們的新聞報導，請前往democracynow.org或
56:08.559–56:15.000
下載我們的新聞應用程式、訂閱我們的電子報、訂閱每日播客，還有更多
0:03.200–0:12.320
This is democracyow [music]democracynow.orgthe War and Peace Report. I'm AmyGoodman. The Empire of AI, that's the
這是民主現在 [音樂] 民主現在.org 的戰爭與和平報導。我是艾美·古德曼。AI帝國，那是
0:12.320–0:20.320
name of a new book by journalist KarenHow, who's been closely reporting on therise of the artificial intelligence
記者Karen How新書的名稱，她一直密切報導人工智慧產業的崛起，並聚焦於Sam Altman的Open AI。
0:20.320–0:31.359
industry with a focus on Sam Alman'sOpen AI. That's the company behind ChatGPT. Karen how compares the actions of
那就是ChatGPT背後的企業。Karen How將這些行動與標誌著這段歷史的同樣公開暴力與殘酷相比較，但他們也掠奪並提取珍貴資源以
0:31.359–0:40.320
the AI industry to those of colonialpowers of the past. She writes, quote,"The empires of AI are not engaged in
開放AI。那就是ChatGPT背後的企業。卡倫·豪將
0:40.320–0:48.399
the same overt violence and brutalitythat marked this history, but they tooseize and extract precious resources to
餵養他們對人工智慧的願景。藝術家與作家的作品、無數個人發布關於他們
0:48.399–0:56.879
feed their vision of artificialintelligence. the work of artists andwriters, the data of countlessindividuals posting about their
超級電腦的數據。她寫道：「過去一年來，川普政府日益擁抱人工智慧產業。
0:56.879–1:05.920
experiences and observations online, theland, energy, and water required tohouse and run massive data centers and
同樣顯著的暴力與殘酷行為，這些行為標誌著這段歷史，但它們同樣
1:05.920–1:15.520
supercomputers. She writes, "Over thepast year, the Trump administration hasincreasingly embraced the AI industry.
超級電腦。她寫道：「過去一年來，川普政府越來越擁抱人工智慧產業。
1:15.520–1:22.880
In December, Trump signed an executiveorder to bar states and localgovernments from enacting their own AI
餵養它們對人工智慧的願景。藝術家與作家的作品、無數個人
1:22.880–1:32.000
regulations. Soon after he signed theorder, his family's company, Trump Mediand Technology, announced a $6 billion
監管規定。簽署行政命令後不久，其家族企業Trump Media and Technology便宣布以60億美元
1:32.000–1:40.880
merger with a firm aiming to build theworld's first viable nuclear fusionplant to power AI projects. Karen How is
與一家旨在建造全球首座可行核融合發電廠以供應人工智慧專案電力的公司合併。Karen How是
1:40.880–1:49.040
a former reporter at the Wall StreetJournal and MIT Technology Review whereshe became the first journalist to
超級電腦所需的土地、能源與水資源。她寫道：「過去一年來，特朗普政府日益擁抱AI產業。
1:49.040–1:58.159
profile open AI. Democracy Now's JuanGonzalez and I spoke to her in May. TheNational Book Critic Circle recently
十二月，特朗普簽署行政命令，禁止各州與地方政府制定自己的AI
1:58.159–2:08.000
named her book The Empire of AI: Dreamsand Nightmares and Sam Alman's Open AIas a finalist for best non-fiction book
監管規定。簽署命令後不久，其家族企業特朗普媒體與科技宣布與一家旨在建造
2:08.000–2:20.160
of 2025.I began by asking Karen How to explainjust what artificial intelligence is. SoAI is a collection of many different
全球首座可行核融合發電廠以供應AI專案的企業進行60億美元
2:20.160–2:27.680
echnologies but most people wereintroduced to it through chatbt and whatI argue in the book and what the title
技術，但大多數人是透過ChatGPT接觸到它的，而我在書中提出的論點以及書名
2:27.680–2:34.160
refers to empire of AI it's actually acritique of the specific trajectory ofAI development that led us to chatbt and
《華爾街日報》與《麻省理工科技評論》的前記者，在那裡她成為首位
2:34.160–2:42.720
has continued since chatbt and that ispecifically Silicon Valley's scale atall costs approach to AI development AI
為開放AI撰寫人物專訪的記者。民主現在新聞的胡安·岡薩雷斯與我在五月採訪了她。國家書評人協會最近
2:42.720–2:50.480
models in modern day they are trained ondata they need computers to train themon that data. But what Silicon Valley
將她的書《AI帝國：山姆·阿爾特曼開放AI的夢想與夢魘》評為
2:50.480–2:57.760
did and what OpenAI did in the last fewyears is they started blowing up theamount of data and the the size of the
2025年最佳非小說類書籍的決選作品。我首先請卡倫·豪解釋
2:57.760–3:03.840
computers that need to do this training.So we are talking about the full Englishlanguage internet being fed into these
需要進行這些訓練的電腦。所以我們談的是將整個英文互聯網的內容輸入到這些
3:03.840–3:10.720
models, books, scientific articles, allof the intellectual property that isbeing created and also massive
技術的集合，但大多數人是透過ChatGPT接觸到它的，而我在書中論述的以及書名
3:10.720–3:17.040
supercomputers that run tens ofthousands even hundreds of thousands ofcomputer chips that are the size of
所指的AI帝國，實際上是对導致我們走向ChatGPT的特定AI發展軌跡的批判，並且
3:17.040–3:25.360
dozens maybe hundreds of football fieldsand use practically the entire energydemands of cities now. So this is an
自ChatGPT以來持續至今，這特別是指矽谷在AI發展上不惜代價追求規模的
3:25.360–3:31.840
extraordinary um type of AI developmenthat is causing a lot of social, laborand environmental harms and that is
非凡的、呃，某種類型的 AI 開發，這造成了許多社會、勞動和環境上的危害，而且這正是
3:31.840–3:40.480
ultimately why I evoke this analogy toempire.>> And Karen, could you talk some moreabout not only the energy requirements
模型是在數據上訓練的，它們需要電腦來在這些數據上進行訓練。但矽谷
3:40.480–3:50.000
but the water requirements of these hugedata centers that are essence in essencethe backbone of of this uh widening uh
在過去幾年裡所做的事，以及開放AI所做的事，是他們開始大幅增加數據量以及進行此類訓練所需電腦的
3:50.000–3:51.030
industry?
規模。所以我們談論的是將整個英文互聯網的內容餵入這些
3:51.040–3:56.560
>> Absolutely. I'll give you two stats onboth the energy and the water. Whentalking about the energy demand,
模型，包括書籍、科學文章、所有正在創作的知識產權，以及運行數萬甚至數十萬顆
3:56.560–4:03.920
McKenzie recently came out with a reporthat said in the next 5 years based onthe current pace of AI computational
晶片的大型超級電腦，這些晶片的面積可能相當於數十個甚至數百個足球場，並且現在實際上佔用了城市的全部能源需求。所以這是
4:03.920–4:13.040
infrastructure expansion, we would needto put as much energy on the global gridas what is consumed by two to six times
一種非凡的AI發展類型，造成了許多社會、勞工和環境危害，而這正是
4:13.040–4:22.479
the energy consumed annually by thestate of California. And that willmostly be serviced by fossil fuels.We're already seeing reporting of uh
我借用帝國這個類比的根本原因。>> 卡倫，你能否再多談談這些龐大資料中心不僅是能源需求
4:22.479–4:29.120
coal plants with their lives beingextended. They were supposed to retirebut now they cannot to support this data
，還有水資源需求，這些資料中心本質上正是這個不斷擴張產業的骨幹？
4:29.120–4:36.800
center development. We are seeingmethane gas turbines, unlicensed onesbeing popped up to uh service these data
>> 絕對可以。我提供兩項關於能源與水資源的數據。談到能源需求時，
4:36.800–4:44.080
centers as well. From a freshwaterperspective, these data centers need tobe trained on freshwater. They cannot be
麥肯錫最近發布一份報告指出，根據目前AI運算基礎設施擴張的速度，在未來五年內，
4:44.080–4:50.000
trained on any other type of waterbecause it can corrode the equipment. Itcan lead to bacterial growth and most of
我們需要在全球電網上架設的能源量，將相當於加州年度能源消耗量的兩到六倍。而且這主要將由化石燃料供應。我們已經看到報導指出，
4:50.000–4:58.080
the time it actually taps directly intoa public drinking water supply becausethat is the infrastructure that has been
煤電廠的壽命被延長。它們本應退役，但現在為了支援這些資料
4:58.080–5:05.919
laid to deliver this clean fresh waterto different businesses to differenthomes. And Bloomberg recently had an
中心的發展，它們無法退役。我們也看到未經許可的天然氣燃氣輪機被安裝，以服務這些資料
5:05.919–5:13.680
alysis where they looked at thexpansion of these data centers aroundthe world and 2thirds of them are being
中心。從淡水角度來看，這些資料中心需要淡水進行訓練。它們不能
5:13.680–5:21.840
placed in water scarce areas. So they'rebeing placed in communities that do nothave access to fresh water. So it's not
使用其他類型的水，因為它會腐蝕設備。它可能導致細菌生長，而且大多數時候它實際上直接接入公共飲用水供應，因為這是已經鋪設好的基礎設施，用於向不同企業和不同家庭提供乾淨的淡水。而彭博最近進行了一項
5:21.840–5:27.520
just the total amount of fresh waterthat we need to be concerned about, butactually the distribution of this
不僅是我們需要關注的淡水總量問題，更是這些淡水在全球的分配問題。
5:27.520–5:37.840
infrastructure around the world.And most people are familiar with chatGPT the consumer aspect of AI but what
基礎設施在全球的分布情況。大多數人熟悉的是 ChatGPT 這種面向消費者的 AI，但
5:37.840–5:45.600
about the military uh aspect of AI wherein essence uh we're finding SiliconValley companies becoming the next
AI 的軍事層面呢？基本上，我們發現矽谷公司正成為下一代
5:45.600–5:53.199
generation of defense contractors.>> One of the reasons why OpenAI and manyother companies are turning to the
國防承包商。>> OpenAI 和許多其他公司轉向國防產業的原因之一是，
5:53.199–5:59.600
defense industry is because they havespent an extraordinary amount of moneyin developing these technologies.
國防工業的原因，因為他們在開發這些技術上花費了巨額資金。
5:59.600–6:07.520
They're spending hundreds of billions totrain these models and they need torecoup those costs and there are only so
它們花費數千億美元來訓練這些模型，並且需要收回這些成本，而能夠支付如此龐大金額的行業和地區有限。因此，
6:07.520–6:15.360
many industries and so many places thathave that size of a paycheck to pay. Andso that's why we're seeing a cozying up
我們看到它們與國防產業的關係日益緊密。我們也看到矽谷利用美國政府來實現其帝國擴張的野心。你可以說美國
6:15.360–6:22.720
to the defense industry. We're alsoseeing Silicon Valley use the USgovernment in their empire buildingambitions. You could argue that the US
轉向國防工業。我們也看到矽谷利用美國政府來實現他們建立帝國的野心。你可以說美國
6:22.720–6:29.039
government is also trying to use SiliconValley vice versa in their empirebuilding ambitions. Um but certainly
政府也在試圖利用矽谷，反過來實現他們建立帝國的野心。呃，但顯然
6:29.039–6:39.039
these technologies are not they are notdesigned to be used in a sensitivemilitary context. And so the theaggressive push of these companies to
試圖獲取國防合約，並將其技術越來越深入地整合到軍事基礎設施中，這確實令人擔憂。我想談談你去過的國家或你報導的故事，因為
6:39.039–6:44.800
try and get those defense contracts andintegrate their technologies more andmore to the into the infrastructure of
這太驚人了，你從肯亞到烏拉圭再到智利的報導深度。嗯，你談到了水的使用，我也想問你關於核能的問題。嗯，但在智利，
6:44.800–6:52.560
the military is really alarming. Iwanted to go to the countries you wento or the stories you covered because I
這些數據中心及其使用的水資源以及對此的抵抗情況如何？>> 是的。智利有一段有趣的歷史，因為它曾
6:52.560–7:00.240
mean this is amazing the depth of youreporting from Kenya to Uruguay toChile. Um you were talking about the use
長期處於獨裁統治之下。在那段時期，大多數公共資源被私有化，包括水資源。但由於一個
7:00.240–7:08.240
of water and I also want to ask youabout nuclear power. Uh but in Chile umwhat is happening there around these
例外情況，大聖地亞哥都會區有一個社區實際上仍然擁有公共淡水資源，該資源在緊急情況下既服務該社區也服務全國其餘地區。
7:08.240–7:15.520
data centers and the water they woulduse and the resistance to that.>> Yeah. So Chile has an interestinghistory in that it's been under it was
這正是 Google 選擇嘗試建立數據中心的社區。>> 並且它是免費的>> 而且，你知道，我完全不知道。這
7:15.520–7:25.440
under a dictatorship for a very longtime. And so during that time, mostpublic resources were privatized,including water. But because of an
在獨裁統治下很長一段時間。所以在這段時間內，大多數公共資源被私有化，包括水資源。但由於
7:25.440–7:31.440
omaly, there's one community in thegreater Santiago metropolitan regionthat actually still has access to a
根據閱讀文件後的信念，認為 Google 繳納的稅款是繳給他們註冊辦公室、行政辦公室所在地的，而不是他們建立數據中心的地方。所以
7:31.440–7:38.960
public freshwater resource that servicesboth that community as well as the restof the country in emergency situations.
他們沒有從這個數據中心獲得任何直接利益，也沒有看到對該數據中心允許提取的淡水有任何限制。因此，這些活動人士說，
7:38.960–7:48.479
That is the exact community that Googlechose to try to put a data center in>> and it would be free>> and it, you know, I have no idea. That
「等等，絕對不行。我們不會允許這個數據中心進入，除非他們給我們一個合理的理由，說明它如何使我們受益。」於是，
7:48.479–7:56.319
is a great question. But what thecommunity told me was they weren't evenpaying taxes for this because they they
這是一個很好的問題。但社區告訴我的是，他們甚至沒有為此繳納稅款，因為他們
7:56.319–8:01.520
believed based on reading thedocumentation that the taxes that Googlewas paying was in fact to where they had
每一個鄰居的門，向社區散發傳單，告訴他們這家公司正在拿走我們的淡水資源，卻沒有給予任何回報。因此，他們的行動如此激烈地升級，
8:01.520–8:08.479
registered their offices, theiradministrative offices, not where theywere putting down the data center. So
甚至升級到了 Google 智利分公司。然後升級到了 Google 山景城總部，順便說一句，後者隨後派出了只會說英語的代表前往智利。
8:08.479–8:14.240
they were not seeing any benefit fromthis data center directly to thatcommunity and they were seeing no checks
但最終，這升級到了智利政府。現在，智利政府舉行圓桌會議，邀請
8:14.240–8:20.960
placed on the fresh water that this datacenter would have been allowed toextract. And so these activists said,
對新鮮水資源施加的限制，這個數據中心被允許抽取的水量。所以這些活動人士說，
8:20.960–8:27.599
"Wait a minute, absolutely not. We'renot going to allow this data center tocome in unless they give us a legitimate
共同討論如何使數據中心的發展對社區更加有益。活動人士表示，戰鬥尚未結束。僅僅因為他們被邀請到談判桌旁，並不意味著一切都突然變好。他們需要
8:27.599–8:35.440
reason for why it benefits us." And sothey started doing boots on the groundactivism, pushing back, knocking on
保持警惕。他們需要繼續嚴格審查這些項目。但到目前為止，他們已經能夠阻止這個
8:35.440–8:40.399
every single one of their neighborsdoors, handing out flyers to thecommunity, telling them this company is
項目長達四到五年，並獲得了坐在談判桌旁的席位。>> 那麼，這些西方公司如何
8:40.399–8:46.800
taking our freshwater resources withoutgiving us anything in return. And sothey escalated so dramatically that it
正在未經回報的情況下奪取我們的淡水資源。因此他們大規模升級行動，
8:46.800–8:51.839
escalated to Google Chile. It escalatedto Google Mountain View, which by theway then sent representatives to Chile
升級到了智利谷歌。接著升級到了加州山景城的谷歌總部，順便一提，谷歌總部隨後派遣了
8:51.839–9:01.680
that only spoke English.But then it eventually escalated to theChilean government. And the Chileangovernment now has roundts where they
只會說英文的代表前往智利。但最終還是升級到了智利政府。而智利政府現在定期舉行會議，
9:01.680–9:06.160
ask these community residents and thecompany representatives andrepresentatives from the government to
互聯網上抓取來的，你不能直接將這些數據輸入到你的 AI 模型中，因為數據中包含很多
9:06.160–9:12.640
come together to actually discuss how tomake data center development morebeneficial to the community. The
東西。數據嚴重受污染。需要進行清理。需要進行標註。所以這就是
9:12.640–9:19.920
activists say it is the fight is notover. Just because they've been invitedto the table doesn't mean thateverything is suddenly better. They need
活動人士表示，這場戰鬥尚未結束。僅僅被邀請到談判桌上，並不意味著一切都突然變好了。他們需要
9:19.920–9:25.519
to stay vigilant. They need to continuescrutinizing these projects. But thusfar they've been able to block this
保持警惕。他們需要繼續嚴格審查這些專案。但到目前為止，他們已成功阻止該
9:25.519–9:36.640
project for four to five years and havegained that seat at the table.>> And how is it that these uh westerncompanies in essence are exploiting
專案長達四到五年，並獲得了談判桌上的席位。>> 那麼，這些西方公司本質上是如何剝削
9:36.640–9:40.240
labor in the global south? You go into
全球南方的勞工呢？你進入
9:40.240–9:44.640
something called data annotation firms.What what are those?
一種稱為資料標註公司的地方。那是什麼？
9:44.640–9:51.279
>> Yeah. So because AI modern day AIsystems are trained on massive amountsof data and they're sc that's scraped
是的。因為現代的人工智慧系統是透過大量數據進行訓練，這些數據是從網路抓取而來的，
9:51.279–9:59.279
from the internet, you can't actuallypump that data directly into your AImodel because there are a lot of things
你無法直接將這些數據輸入到人工智慧模型中，因為數據中包含許多
9:59.279–10:05.279
within within that data. It's heavilypolluted. It needs to be cleaned. Itneeds to be annotated. So this is where
在那些數據中，它受到嚴重污染，需要進行清理和標註，這就是
10:05.279–10:14.160
data annotation firms come in. These aremiddleman firms that hire contract laborto provide to these AI companies to do
資料標註公司就此介入。這些是中間商公司，雇用合約勞工來為這些 AI 公司提供服務，以進行
10:14.160–10:21.600
that kind of data preparation. And openAI when it was starting to think aboutcommercializing its products and
這類型的資料準備工作。而當 OpenAI 開始考慮將其產品商業化，
10:21.600–10:28.880
thinking about let's put text generationmachines that can spew any kind of textinto the hands of millions of users.
並思考讓我們將能生成各種類型文字的文本生成機器交到數百萬用戶手中時。
10:28.880–10:34.880
They realized they needed to have somekind of content moderation. They neededto develop a filter that would wrap
他們意識到需要某種內容審核機制。他們需要開發一種過濾器，將這些模型包裹起來，
10:34.880–10:42.320
around these models and prevent thesemodels from actually spewing racist,hateful, and harmful speech to users
以防止這些模型實際上向用戶噴出種族歧視、仇恨和有害的言論，
10:42.320–10:48.640
that would not make a very goodcommercially viable product. And so theycontracted these middleman firms in
否則這將無法成為一個非常具有商業可行性的產品。因此，他們與這些中間商公司簽約，
10:48.640–10:56.320
Kenya where the Kenyan workers had toread through reams of the worst text onthe internet as well as AI generated
在肯亞，肯亞工人必須閱讀大量網際網路上的最糟文字，以及 AI 生成的
10:56.320–11:03.600
text where open AI was prompting its ownAI models to imagine the worst text onthe internet and then telling these
文字，其中 OpenAI 提示其自己的 AI 模型去想像網際網路上的最糟文字，然後告訴這些
11:03.600–11:11.839
Kenyan workers to detail to categorizethem in detailed taxonomies of is thisexual content, is this violent content,
肯亞工人詳細地將它們分類到詳細的分類法中，例如這是色情內容嗎，這是暴力內容嗎，
11:11.839–11:18.480
how graphic is that violent content inorder to teach its filter all thedifferent categories of content it had
該暴力內容有多麼具視覺衝擊力，以便教導其過濾器所有它必須
11:18.480–11:25.120
to block. And this is incrediblyuncommon form of labor. There are lotsof other different types of contract
封鎖的不同類型內容。這是一種極其不尋常的勞動形式。他們使用了許多其他類型的合約
11:25.120–11:32.480
labor that they use. But these workers,they're paid a few bucks an hour, if atall. And just like the era of social
勞工。但這些工人，如果有的話，每小時只賺幾美元。就像社交媒體時代一樣，
11:32.480–11:40.079
media, these content moderators are leftvery deeply psychologically traumatized.And ultimately, there is no real
媒體、這些內容審核員所面臨的情況，他們在心理上遭受了極深的創傷。而最終，並沒有真正的
11:40.079–11:46.640
philosophy behind why these workers arepaid a couple bucks an hour and havetheir lives destroyed. And why AI
哲學來解釋為什麼這些工人每小時只賺幾美元，卻讓他們的生活毀滅。以及為什麼
11:46.640–11:52.240
researchers who also contribute to thesemodels are paid million-dollcompensation packages simply because
研究人員，他們為這些模型貢獻心力，卻僅獲得百萬美元的薪酬，僅僅因為
11:52.240–12:00.079
they sit in Silicon Valley in OpenAI'soffices. That is the logic of Empire.And that hearkens back to my title,
研究人員，僅僅因為他們坐在矽谷 OpenAI 的辦公室裡，就能獲得百萬美元的薪酬包。這就是帝國的邏輯。這也呼應了我的標題，
12:00.079–12:10.320
Empire of AI.>> So, let's go back to your title, Empireof AI. the subtitle dreams andnightmares in Sam Alman's Open AI. So
AI 帝國。>> 所以，讓我們回到你的標題，AI 帝國。副標題是薩姆·阿爾特曼的 OpenAI 中的夢想與噩夢。所以
12:10.320–12:17.839
tell us the story of Sam Alman and whatOpen AI is all about right through tothe deal he just made in the Gulf when
告訴我們關於薩姆·阿爾特曼的故事，以及 OpenAI 到底是怎麼回事，一直到他剛剛在中東
12:17.839–12:27.040
President Trump uh Sam Alman and ElonMusk were there. Alman is very much aproduct of Silicon Valley. His career
與特朗普總統、薩姆·阿爾特曼和伊隆·馬斯克在場時達成的協議。阿爾特曼非常具有矽谷特色。他的職業生涯
12:27.040–12:32.880
was first as a founder of a startup andthen as the president of Y Combinator,which is one of the most famous startup
首先是作為一家初創公司的創始人，然後是 Y Combinator 的總裁，這是矽谷最著名的初創公司
12:32.880–12:40.800
accelerators in Silicon Valley, and thenthe CEO of OpenAI. And there's nocoincidence that OpenAI ended up
加速器之一，之後是 OpenAI 的首席執行官。OpenAI 最終向世界介紹了
12:40.800–12:47.519
introducing the world to the scale atall cost approach to AI developmentbecause that is the way that Silicon
不惜一切代價的 AI 發展規模，這並非巧合，因為這就是矽谷
12:47.519–12:59.839
Valley has operated in the entire timethat Altman came up in it. And so he isa very strategic person. He isincredibly good at telling stories about
在阿爾特曼成長期間一直運作的模式。所以他是一個非常有策略性的人。他非常擅長講述關於
12:59.839–13:06.639
he future and painting these sweepingvisions that investors and employeeswant to be a part of. And so early on at
未來的故事，並描繪這些宏大的願景，讓投資者和員工希望成為其中的一部分。所以在早期
13:06.639–13:14.959
YC, he identified that AI would be oneof the trends that could take off. Andhe was trying to build a portfolio of
於 YC 時，他識別出 AI 將是可能爆發的趨勢之一。而他試圖建立一個
13:14.959–13:20.639
different investments and differentinitiatives to place himself in thecenter of various different trends
不同的投資組合和不同的計劃，以便根據哪一個爆發，將自己置於各種不同趨勢的
13:20.639–13:26.160
depending on which one took off. He wasinvesting in quantum computing. He wasinvesting in nuclear fusion. He was
取決於哪一項技術獲得了成功。他投資於量子計算、核融合，以及
13:26.160–13:33.360
investing in self-driving cars. And hewas developing a fundamental AI researchlab. Ultimately, the AI research lab was
自駕車。他還建立了一個基礎人工智慧研究實驗室。最終，該人工智慧研究實驗室
13:33.360–13:43.760
the ones that started acceleratingreally quickly. So he makes himself theCEO of that company. Um and andoriginally he started it as a nonprofit
開始迅速發展。因此，他成為該公司的執行長。最初，他將其設立為非營利組織
13:43.760–13:52.800
o try and position it as a counter toforprofit driven incentives in SiliconValley. But within one and a half years,
開始加速發展。所以他讓自己成為該公司的 CEO。嗯，最初他將其作為非營利
13:52.800–14:00.240
OpenAI's executives identified that ifthey wanted to be the lead in thispace, they had to go for this scale at
組織，試圖將其定位為對矽谷以利潤為驅動力的激勵措施的制衡。但在一年半內，
14:00.240–14:05.519
all cost approach and had to should bein quotes. They thought that they had todo this. There are actually many other
OpenAI 的高管們識別出，如果他們想成為這個領域的領先者，他們必須採取這種不惜一切代價的規模化方法，並且應該加引號。他們認為他們必須
14:05.519–14:11.920
ways to develop AI and to have progressin AI that does not take this approach.But once they decided that, they
這樣做。實際上還有其他許多方法來開發 AI 並在 AI 方面取得進展，而不採取這種方法。但一旦他們決定這樣做，他們
14:11.920–14:19.440
realized the bottleneck was capital. Itjust so happens Sam Alman is a once in ageneration fundraising talent. He
就意識到瓶頸是資本。巧合的是，薩姆·阿爾特曼是一代難得的募資人才。他
14:19.440–14:26.880
created this new structure nesting afor-profit arm within the nonprofit tobecome this fundraising vehicle for the
創建了這種新結構，在非營利組織內部嵌套一個營利部門，以成為這個募資工具，
14:26.880–14:33.440
tens of billions and ultimately hundredsof billions that they needed to pursuethe approach that they decided on. And
為數百億乃至最終數千億美元提供資金，以追求他們決定採取的方法。而
14:33.440–14:41.600
that is how we ultimately get to presentday OpenAI, which is one of the mostcapitalistic companies in the history of
這就是我們最終到達當前的 OpenAI，它是矽谷歷史上最資本主義的公司之一，繼續募集數百億，甚至阿爾特曼開玩笑說數萬億，以生產一種
14:41.600–14:49.600
Silicon Valley, continuing to raisehundreds of billions and and Altman hasjoked even trillions to produce a
矽谷，持續籌措數千億美元，甚至艾爾曼開玩笑說要籌措數兆美元，以生產出
14:49.600–14:58.399
technology that ultimately has amiddling economic impact thus far. We'llreturn to our conversation in a minute
與《AI 帝國：薩姆·阿爾特曼的 OpenAI 中的夢想與噩夢》一書的作者卡倫·霍（Karen Ho）繼續我們的對話。請繼續收看。
14:58.399–15:10.639
with Karen How, author of the new bookEmpire of AI: Dreams and Nightmares inSam Alman's Open AI. Stay with us.
與《AI帝國：艾爾曼的OpenAI中的夢想與夢魘》新書作者Karen How一起，請繼續收看。
15:10.639–15:14.160
This is Democracy Now! democracynow.org.
這是《Democracy Now!》，網站是 democracynow.org。
15:14.160–15:22.000
I'm Amy Goodman. In this holidayspecial, we continue with the journalistKaren How, author of the new book Empire
我是艾美·古德曼。在這次假日特別節目中，我們繼續採訪記者卡倫·豪（Karen Ho），她是新書《AI帝國：山姆·阿爾曼的Open AI中的夢想與夢魘》的作者。
15:22.000–15:32.000
of AI: Dreams and Nightmares, and SamAlman's Open AI. Karen came into ourstudio in May when she discussed how AI
卡倫於五月進入我們的攝影棚，當時她討論了AI將如何影響勞工。我們看到的一項事實是，這項技術已經對就業產生了巨大影響。
15:32.000–15:41.519
will impact workers.One of the things that we have seen isthis technology is already having a hugeimpact on jobs.
這不一定是因為技術本身真的有能力取代工作，而是人們認為它有能力做到這一點，因此高層主管正在裁員。
15:41.519–15:48.480
Not necessarily because the technologyitself is really capable of replacingjobs, but it is perceived as capable
並非因為技術本身真的有能力取代工作，而是人們認為它有能力
15:48.480–15:59.600
nough that executives are laying offworkers. And we need more some kind ofmore guard rails to actually prevent
這些公司繼續嘗試開發自動化勞工的技術，並試圖將它們轉向生產輔助勞工的技術。
15:59.600–16:10.399
hese companies from continuing to tryand develop labor automatingtechnologies and try to shift them toproducing labor assistive technologies.
這些公司繼續嘗試開發自動化勞動力的技術，並試圖將它們轉向生產輔助勞動力的技術。
16:10.399–16:11.519
What do you mean?
>> 所以，Open AI對他們所稱的「人工通用智能」的定義是高度自主的系統，在大多數
16:11.519–16:20.720
>> So, open AI, their definition of whathey call artificial generalintelligence is highly autonomousystems that outperform humans in most
經濟上有價值的工作中表現優於人類。所以，他們明確表示他們正試圖自動化地取代工作。我的意思是，什麼是
16:20.720–16:28.160
economically valuable work. So, theyexplicitly state that they are trying toautomate jobs away. I mean, what are
經濟上有價值的工。所以，他們明確表示他們正試圖自動化掉工作。我的意思是，
16:28.160–16:30.079
what is economically valuable work? But
人們為了獲得報酬所做的事情。嗯，但麻省理工學院經濟學家傑羅姆·奧斯汀（Jaron Austin）和西蒙·
16:30.079–16:39.279
he things that people do to get paid.Um but there's this really great bookcalled Power in Progress by MITeconomists Jeron Austin and Simon
約翰遜（Simon Johnson）撰寫了一本非常棒的書《進步中的權力》，他們提到技術發展、所有技術革命都採取自動化勞工的方法
16:39.279–16:46.399
Johnson who mention that technologydevelopment all technology revolutionsthey take a labor automating approach
這不是因為不可避免，而是因為頂層的人選擇自動化這些工作。他們選擇設計
16:46.399–16:54.480
not because of inevitability but becausethe people at the top choose to automatethose jobs away. They choose to design
技術，以便他們能將其銷售給高層主管，並說你可以通過裁掉所有這些員工並使用我們的AI服務
16:54.480–17:03.839
the technology so that they can sell ito executives and say you can shrinkyour costs by laying off all theseworkers and using our AI services
來縮減成本。但在過去，我們看過一些研究，例如建議如果你開發一個醫生使用的AI工具
17:03.839–17:12.400
instead. But in the past, we've seenstudies that for example suggest that ifyou develop an AI tool that a doctor
而不是取代醫生，你實際上會為病人提供更好的醫療保健。你會得到更好的癌症
17:12.400–17:19.280
uses rather than replacing the doctor,you will actually get better health carefor patients. you will get better di
診斷。如果你開發一個教師可以使用的AI工具，而不僅僅是取代教師的AI導師，你的孩子將獲得更好的
17:19.280–17:25.600
cancer diagnoses. If you develop an AItool that teachers can use rather thanjust an AI tutor that replaces the
教育成果。這就是我所說的輔助勞工而非自動化勞工的意思。
17:25.600–17:32.559
teacher, your kids will get bettereducational outcomes. And so that's whatI mean by labor assistive than labor
>> 請解釋一下你的意思，因為我認為很多人甚至不理解人工智慧。而且
17:32.559–17:38.720
>> and explain uh what you mean because Ithink a lot of people don't evenunderstand artificial intelligence. And
>> 並解釋一下你的意思，因為我認為很多人甚至不了解人工智慧。而且
17:38.720–17:42.720
when you say replace the doctor, whatare you talking about?
>> 沒錯。所以這些公司試圖開發一種技術，他們將其定位為一台什麼都能做的
17:42.720–17:49.600
>> Right. So these companies they try todevelop a technology that they positionas an everything machine that can do
萬能機器。嗯，所以他們會嘗試說你可以使用這個，你可以跟ChatGPT進行心理治療。不，你不行。ChatGPT
17:49.600–17:59.919
anything. Um and so they will try to sayou can use this you can talk to chatbtfor therapy. No, you cannot. Chat GBT is
不是持照的心理治療師。事實上，這些模型實際上會噴出大量醫療誤導資訊。而且已經有
17:59.919–18:05.919
not a licensed therapist. And in fact,these models actually spew lots ofmedical misinformation. And there have
許多例子顯示，用戶實際上因模型而受到心理傷害，因為模型會
18:05.919–18:14.559
been lots of um examples of actuallyusers being psychologically harmed bythe model because the model will
繼續強化自我傷害行為。我們甚至見過一些案例，兒童與聊天機器人對話
18:14.559–18:22.480
continue to reinforce um selfharmingbehaviors. And we've even had caseswhere uh children who speak to chatbots
並與這些聊天機器人發展出深厚的情感關係，在使用這些聊天機器人系統後實際上自殺了。
18:22.480–18:29.120
and develop huge emotional relationshipswith these chatbots have actually killedthemselves after using these chatbot
嗯，但這就是我的意思，當這些公司試圖開發自動化勞工的工具時。他們
18:29.120–18:34.559
systems. Um but that's what I mean whenthese companies are trying to developlabor automating tools. They're
將其定位為你可以現在僱用這個工具來取代僱用勞工。我的意思是，最近，山姆·阿爾曼在
18:34.559–18:42.480
positioning it as you can now hire thistool instead of hire a worker. I mean,most recently, Sam Alman was speaking at
一次會議上說：「我們最初說這些模型是律師行的初級合夥人，現在我們認為
18:42.480–18:50.080
a conference and said, "We originallysaid that these models were junior levelpartners at a law firm, and now we think
他們實際上可以成為律師行的高級同事。」他的意思是不要僱用初級合夥人，不要僱用高級
18:50.080–18:59.280
that they can really be more seniorcolleagues at a law firm." What he'saying is don't hire the junior levelpartners, don't hire the senior
同事，而只是使用我們的AI模型。我們已經看到職業階梯正在崩壞，因為許多不同的
18:59.280–19:08.160
colleagues, and just use our AI models.And we are already seeing the careerladder breaking because many different
白領工作、白領服務業以及其他行業都逐漸相信他們不需要
19:08.160–19:15.280
white collar job uh white collar serviceindustries as well as other industriesare becoming convinced that they do not
僱用實習生。他們不需要僱用入門級職位，他們只需要這些AI模型，而新
19:15.280–19:21.360
need to hire interns. They do not needto hire entry-level positions that theyjust need these AI models and new
的大學畢業生現在正努力尋找工作機會，以幫助他們在這些行業中立足。所以，
19:21.360–19:28.559
college graduates are struggling now tofind job opportunities to help them geta foothold into these industries. So,
你談到了山姆·阿爾曼，在第一部分中，我們觸及了他是誰，但我希望你能更深入地探討
19:28.559–19:36.640
you've talked about Sam Alman, and inpart one, we touched on uh who he is,but I'd like you to go more deeply into
山姆·阿爾曼是誰，他是如何突然出現在美國舞台上，在國會作證，實際上
19:36.640–19:46.559
what uh who Sam Alman is, how hexploded onto the um US scene,testifying before Congress, actuallywarning about the dangers of AI. So,
警告AI的危險。所以，這在某種程度上保護了他。>> 嗯，人們把他視為先知。那是P-O-P。但現在，我們可以談談另一種先知，P-O-F-I-T。
19:46.559–19:56.960
that really protected him in a way.>> Um people seeing him as a prophet.That's a P O P. But now, we can talkabout the other kind of prophet, P O Fit
並探討Open AI是如何成立的。Open AI與AI有何不同？
19:56.960–20:05.360
T.um and how open AI was formed. How isopen AI different from AI?
Open AI是一家公司，我的意思是它最初是作為非營利組織成立的，正如我所提到的，阿爾曼特別是在他
20:05.360–20:15.440
OpenAI is a com I mean it was originallyfounded as a nonprofit as I mentionedand Alman specifically when he was
思考如何建立一個將引起轟動的根本AI研究實驗室時，他選擇將其設為非營利組織，因為他
20:15.440–20:25.120
thinking about how do I make afundamental AI research lab that isgoing to make a big splash he chose tomake it a nonprofit because he
思考如何建立一個能引起巨大轟動的根本性 AI 研究實驗室時，他選擇將其設為非營利組織，因為他
20:25.120–20:33.200
identified that if he could not competeon capital and he was relatively late tothe game Google already had a monopoly
認清了自己無法在資金上競爭，而且他入局較晚，當時 Google 已經壟斷了
20:33.200–20:41.520
on a lot of top AI research talent athe time if he could not compete oncapital and he could not compete um interms of being a first mover he
許多頂尖 AI 研究人才。如果他在資金上無法競爭，也無法在成為先驅者方面競爭，他
20:41.520–20:49.600
needed some other kind of ingredienthere to really recruit talent recruitum public goodwill and establish a name
就需要其他類型的要素來真正招募人才、爭取公眾善意並建立名聲
20:49.600–20:58.159
for open AAI so he identified a missionhe identified let me make this anonprofit and let me give it a reallycompelling mission so the mission of
以打造 OpenAI。因此他確立了一項使命，他確立了讓我將其設為非營利組織，並賦予它一項極具吸引力的使命，所以 OpenAI 的使命是
20:58.159–21:06.880
openai is to ensure artificial generalintelligence benefits all of humanityAnd one of the quotes that I open my
確保通用人工智慧造福全人類。我在書中引用的其中一句話是 Sam Altman
21:06.880–21:16.000
book with is this quote that Sam Almancited himself in 2013 um in his blog. Hewas an avid blogger back in the day
在 2013 年他自己部落格中引用的這句話。他當時是一位熱衷於部落格寫作的
21:16.000–21:23.600
talking about his learnings on businessand strategy and Silicon Valley startuplife. And the quote is successful people
談論他在商業、策略和矽谷創業生活方面的學習心得。這句話是：成功的人
21:23.600–21:31.200
built companies, more successful peoplebuild countries. The most successfulpeople build religions. And then he
建立公司，更成功的人建立國家。最成功的人建立宗教。然後他
21:31.200–21:37.520
reflects on that quote in his blogsaying, "It appears to me that the bestway to build a religion is actually to
在他的部落格中反思這句話，說：「對我來說，建立宗教的最佳方式其實是
21:37.520–21:46.240
build a company.">> And so talk about how Alman was thenforced out of the company and then cameback. And also I just found it so
建立一家公司。」>> 接著談論 Altman 如何被公司驅逐，然後又回歸。還有，我覺得非常
21:46.240–21:52.480
fascinating that you were able to speakwith so many Open AI workers. Youthought there was a kind of total ban on
迷人的是，你能與這麼多 OpenAI 的員工交談。你認為有一道全面的禁令
21:52.480–21:59.360
you.>> Yes. Yeah. Exactly. So I was the firstjournalist to profile OpenAI. Um, Iembedded within the company for 3 days
阻止你。>> 是的。沒錯。正是如此。所以我是第一位為 OpenAI 撰寫專訪的記者。我在
21:59.360–22:06.880
in 2019 and then my profile published in2020 for MIT technology review and athe time I identified in the profile
2019 年於該公司內部進行了 3 天的嵌入式採訪，隨後我的專訪於
22:06.880–22:13.760
this tension that I was seeing where itwas a nonprofit by name but behind thescenes a lot of the public values that
2020 年發表於 MIT Technology Review。當時我在專訪中指出
22:13.760–22:18.960
hey exposed were actually the oppositeof how they operated. So they espousedtransparency but they were highly
我所觀察到的這種張力：該組織名義上是非營利組織，但背地裡所展現的許多公共價值
22:18.960–22:26.720
secretive. They espousedcollaboriveness. They were highlycompetitive. and they espoused that theyhad no commercial intent, but in fact it
秘密。他們宣揚合作精神。他們高度競爭。他們宣稱沒有商業意圖，但實際上它
22:26.720–22:32.159
seemed like they had just gotten a $1billion investment from Microsoft. Itseems like they were rapidly going to
保密。他們宣稱協作，卻高度競爭。他們宣稱沒有商業意圖，但事實上看來
22:32.159–22:38.320
develop commercial intent. And so Iwrote that into the profile and OpenAIwas deeply unhappy about it and they
他們剛剛獲得了微軟的 10 億美元投資。看來他們正迅速發展
22:38.320–22:44.640
would not refuse to talk to me for 3years. But when I started working on thebook, when I started reaching out to
商業意圖。因此我將這些寫入專訪中，OpenAI 對此深感不滿，他們
22:44.640–22:50.880
employees, current and former, Idiscovered that many employees actuallyreally liked the profile and they
拒絕與我交談長達 3 年。但當我開始撰寫這本書，開始聯絡
22:50.880–22:59.200
specifically wanted to talk to mebecause they thought that I would dojustice to the truth of what hadactually happened within the company and
特別想和我談話，因為他們認為我會公正地呈現公司內部實際發生的真相，並
22:59.200–23:07.520
be able to discover behind what thexecutives mythologized and narrativizedabout this technology [clears throat]
特別希望與我交談，因為他們認為我會公正地呈現公司內部實際發生的真相，
23:07.520–23:13.679
and about the course of this company. Iwould be able to actually get beneathat to the real heart of the matter.
並了解這家公司的發展歷程。我能夠真正深入探究，觸及問題的核心。
23:13.679–23:21.760
And so um one of the things that youreally have to understand about AIdevelopment today is that there are what
和這家公司發展歷程神話化、敘事化的背後，發現真相。我能夠真正深入核心，
23:21.760–23:29.200
I call quasi religious movements thathave developed within Silicon Valley.The concept of artificial general
我稱之為在矽谷內部發展出來的準宗教運動。人工通用
23:29.200–23:36.480
intelligence is not one that'scientifically grounded. It is this ideathat we can fundamentally recreate human
智能的概念並非基於科學基礎。這是一個我們可以根本性地重現人類
23:36.480–23:42.400
intelligence in computers. And this ideahas been around for actually a reallylong time. The field of AI was founded
智能於電腦中的想法。這個想法其實存在了很長一段時間。人工智慧領域成立於
23:42.400–23:48.320
all the way back in the 1950s and thatwas the original intent of the field.How do we recreate intelligence in
早在 1950 年代，這也是該領域最初的意圖。我們如何在
23:48.320–23:50.080
computers? Can machines think? That was
智能於電腦中的想法。這個想法其實存在了很長一段時間。AI 領域早在
23:50.080–23:58.159
the famous question that Britishmathematician Alan Turing asked. But weto this day do not have scientific
1950 年代就已創立，那也是該領域最初的意圖。我們如何在
23:58.159–24:05.840
onsensus around even what humanintelligence is. And so to peg an entiresearch field and a technology to the
甚至對於人類智能是什麼都無法達成共識。因此，將整個搜尋領域和技術與
24:05.840–24:12.640
basis of human intelligence is a verytricky endeavor because there are nogood metrics to assess have we actually
英國數學家艾倫·圖靈提出的著名問題。但直到今天，我們甚至對人類
24:12.640–24:22.159
gotten there yet and there's noblueprint to say what should AI looklike and how should it work andultimately who should it serve. And so
智能是什麼都還沒有科學共識。因此，將整個研究領域和技術建立在
24:22.159–24:29.440
when OpenAI took up this mission ofartificial general intelligence, theywere able to essentially shape and mold
人類智能的基礎上是一項非常棘手的任務，因為沒有好的指標來評估我們是否真的
24:29.440–24:36.640
what they wanted this technology to bebased on what is most convenient forthem. But when they identified it, it
達成了目標，也沒有藍圖來說明 AI 應該是什麼樣子、如何運作，以及最終應該為誰服務。因此
24:36.640–24:45.279
was at a time when scientists reallylooked down on this term even AGI. Andso they absorbed just a small group of
當 OpenAI 承擔起通用人工智慧的使命時，他們能夠根據對他們最方便的方式， essentially 塑造和塑造
24:45.279–24:52.159
self-identified AGI believers. This iswhy I call it quasi religious becausethere's no scientific evidence that we
他們希望這項技術成為的樣子。但當他們確立這一概念時，
24:52.159–24:59.279
can actually develop AGI. The people whoare strongly con have this strongconviction that they will do it and that
科學家們真的輕視這個詞，甚至 AGI。因此他們只吸收了一小群
24:59.279–25:05.279
it's going to happen soon. It is justpurely based on belief and they talkabout it as a belief too. But there are
自我認同的 AGI 信仰者。這就是我稱之為準宗教的原因，因為沒有科學證據表明我們
25:05.279–25:11.919
two factions within this belief systemof the AGI religion. There are peoplewho think AGI is going to bring us to
這個信仰體系內的兩個派系，即 AGI 宗教的兩派。有些人認為 AGI 將為我們帶來
25:11.919–25:20.559
utopia and there are people who thinkAGI is going to destroy all of humanity.Both of them believe that it ispossible. It's coming soon. And
烏托邦，而有些人則認為 AGI 將毀滅全人類。兩派都相信這有可能發生。它即將到來。而且
25:20.559–25:27.520
therefore they conclude that they needto be the ones to control the technologyand not democratize it. And this is
因此他們認為必須由他們來掌控這項技術，而非將其民主化。而這正是
25:27.520–25:33.760
ultimately what leads to your questionof what happened when Sam Alman wasfired and rehired through the history of
這最終引出了你的問題：在 OpenAI 的歷史中，當山姆·阿爾特曼被解僱又復聘時，究竟發生了什麼事。
25:33.760–25:39.840
OpenAI. There's been a lot of clashingbetween the boomers and doomers aboutwho should actually>> the boomers and doomers.
在 OpenAI 內部，嬰兒潮世代與末日論者之間就誰才應該主導大局產生了許多衝突。
25:39.840–25:47.120
>> The boomers and the doomers. [laughter]>> Those that say it'll bring us theapocalypse. topia boomers and those that
>> 嬰兒潮世代和末日論者。[笑聲] >> 那些認為它會帶來世界末日的人。烏托邦的嬰兒潮世代和那些
25:47.120–25:54.960
say it'll destroy humanity. The doomersand they have clashed relentlessly andaggressively about how quickly to build
說它會毀滅人類。末日論者與他們 relentless 且具攻擊性地爭論著技術建構的速度
25:54.960–26:04.799
the technology, how quickly to releasethe technology and ultimately Altman isone that he is really good at saying to
這項技術、多快釋出這項技術，以及最終奧爾特曼非常擅長對人們說他們想聽的話
26:04.799–26:10.960
people what they need to hear and hewill say different things to differentpeople if he thinks they need to hear
對人們說他們想聽的話，並且如果他认为某些人需要聽不同的內容，他會對不同的人說不同的話
26:10.960–26:14.320
different things. So when I askedboomers, is Altman a boomer? They said
不同的說法。所以當我問嬰兒潮世代，奧特曼是嬰兒潮世代嗎？他們說
26:14.320–26:19.279
yes. When I asked doomers, is Altman adoomer? They said yes. And I want to
是的。當我問千禧世代，奧特曼是千禧世代嗎？他們說是的。而我想
26:19.279–26:28.799
take this up until today to um inJanuary, the Trump administrationannouncing the Stargate project, a $500
這件事一直持續到今天，直到今年一月，特朗普政府宣布星門計畫，一個價值五千億美元的
26:28.799–26:39.520
billion project to boost AIinfrastructure in the United States.This is Open AI Sam Alman speakingalongside President Trump.
千億美元計畫，旨在提升美國的人工智慧基礎建設。這是 OpenAI 的山姆·奧爾特曼與特朗普總統共同發表的講話。
26:39.520–26:45.919
I think this will be the most importantproject of this era and as Masa said forAGI to get built here to create hundreds
我認為這將是本時代最重要的計畫，正如馬薩所說，為了在這裡建構通用人工智慧，創造數十萬個工作機會
26:45.919–26:51.830
of thousands of jobs to create a newindustry centered here. Uh we wouldn'tbe able to do this without you Mr.President.
創造數以萬計的就業機會，建立以這裡為中心的新產業。總統先生，沒有您我們無法做到這一點。
26:51.840–27:01.679
>> He also there referred to AGI um uhartificial general intelligence. Explainwhat happened here and what this is and
>> 他也在那裡提到了人工通用智能，也就是AGI。請解釋這裡發生了什麼，這究竟是什麼，以及它是否真的實現了。
27:01.679–27:14.080
has it actually happened. So Altmanbefore Trump was electedum he already was sensing throughobservation that it was possible that
事實上有沒有發生呢？所以在特朗普當選之前，奧爾特曼就已經透過觀察察覺到這種可能性
27:14.080–27:21.279
he administration would shift and thathe would need to start politicking quiteheavily to ingruiate himself to a new
政府會轉向，而他必須開始大力進行政治遊說，以贏得新政府的青睞
27:21.279–27:30.480
administration.Alman is very strategic. Um he was undera lot of pressure at the time as wellbecause his original co-founder Elon
政府。奧爾特曼非常具有策略性。當時他也承受著巨大壓力，因為他的共同創始人伊隆·馬斯克
27:30.480–27:38.240
Musk now has great beef with him. UhMusk feels like Alman used his name andhis money to set up OpenAI and then he
現在馬斯克跟他有嚴重的糾紛。馬斯克覺得奧爾特曼利用他的名聲和資金來創立 OpenAI，然後他
27:38.240–27:46.080
got nothing in return. So Musk had beensuing him, still suing him and suddenlybecame first buddy of the Trump
一無所獲，所以 Musk 一直在對提告，現在卻突然成為特朗普新政府的親密夥伴
27:46.080–27:56.559
administration. So Altman basicallycleverly orchestrateda um this announcement where by the waythe the announcement is quite strange
行政團隊。所以奧特曼基本上巧妙地策劃了這項公告，順便一提，這項公告相當奇怪
27:56.559–28:03.600
because the Trump President Trump is notit's not the US government giving $500billion. It's private investment coming
因為特朗普總統並非美國政府提供五千億美元，而是來自軟銀等地區的私人投資
28:03.600–28:13.840
into the US um from places like SoftBank>> which is>> uh which is one of the largestinvestment funds um run by Masay Yoshi
進入美國，嗯，來自軟銀等地方，也就是，呃，也就是由孫正義經營的其中一個最大投資基金
28:13.840–28:22.240
Son a Japanese businessman who made alot of his wealth from the previous techera. So, so it's not even the USgovernment that's that's providing this
孫正義是一位日本商人，他的財富主要來自上一個科技時代，所以提供這些資金的甚至不是美國政府
28:22.240–28:33.520
money.>> And take that right through to now thatGulf trip that um Elon Musk was on, butso was Sam Alman to the fury of Elon
錢>>然後一路延續到現在的沙特阿拉伯之行，當時伊隆·馬斯克去了，山姆·阿爾特曼也去了，這讓伊隆·馬斯克非常憤怒
28:33.520–28:45.200
Musk and then a deal was sealed in AbuDhabi.>> Yeah.>> It didn't include Elon Musk but wasabout open AI.>> Exactly. So Altman has continued to try
然後在阿布達比敲定了交易。對，雖然沒有伊隆·馬斯克參與，但談的是開放人工智能。沒錯，所以阿爾特曼繼續嘗試
28:45.200–28:59.919
and use the US government as a way to toget access to more places and uh morepowerful spaces to build out thisempire. And one of the one of the things
並利用美國政府作為管道，以取得更多地點與更具權力的空間，來擴張這個帝國。而其中一項
28:59.919–29:06.720
because OpenAI's computationalinfrastructure needs are so aggressive.You know, I had an OpenAI employee tell
因為 OpenAI 的運算基礎設施需求非常龐大。你知道，我有一位 OpenAI 的員工告訴我
29:06.720–29:13.360
me we're running out of land and power.So they are running out of resources inthe US which is why they're trying to
我有一位 OpenAI 的員工告訴我，我們即將耗盡土地和電力資源，因此他們在美國面臨資源匱乏，這也是他們試圖
29:13.360–29:19.679
get access to land and energy in otherplaces. The Middle East has a lot ofland and has a lot of energy and they're
在其他地方取得土地和能源的存取權。中東擁有大量的土地和能源，並且他們願意進行交易，這就是為什麼奧特曼會參與那次行程，以尋求達成協議，而他們所達成的協議是，利用當地的土地和能源，在中東建造一座或數座大型資料中心。但最近開放人工智能推出的一項新計畫，他們稱之為「面向國家的開放人工智能」計畫
29:19.679–29:26.399
willing to strike deals and that is whyAltman was part of that trip looking tostrike a deal and what they the deal
願意簽訂協議，這也是為什麼奧特曼參與了那次行程，旨在敲定協議，而他們所達成的協議
29:26.399–29:35.360
that they struck was to build a massivedata center or multiple data centers inthe Middle East using their land and
他們所簽署的協議，是要利用當地的土地和能源，在中東建設大型或多個數據中心
29:35.360–29:42.559
their energy. But one of the things thatOpenAI has recently rolled out, theycall it the OpenAI for countries program
但 OpenAI 最近推出的一項計畫，他們稱之為 OpenAI for countries 計畫
29:42.559–29:53.440
and it is this idea that they want toinstall OpenAI hardware and software inplaces around the world and explicitly
而他們正是基於這個理念，希望在全球各地安裝 OpenAI 的硬體與軟體，並且明確地
29:53.440–30:05.039
says we want to build democratic AIrails.We want to install our hardware andsoftware as a foundation of democratic
表示我們希望建立民主的 AI 規範。我們希望在全球安裝我們的硬體和軟體，作為民主的基礎
30:05.039–30:15.919
AI globally so that we can stop Chinafrom installing authoritarian AIglobally. But the thing that he does not
全球部署AI，以便阻止中國在全球安裝專制AI。但他沒有提到的是
30:15.919–30:23.360
acknowledge is that there is nothingdemocratic about what he's doing. Youknow, the Atlantic executive editor says
他沒有承認的是，他所做的毫無民主可言。你知道，大西洋雜誌執行編輯說
30:23.360–30:31.360
we need to call these companies for whathey are. They are technoauthoritarians. They do not ask thepublic for any perspective on how they
我們必須揭穿這些公司的真面目，它們是技術專制主義者，它們不會就技術開發、訓練數據以及數據中心的建設地點向公眾徵求任何意見
30:31.360–30:36.559
develop the technology, what data theytrain the technology on, where theydevelop these data centers. In fact,
開發技術、使用哪些數據來訓練技術，以及這些數據中心設在哪裡。事實上
30:36.559–30:44.159
these data centers are often developedin the cover of night um under shellcompanies like Meta recently entered New
這些數據中心通常是在夜幕掩護下開發的，例如 Meta 最近透過殼公司進入新墨西哥州。
30:44.159–30:54.720
Mexico under the Shell company namedGreater Kudu LLC.>> Greater Kudu.LLC. And once the deal wasactually closed and the residents
殼公司名為 Greater Kudu LLC。>> Greater Kudu LLC。一旦交易實際完成，居民
30:54.720–30:58.960
couldn't do anything about anymore,that's when it was revealed, surprise,we're Meta and you're going to get a
再也無能為力時，才揭露真相，驚喜的是，我們是 Meta，而你將獲得一座
30:58.960–31:04.320
data center that drinks all of yourfresh water.>> And then there was this wholecontroversy in Memphis around a data
耗盡你所有新鮮水源的數據中心。>> 然後在孟菲斯發生了一場關於數據中心
31:04.320–31:12.159
center.>> Yes. So that is the data center thatElon Musk is building. So meanwhile,Musk is saying Alman is terrible.
的爭議。>> 是的。這就是伊隆·馬斯克正在建造的數據中心。與此同時，
31:12.159–31:19.120
Everyone should use my AI. And ofcourse, his AI is also being developedusing the same environmental and public
每個人都應該使用我的 AI。當然，他的 AI 也是同樣的環境和公共
31:19.120–31:25.039
health costs. So he built this massivesupercomputer called Colossus inMemphis, Tennessee that's training
健康成本下開發的。所以他在田納西州孟菲斯建造了一個名為 Colossus 的大型
31:25.039–31:36.880
Grock, the chatbot that people canaccess through X and that is beingpowered byaround 35 unlicensed methane gas
超級電腦，用於訓練 Grock，人們可以透過 X 存取這個聊天機器人，而它是由
31:36.880–31:45.440
turbines that are pumping thousands oftons of toxic air pollutants into thegreater Memphis community. And that
大約 35 部未獲許可的天然氣渦輪機驅動，向大孟菲斯社區排放數千噸有毒空氣污染物。而該
31:45.440–31:54.640
community has long suffered a lack ofaccess to clean air, a fundamental humanright.>> So I want to go to interestingly Sam
社區長期以來缺乏獲得清潔空氣的權利，這是一項基本人權。>> 所以我想轉向一個有趣的議題，Sam
31:54.640–32:03.279
Alman testifying in front of Congressabout solutions to the high energyconsumption of artificial intelligence.
Alman 在國會作證，討論人工智慧高能源消耗的解決方案。
32:03.279–32:09.360
>> In the short term, I think this probablylooks like more natural gas. Um althoughthere are some applications where I
>> 短期內，我認為這可能看起來像是更多的天然氣。嗯，雖然在某些應用場景中，我認為
32:09.360–32:16.000
think solar can really help. In themedium-term, I hope it's advancednuclear uh fish and fusion. More energy
太陽能確實能提供幫助。在中長期，我希望是先進核能、氫能和核融合。更多能源
32:16.000–32:24.320
is important well beyond AI.>> So that's open AI's Sam Alman. This istestifying before the Senate and talking
的重要性遠超 AI 範疇。>> 所以這是 Open AI 的 Sam Alman。他在參議院作證，談論
32:24.320–32:31.600
about everything from uh solar tonuclear power. Something that was foughtin the United States by environmental
從太陽能到核能等各種議題，這是在美國被環保人士數十年來爭論的話題
32:31.600–32:38.480
activists for decades. So you have thesehuge old uh nuclear power plants, butmany say you can't make them safe no
活動人士數十年來反對的議題。所以你有這些巨大的舊核電廠，但
32:38.480–32:45.279
matter how small. and smart you makethem.>> This is one of the things of the manythings that I'm concerned about with the
許多人說，無論你把它們做得多小、多聰明，都無法保證它們的安全。>> 這是讓我擔憂的許多事情之一，關於
32:45.279–32:52.000
current trajectory of AI development.This is a second order tertiary ordereffect is that because these companies
AI發展的當前軌跡，這是一種二階或三階效應，因為這些公司
32:52.000–32:59.679
are trying to claim that the AIdevelopment approach they took doesn'thave climate harms. They are explicitly
試圖宣稱他們採取的 AI 開發方法沒有氣候危害。他們明確地
32:59.679–33:05.360
evoking nuclear again and again andagain as nuclear will solve the problem.And it has been effective. I have talked
一次又一次地提及核能，因為核能將解決問題。而且這很有效。我與某些 AI 研究人員談過，他們認為
33:05.360–33:13.519
with certain AI researchers who thoughthe problem was solved because ofnuclear and in order to try and actually
與某些認為核能已解決該問題的人工智慧研究人員進行交流，並且為了實際推動更多核電廠的建設，他們正在遊說政府試圖鬆綁核電廠建設的監管架構。
33:13.519–33:24.000
build more and more nuclear plants, theyare lobbying governments to try andunwind the regulatory structure around
建造越來越多的核電廠，他們正在遊說政府試圖鬆綁核電廠建設的監管架構
33:24.000–33:32.159
nuclear power plant building. I meanthis is this is like crazy on so manylevels that they're not just trying to
核電廠建設的監管結構。我的意思是，這在許多層面上都瘋狂，他們不僅試圖
33:32.159–33:42.320
develop these the AI technologyrecklessly. They are also trying to laydown infrastructure and nuclearinfrastructure in this move fast break
魯莽地開發這些 AI 技術。他們還試圖在「快速行動，打破常規」的意識形態下，
33:42.320–33:50.480
things ideology. But for those who umare environmentalists and have longopposed nuclear will they be sucked in
鋪設基礎設施和核基礎設施。但對於那些長期反對核能的環境主義者來說，他們會被
33:50.480–33:54.320
by the solar alternative? But that exact
太陽能替代方案所吸引嗎？但事實正是如此
33:54.320–34:02.799
so data centers have to run 247. So theycannot actually run on just renewables.That is why the companies keep trying to
數據中心必須 24/7 運行。所以它們實際上不能僅靠再生能源運行。這就是為什麼公司不斷試圖
34:02.799–34:11.599
evoke nuclear as the solve all but solardoes not actually work when we do nothave sufficient enough energy storage
將核能塑造成萬靈丹，但當我們缺乏足夠的能源儲存解決方案時，太陽能實際上並不可行
34:11.599–34:18.879
solutions for that 24/7 operation. We'llreturn to our conversation in a minutewith Karen How, author of the new book
解決方案來支持 24/7 運行時，太陽能實際上並不起作用。我們將在一分鐘後回到與 Karen Ho 的對話，她是新書
34:18.879–34:32.800
Empire of AI: Dreams and Nightmares inSam Alman's Open AI. Stay with us.This is Democracy Now! Democracynow.org.
《AI 帝國：Sam Alman 的 Open AI 中的夢想與噩夢》的作者。請繼續收看。這是《Democracy Now!》。Democracynow.org。
34:32.800–34:42.480
I'm Amy Goodman. In this holidayspecial, we're speaking with thejournalist Karen How, author of the newbook Empire of AI: Dreams and Nightmares
我是 Amy Goodman。在這個假日特別節目中，我們正在採訪記者 Karen Ho，新書《AI 帝國：Sam Alman 的 Open AI 中的夢想與噩夢》的作者。
34:42.480–34:53.200
in Sam Alman's Open AI. She came intour studio in May. She lives in HongKong. I asked her to talk about what'shappening in China around artificial
在薩姆·阿爾曼的開放人工智能中。她於五月進入我們的攝影棚。她住在香港。我請她談談中國在人工智慧領域正在發生的事情
34:53.200–35:02.079
intelligence.>> China and the US are the largest hubsfor AI research. They are the largestconcentration of AI research talent
人工智慧的發展情況。>> 中國和美國是最大的人工智慧研究樞紐。它們是全球最大
35:02.079–35:09.440
globally. Um, China other than SiliconValley, China really is the only otherival in terms of talent density and the
的人工智慧研究人才聚集地。嗯，除了矽谷之外，中國在人才密度方面確實是唯一的競爭對手，以及
35:09.440–35:14.800
amount of capital investment and theamount of infrastructure that is goinginto AI development. In the last few
資本投資的數量和進入 AI 開發的基礎設施數量。在最近幾
35:14.800–35:25.440
years, what we have seen is the USgovernment has been aggressively tryingto stay number one and one of themechanisms that they have used is export
年裡，我們看到美國政府積極試圖保持第一，其中一種機制是出口管制。這些 AI
35:25.440–35:34.480
controls. A key input into these AImodels is the computationalinfrastructure and the computer chipsfor installing into the data centers for
管制。這些 AI 模型的一個關鍵輸入是運算基礎設施，以及安裝在資料中心用於訓練這些模型的電腦晶片。為了開發 AI 模型，企業正在使用最先進的電腦晶片技術。這就像每兩年就會推出新款晶片，他們便立即開始使用該晶片來訓練下一代 AI 模型。這些電腦晶片由美國公司設計，其中最著名的是
35:34.480–35:43.040
training these models. And thesecomputer chips are the in order todevelop the AI models. Companies areusing the most bleeding edge computer
訓練這些模型的電腦晶片。而這些電腦晶片是開發 AI 模型所必需的。公司正在使用最先進的電腦
35:43.040–35:48.640
chip technology. It's like the every twoyears a new chip comes out and theyimmediately start using that to train
晶片技術。就像每兩年出一款新晶片，他們立即開始使用它來訓練
35:48.640–35:55.760
the next generation of AI models. Thosecomputer chips are designed by Americancompanies, the most prominent one being
下一代 AI 模型。這些電腦晶片由美國公司設計，其中最著名的是
35:55.760–36:03.920
Nvidia in California. And so the USgovernment has been trying to use exportcontrols to prevent Chinese companies
加州的輝達（Nvidia）。因此，美國政府一直試圖利用出口管制，阻止中國公司
36:03.920–36:12.320
from getting access to the most cuttingedge computer chips. That has all beenunder the recommendation of Silicon
取得最尖端的電腦晶片。這一切都是基於矽谷的建議，說這是防止
36:12.320–36:23.119
Valley saying this is the way to preventChina from being number one. and likeput export controls on them and don't
中國成為第一名的方法。並對他們實施出口管制，完全不監管我們，讓我們保持第一名，而他們則落後。
36:23.119–36:32.800
regulate us at all so we can stay numberone and they will fall behind. What hashappened insteadis because there is a strong base of
實際發生的情況是，由於中國擁有強大的AI研究人才基礎，在計算資源較少的限制下，中國公司實際上已經能夠
36:32.800–36:41.920
talent of AI research talent in Chinaunder the constraints of fewercomputational resources, Chinesecompanies have actually been able to
創新並開發出與美國公司相同水平的AI模型能力，但計算資源少了兩個數量級
36:41.920–36:50.079
innovate and develop the same level ofAI model capabilities as Americancompanies with two orders of magnitude
，更少的能源、更少的數據。所以，我特別指的是中國公司Highfly，它今年早些時候開發了這個
36:50.079–37:00.240
less computational resources, lessenergy, less data. So, I'm talkingspecifically about um the Chinesecompany Highfire, which developed this
名為Deep Seek的模型，該模型曾短暫導致全球經濟震盪，因為該公司表示他們
37:00.240–37:09.040
model called Deep Seek earlier this yearthat briefly tanked the global economybecause the company said that their
訓練這一個AI模型的費用大約為600萬美元，而OpenAI訓練模型的費用數以億計，甚至超過數百億
37:09.040–37:21.680
um training this one AI model costaround $6 million when OpenAI wastraining models that cost hundreds ofmillions if not over tens of billions of
美元。這種差異向人們證明，矽谷過去幾年試圖說服大家的觀點——即獲得更多AI能力的唯一途徑——是完全錯誤的，實際上，中國公司使用的技術
37:21.680–37:30.640
dollars. And that delta demonstrated topeople that this what Silicon Valley hastried to convince everyone for the last
是文獻中已經存在的技術，只需進行組合。他們使用了大量的工程技巧來完成這項工作，但他們並沒有使用
37:30.640–37:39.040
few years that this is the only path togetting more AI capabilities is totallyfalse and actually the techniques that
根本性的新技術。他們使用的是實際上已經存在的技術。>> 所以請進一步解釋，因為我認為很多人無法理解這一點。
37:39.040–37:46.560
chi the Chinese company was using wereones that existed in the literature andjust had to be assembled. They used a
奇虎（Chi）使用的技術是文獻中已存在的，只需進行組合。他們投入了
37:46.560–37:52.800
lot of engineering sophistication to dothat, but they weren't actually usingfundamentally new techniques. They were
>> 有一種稱為神經網絡的軟件，本質上是一個巨大的統計引擎，它進行大量複雜的統計計算，試圖確定數據集中存在哪些類型的模式。所以
37:52.800–37:58.720
ones that actually already existed.>> So explain it further because I think alot of people just can't get their minds
通常，在我們轉向大型語言模型之前，它會做一些事情，比如查看MRI
37:58.720–38:02.550
around this. How do you do thistraining?
掃描，並檢查癌症在MRI掃描中的模式。現在，對於GPT，它查看的是
38:02.560–38:11.280
>> So there's software called neuralnetworks which is essentially a massivestatistical engine. is doing lots and
>> 因此有一種稱為神經網路的軟體，它本質上是一個龐大的統計引擎，正在進行大量
38:11.280–38:19.599
lots of sophisticated statisticalcomputation to try and ascertain whatkinds of patterns exist in data sets. So
結構、通常使用的修辭手法，然後它使用這些模式來構建新句子。這就是生成式AI
38:19.599–38:26.000
typically in in the past before we goto large language models it would bedoing something like um looking at MRI
的工作原理。之所以計算成本如此高昂，是因為它在處理這些模式的數字。你輸入的數據越多，它需要處理的數字就越多。因此，我們過去訓練這些AI模型時，使用的是
38:26.000–38:34.560
scans and checking the patterns of whatdoes cancer look like in an MRIscan. Um now with GBT what it's looking
你知道的強大筆記型電腦，也許只有一個電腦晶片，也許是像麻省理工學院這樣的富裕實驗室，他們會在幾個或幾十個電腦晶片上進行訓練，而像Google這樣的公司
38:34.560–38:38.320
at is what are the patterns of theEnglish language? what is the syntax,
可能會在幾百個電腦晶片上進行訓練。我們現在談論的是數十萬個電腦晶片訓練單一模型。
38:38.320–38:48.640
the structure, figures of speech thatare typically used and then it usesthose patterns to construct newsentences. That's how generative AI
這就是OpenAI所說的構建這些技術所必需的，而Deep Seek證明了這是錯誤的。所以，讓我問你一個問題，Karen。你最近在美國旅行期間，
38:48.640–38:55.280
works. And the reason why it's socomputationally expensive is becauseit's crunching the numbers for those
在返回香港之前，關於特朗普對學術界的攻擊，這如何契合。特朗普對
38:55.280–39:03.680
patterns. And the more data you feed in,the more it has to crunch. And so it weused to train these AI models on you
模式。你輸入的資料越多，它需要運算的數量就越大。因此，我們過去在訓練這些 AI 模型時，使用的是
39:03.680–39:12.640
know a powerful laptop like maybe onecomputer chip maybe the richest labsacademic labs like MIT they would be uh
，並撤銷他們的簽證，將如何影響AI行業的未來？但不僅是中國學生，因為現在發生的事情讓世界各地的學生感到恐懼，而且實驗室以各種方式關閉，美國
39:12.640–39:20.240
training on a couple or a dozen computerchips and companies like Google theywould be training maybe on a couple
學生也決定出國。這只是美國政府過去幾年採取的最新行動，旨在真正疏遠美國創新的一個關鍵人才庫。最初，在美國工作的中國研究人員
39:20.240–39:27.680
hundred computer chips. We are nowtalking about hundreds of thousands ofcomputer chips training a single model.
對美國AI的貢獻比在中國的多，因為[噴鼻聲]就在幾年前，中國研究人員渴望在
39:27.680–39:37.119
Um and that is the you know that is whatopen AI says is necessary to build thesetechnologies and that is what deepseek
美國公司工作。他們想搬到美國。他們想為美國經濟做出貢獻。他們不想回到自己的
39:37.119–39:45.119
proved wrong. So, let me ask yousomething, Karen. uh the latest news umas you're traveling in the United States
祖國。但由於所謂的「中國行動計劃」，這是第一屆特朗普政府試圖將中國學者或華裔學者刑事化的行動，其中一些人是美國人，僅因
39:45.119–39:55.760
before you go back to Hong Kong uh ofTrump's attack on academia, how thisfits in. Um how could Trump's attack on
文件錯誤。他們指控他們是間諜。這是其中一項最初的行動。然後當然
39:55.760–40:04.240
international students specificallytargeting the what more than 250,000 aquarter of a million Chinese students
疫情爆發，美中貿易升級開始加劇反華言論。所有這些都導致
40:04.240–40:11.839
and revoking their visas impact thefuture of the AI industry. But not justChinese students because what's going on
以及撤銷他們的簽證，將如何影響人工智慧產業的未來，但受影響的不僅是中國學生，因為目前發生的事
40:11.839–40:21.280
here now is terrifying students aroundthe world and because labs are shuttingdown in all kinds of ways here uh US
現在的情況讓全球學生感到恐懼，因為實驗室正在以各種方式關閉，美國
40:21.280–40:29.920
students as well uh deciding to goabroad.This is just the latest action that theUS government has taken over the last
以及決定出國留學的學生。這是美國政府過去幾年來為疏遠美國創新關鍵人才庫所採取的最新行動
40:29.920–40:42.480
few years to really alienate a keytalent pool for US innovation.Originally, there were more Chineseresearchers working in the US
近年來美國政府採取的這項最新舉措，旨在疏遠對美國創新至關重要的關鍵人才庫。最初，在美國工作的中國研究人員比在中國的多
40:42.480–40:52.160
contributing to US AI than there were inChina because [snorts] just a few yearsago, Chinese researchers aspired to work
對美國 AI 的貢獻比在中國還多，因為 [噴鼻息聲] 就在幾年前，中國研究人員還渴望在美國工作
40:52.160–41:02.960
for American companies. They wanted tomove to the US. They wanted tocontribute to the US economy. Theydidn't want to go back to their home
為美國企業工作。他們希望移居美國，希望為美國經濟做出貢獻。他們不想回到祖國
41:02.960–41:11.760
country. But because of what was calledthe China Initiative, which was the afirst Trump era initiative to try and
國家。但由於所謂的中國行動計劃，這是特朗普第一任期內試圖
41:11.760–41:18.960
criminalize Chinese academics orethnically Chinese academics, some ofwhom were actually Americans um based on
將中國學者或具有中國族裔背景的學者刑事化，其中一些學者實際上已是美國人，基於
41:18.960–41:26.480
just paperwork errors. They would accusethem of being spies. That was one of thefirst actions. Then of course the
只是文件上的疏失。他們會指控這些人是間諜。這是其中一項最早的行動。然後當然
41:26.480–41:34.720
pandemic happened and the USChina tradescalations started amplifyinganti-Chinese rhetoric. All of these led
疫情爆發後，美中貿易升級開始加劇反華言論。這些因素共同導致
41:34.720–41:44.800
and now with the potential ban oninternational students. All of thesehave led more and more Chineseresearchers to just opt for staying at
而現在加上可能禁止國際學生入境，所有這些因素導致越來越多的中國研究人員選擇留在
41:44.800–41:54.319
home and contributing to the Chinese AIecosystem. And this was a prerequisiteto High-fly pulling off Deepseek. If
國內，為中國的人工智慧生態系統做出貢獻。而這是High-fly成功推出Deepseek的前提條件。如果
41:54.319–42:03.280
there had not been that concentrationand buildup of AI talent in China, theyprobably would have had a much harder
中國沒有這樣集中和積累的人工智慧人才，他們在創新以繞過美國政府對他們實施的出口管制時，
42:03.280–42:11.040
time innovating around circumventingthese export controls that the USgovernment was imposing on them. But
在創新以繞過美國政府對他們實施的出口管制方面所花費的時間。但
42:11.040–42:21.680
because they now have a highconcentration of top talent, some of thetop talent globally,when those restrictions were imposed,
因為他們現在擁有高度集中的人才，其中一些是全球頂尖人才，當這些限制措施實施時，
42:21.680–42:29.839
they were able to innovate around them.So Deepseek is literally a product ofthis continuation of that alienation and
他們能夠在這些限制下進行創新。因此，Deepseek literally 是這種疏離感延續的產物，
42:29.839–42:38.160
with the US continuing to take thistance, it is just going to get worse.And as you mentioned, it's not justChinese researchers. I literally just
隨著美國繼續採取這種態度，情況只會越來越糟。正如你所提到的，這不僅僅是中國研究人員的問題。我剛剛
42:38.160–42:45.680
alked to a friend in academia that saidshe's considering going to Europe nowbecause she just cannot survive without
和一位學術界的朋友聊過，她說她現在考慮去歐洲，因為她無法在沒有
42:45.680–42:52.400
hat public funding. And Europe Europeancountries are seeing a criticalopportunity offering milliondoll
公共資金的情況下生存。歐洲國家正看到一個關鍵機會，提供百萬美元
42:52.400–42:59.599
packages. Come here, we'll give you alab. We'll give you millions of dollarsof funding. I mean this is the fastest
的套餐。來這裡，我們會給你實驗室，我們會給你數百萬美元的資金。我的意思是，這是這個國家人才外流最快的方式。>> 我的意思是，許多人說美國的人才外流就是他們的人才增益。是的。>> 這也讓我們想起歷史。你
42:59.599–43:09.839
way to brain drain this country.>> I mean what many are saying US's braindrain is their brain gain. Yes.>> And this also reminds us of history. You
這將是這個國家人才外流最快的方式。>> 我的意思是，許多人說美國的人才外流就是他們的人才流入。是的。>> 這也讓我們想起歷史。你
43:09.839–43:20.560
have the Chinese rocket scientist ChenShuen who in the 1950s was inexplicablyheld under house arrest for years and
然後艾森豪威爾將他驅逐回中國。他成為火箭科學之父，以及中國進入太空領域的關鍵人物。
43:20.560–43:28.800
then Eisenhower has him deported toChina. He becomes the father of rocketscience and uh China's entry into space.
他說他永遠不會再踏足美國，儘管最初那裡是他唯一想居住的地方。>> 是的。還有一位我認為是
43:28.800–43:34.400
And he said he would never again stepfoot into the United States even thoughoriginally that was the only place he
政府官員，一位美國政府官員說，這是美國犯過最愚蠢的錯誤。>> 嗯，我們談論人才外流和人才增益。好的。再次，嗯，一些更多的押韻，末日論者和
43:34.400–43:41.920
wanted to live.>> Yes. And there was a I believe agovernment official, a US governmentofficial who said that was the dumbest
嬰兒潮一代。嗯，我想談談人工智慧末日看起來是什麼樣子，意味著它如何將我們帶向末日，但也
43:41.920–43:53.040
mistake the US ever made.>> Um you we talk about the brain drain andthe brain gain. Okay. Again, uh somemore rhyming, the doomers and the
美國犯過最愚蠢的錯誤。我們談到人才外流和人才回流，好吧，再來一些押韻詞，末日派和嬰兒潮世代。
43:53.040–44:04.560
boomers. Um, I want to talk about whatan AI apocalypse looks like, meaning howit brings us to apocalypse, but also um
>> 這是一個很好的問題，我經常問嬰兒潮一代和末日論者這個問題。你能向我精確地闡述我們如何到達
44:04.560–44:13.349
how uh people say it could lead us to autopia. What are the two trackstrajectories?
那麼人們說它可能會帶我們走向烏托邦，這兩條軌跡是什麼呢
44:13.359–44:19.680
>> It's a great question and I ask boomersand doomers this all the time. Can youarticulate to me exactly how we get
做不到。這就是為什麼我稱之為類宗教。它確實基於信念。我的意思是，我和一位自認為是嬰兒潮一代的研究人員談話。
44:19.680–44:21.200
here? And the issue is that they
我說，你知道，他的眼睛睜得大大的，他真的興奮起來說，你知道，一旦我們達到通用人工智慧（AGI），遊戲結束。一切都變得完美。
44:21.200–44:30.079
cannot. And this is why I call it quasireligious. It really is based on belief.I mean, I was talking with oneresearcher who identified as a boomer.
我問他，我說，你能解釋一下通用人工智慧如何餵養那些現在桌上沒有食物的人嗎？而他說，「哦，
44:30.079–44:39.359
And I said, you know, he his his eyeswere wide and he he really lit upsaying, you know, once we get to AGI,game over. Everything becomes perfect.
你是在談論地板以及如何提升他們的生活質量。」我說，「是的，因為他們也是全人類的一部分。」
44:39.359–44:49.760
And I asked him, I was like, can youexplain to me how does AGI feed peoplethat haven't don't have food on thetable right now? And he was like, "Oh,
然後我問他，我說，你能不能解釋一下，AGI 怎麼餵飽那些現在連飯都吃不飽的人？然後他說，「喔，
44:49.760–44:59.680
you're talking about like the floorand how to elevate their qualityof life." And I was like, "Yes, becausethey are also part of all of humanity."
會幫助中產階級獲得更多的經濟機會。」我說，「好的，但這也是如何發生的呢？」他說，「嗯，一旦
44:59.680–45:04.640
And he was like, "I'm not really surehow that would happen, but I think itcould it could help the middle class get
這些到來，一旦我們擁有通用人工智慧，它僅僅創造數萬億美元的經濟價值，我們就可以直接給他們
45:04.640–45:09.920
more economic opportunity." And I waslike, "Okay, but how does that happen aswell?" And he was like, "Well, once
更多經濟機會。」然後我說：「好吧，但這是怎麼發生的呢？」他說：「嗯，一旦
45:09.920–45:15.839
these come once we have AGI and it canjust create trillions of dollars ofeconomic value, we can just give them
這些情況會在我們擁有通用人工智慧且它能創造數兆美元經濟價值時出現，我們可以就此給予他們
45:15.839–45:19.119
cash payouts." And I was like, who'sgiving them cash payouts? What
就像當你實際測試他們的邏輯時，它並不完全成立。對於末日論者來說，我的意思是，這
45:19.119–45:20.560
institutions are giving them? You know,
是一樣的，他們的信念最終是我在報導這本書時意識到的，他們相信通用人工智慧是可能的
45:20.560–45:28.319
like it it doesn't when you actuallytest their logic, it doesn't reallyhold. And with the doomers, I mean, it's
因為他們對人類大腦運作方式的信念。他們相信人類智能本質上是完全計算性的。所以如果你擁有足夠的
45:28.319–45:37.920
the same thing. like their belief isultimatelywhat I realized when reporting on thebook is they believe AGI is possible
數據和足夠的計算資源，你最終將能夠重建人類智能。這只是時間問題。對他們來說，
45:37.920–45:46.480
because of their belief of how the humanbrain works. They believe humanintelligence is inherently fullycomputational. So if you have enough
導致末日場景的原因是人類，我們通過溝通來學習和提高我們的智能。而溝通是低效的。我們經常誤解彼此。
45:46.480–45:53.040
data and you have enough computationalresources, you will inevitably be ableto recreate human intelligence. It's
因此對於人工智慧智能來說，它們將能夠通過作為數字智能彼此之間完美的溝通，迅速變得更聰明、更聰明、更聰明。所以
45:53.040–46:01.200
just a matter of time. And to them, thereason why there would that would leadto an apocalyptic scenario is humans, we
許多自認為是末日論者的人說，在宇宙歷史上，從未有一個物種
46:01.200–46:07.920
learn and improve our intelligencethrough communication. And communicationis inefficient. We miscommunicate all
優於另一個物種，一個能夠統治更優越物種的物種。所以他們認為最終人工智慧將進化為更高級的物種，然後開始統治我們
46:07.920–46:19.280
the time. And so for AI intelligences,they would be able to rapidly getsmarter and smarter and smarter byhaving perfect communication with one
然後也許決定徹底消滅我們。我在想你是否能談論任何國家的模型，而不是
46:19.280–46:26.800
another as digital intelligences. And somany of these people who selfidentify asdoomers say there has never been in the
作為數位智能彼此之間進行完美溝通。因此，許多自稱為末日論者的人表示，在宇宙歷史中，從未出現過一個優越於其他物種、能夠統治更優越物種的物種。所以他們認為，人工智能最終會進化成為一個更高級的物種，然後開始統治我們，接著可能決定將我們徹底消滅。我在想，你是否能談論任何國家的模型，而非
46:26.800–46:36.640
history of the the universe a speciesthat was superior to another species a aspecies that was able to rule over um a
公司，>> 正在開創民主控制人工智慧的方式。>> 我不認為這正在積極發生
46:36.640–46:45.280
more superior species. So they thinkthat ultimately AI will evolve into ahigher species and then start ruling us
現在。歐盟已經有了歐盟人工智慧法案，這是他們主要的立法，試圖開發基於風險、基於權利
46:45.280–46:54.319
and then maybe decide to get rid of usaltogether. I'm wondering if you cantalk about any model of a country, not a
的人工智慧部署治理框架。但對我來說，民主人工智慧治理的關鍵之一也是民主地
46:54.319–47:05.599
company,>> that is pioneering a way ofdemocratically controlled artificialintelligence.>> I don't think it's actively happening
公司，>> 正在開創一種民主控制人工智慧的途徑。>> 我不認為這正在積極發生
47:05.599–47:15.599
right now.The EU has had the EU AI act, which istheir major piece of legislation tryingto develop a riskbased, rightsbased
現在。歐盟已有歐盟人工智能法案，這是他們主要的立法，試圖開發基於風險、基於權利
47:15.599–47:27.760
framework for governing AI umdeployment.Buto me, one of the keys of democratic AIgovernance is also democratically
的框架來監管人工智慧部署。但對我來說，民主人工智慧治理的關鍵之一也是民主地
47:27.760–47:38.480
developing AI. And I don't think anycountry is really doing that. And what Imean by that is there are AI has asupply chain. It needs data. It needs
開發人工智慧。我不認為任何國家真正在做這件事。我的意思是，人工智慧有供應鏈。它需要數據。它需要
47:38.480–47:45.599
land. It needs energy. It needs water.And it also needs spaces in which thesecompanies need access to to then deploy
土地。它需要能源。它需要水。它也需要空間，讓這些公司能夠進入並部署
47:45.599–47:52.240
their technology. Schools, hospitals,government agencies. Silicon Valley hasdone a really good job over the last
他們的技術。學校、醫院、政府機構。矽谷過去十年做得很好，讓人們覺得他們集體擁有的資源是矽谷的。你知道，我經常與
47:52.240–47:59.599
decade of making people feel that theircollectively owned resources are SiliconValleys. You know, I have I talk with
朋友談話，他們說：「我們不再有數據隱私了。」所以，像什麼，什麼是更多數據給
47:59.599–48:05.920
friends all the time who say, "We don'thave data privacy anymore." So, like,what's more what's what is more data to
我時常和朋友聊天，他們說：「我們已經沒有資料隱私了。」所以，也就是說，對這些公司來說，更多的資料意味著什麼？
48:05.920–48:07.359
these companies? Like, I'm fine just
把他們所有的數據都給他們。但這些數據是你的。你知道，那些知識產權是作家和
48:07.359–48:16.560
giving them all of my data. But thatdata is yours. You know, thatintellectual property is the writers andartists intellectual property. That land
藝術家的知識產權。那些土地是社區的土地。那些學校是學生和老師的學校。那些醫院是醫生、護士和
48:16.560–48:24.480
is a community's land. Those schools arethe students and teachers schools. Thehospitals are the doctors and nurses and
是社區的土地。那些學校是學生和老師的學校。醫院是醫生、護士和
48:24.480–48:32.640
patients hospitals. These are all sitesof democratic contestation in thedeployment in the development and the
醫院是病患的醫院。這些都是民主爭奪的場域，在人工智慧的部署、開發與部署中。
48:32.640–48:38.720
deployment of AI. And just like thoseChilean water activists that we talkedabout who aggressively understood that
部署中。就像我們談到的那些智利水資源活動人士，他們強烈地理解
48:38.720–48:47.680
fresh water was theirs and theywere not willing to give it up unlessthey got some kind of mutuallybeneficial agreement for it. We need to
淡水是他們的，他們不願意放棄，除非他們獲得某種互惠互利的協議。我們需要
48:47.680–49:00.559
have that spirit in protectingour data, our land, our water, and ourschools so that companies inevitablywill have to adjust their approach
在保護我們的數據、土地、水和學校時擁有這種精神，這樣公司最終將不得不調整他們的方法
49:00.559–49:07.200
because they will no longer get accessto the resources they need or the spacesthat they need to deploy in. In 2022,
因為他們將不再獲得他們需要的資源或他們需要部署的空間。在2022年，
49:07.200–49:14.079
Karen, you wrote a piece for MITechnology Review headlined a new visionof artificial intelligence for the
卡倫，你為《麻省理工技術評論》寫了一篇標題為「為人民的人工智慧新願景」的文章。在新
49:14.079–49:22.160
people. In a remote rural town in NewZealand, an indigenous couple ischallenging what AI could be and who it
西蘭偏遠的鄉村小鎮，一對原住民夫婦正在挑戰人工智慧可能成為什麼以及它應該服務誰。他們是誰？這是一篇
49:22.160–49:24.640
should serve. Who are they? This was a
精彩的報導，我做的，這對夫婦經營著媒體。它是新西蘭的一家非營利性毛利語廣播電台。而毛利人已經
49:24.640–49:36.400
wonderful story that I did where thecouple um they run to media. It's anonprofit MAI radio station in NewZealand. And the Maui people have
這是一個很棒的報導，我做了這個報導，那對夫婦他們經營一家非營利的毛利語廣播電台，位於紐西蘭。而毛利人經歷了許多與全球其他原住民族相同的挑戰。
49:36.400–49:43.440
suffered a lot of the same um challengesas many indigenous peoples around theworld. The history of colonization led
他們迅速失去語言，現在世界上說毛利語的人非常少。所以在過去幾年裡
49:43.440–49:49.200
them to rapidly lose their language andthere are very few Mauy speakers in theworld anymore. And so in the last few
年，有人試圖復興這種語言，新西蘭政府試圖通過嘗試鼓勵這種語言的復興來贖罪。但這家非營利性廣播電台，他們擁有所有這些精彩的
49:49.200–49:54.960
years there's been an attempt to revivethe language and the New Zealandgovernment has tried to repent by by
近年來已有人嘗試復興該語言，而紐西蘭政府也試圖透過努力促進該語言的復興來贖罪
49:54.960–50:02.079
trying to encourage that the revival ofthat language. But this nonprofit radiostation, they had all this wonderful
世界各地的毛利語說者、毛利語學習者作為教育資源。問題是，為了做到這一點，他們
50:02.079–50:10.079
archival material, archival audio oftheir ancestors speaking the Mailanguage that they wanted to provide to
檔案資料，祖先說馬語的音檔，他們希望提供給
50:10.079–50:16.319
Maui speakers, ma Mai learners aroundthe world as an educational resource.The problem is in order to do that they
全球毛利語說客和學習者作為教育資源。問題在於，為了做到這一點，他們
50:16.319–50:22.800
needed to transcribe the audio so thatMai learners could actually listen, seewhat was being said, click on the words,
理解翻譯，並實際上將其變成一個積極的學習工具。但說這種高級別毛利語的人太少了，以至於
50:22.800–50:29.760
understand the translation and actuallyturn it into an active learning tool.But there were so few Maui speakers that
他們意識到自己必須轉向人工智慧。這是我書中論點的一個關鍵部分，我並非批評所有人工智慧
50:29.760–50:39.200
can speak at that advanced level thathey realized they had to turn to AI.And this is a key part of my book'sargument is I'm not critiquing all AI
能達到那個進階程度的茂宜語使用者少之又少，以致於他們不得不轉向人工智慧，而這也是我書中論點的核心：我並非批判所有的人工智慧
50:39.200–50:44.800
development. I'm specifically critiquingthe scale at all cost approach thatSilicon Valley has taken. But there are
許多不同類型的有益人工智慧模型，包括他們最終所做的。所以他們採取了根本不同的方法。首先，他們詢問社區，我們想要這個
50:44.800–50:50.880
many different kinds of beneficial AImodels, including what they ended updoing. So they took a fundamentally
許多不同類型的有益人工智慧模型，包括他們最終所採取的行動。因此，他們採取了根本不同的方法。
50:50.880–50:57.760
different approach. First and foremost,they asked their community, do we wanthis AI tool? Once the community said
是，然後他們進入下一步，要求人們完全同意捐贈數據以訓練這個
50:57.760–51:06.480
yes, then they moved to the next step ofasking people to fully consent todonating data for the training of this
工具。他們向社區解釋了這些數據的用途、如何使用，以及他們將如何保護該
51:06.480–51:12.559
tool. They explained to the communitywhat this data was for, how it would beused, how they would then guard that
數據，確保它不會被用於其他目的。他們在短短幾天內收集了大約幾百小時的音頻數據
51:12.559–51:20.640
data and make sure that it wasn't usedfor other purposes.They collected around a couple hundredhours of audio data in just a few days
並確保資料受到保護，且未被用於其他目的。他們在短短幾天內就收集了約數百小時的音訊資料
51:20.640–51:27.119
because the community rallied supportaround this project and only a couplehundred hours was enough to create a
因為社群為這個專案凝聚了支持，僅需幾百小時的數據就足以建立一個
51:27.119–51:33.040
performant speech recognition modelwhich is crazy when you think about thescales of data that these Silicon Valley
所需的數據規模時是瘋狂的。而這再次是一個可以吸取的教訓，實際上有很多研究
51:33.040–51:39.200
companies require. And that is onceagain a lesson that can be learned isactually there's plenty of research that
表明，當你擁有高度策劃的小數據集時，你實際上可以創建非常強大的人工智慧模型，然後一旦他們
51:39.200–51:46.800
shows when you have highly curated smalldata sets, you can actually create verypowerful AI models and then once they
有了這個工具，他們就能夠做他們想做的：開源並向他們的社區提供這個教育資源。
51:46.800–51:55.040
had that tool, they were able to doexactly what they wanted to open sourceand resour uh open source thiseducational resource to their community.
有了這個工具後，他們得以完全按照自己的意願，將這項教育資源以開源方式分享給社群。
51:55.040–52:08.800
And so my vision for AI development inthe future is to have more smalltaskspecific AI models that are notrained on vast polluted data sets but
因此，我對未來人工智慧發展的願景是擁有更多針對特定小型任務的人工智慧模型，它們並非在龐大且受污染的數據集上訓練而成，而是
52:08.800–52:18.240
small curated data sets and thereforeonly need small amounts of computationalpower and can be deployed in challenges
小型精心策劃的資料集，因此只需少量運算能力，且能部署於我們人類亟需克服的挑戰中
52:18.240–52:27.920
that we actually need to tackle forhumanity.mitigating climate change by integratingmore renewable energy into the grid,
我們人類真正需要解決的問題，包括透過將更多再生能源整合進電網來緩解氣候變遷，
52:27.920–52:37.599
improving health care, by doing moredrug discovery.>> So, as we finally do wrap up, what youwere most shocked by, you've been doing
改善醫療保健，透過進行更多藥物研發。>> 所以，當我們終於要結束時，你感到最震驚的是什麼，你一直在
52:37.599–52:49.599
uh this journalism, this research foryears, what you were most shocked by inwriting Empire of AI.I originally thought that I was going to
這些年來你從事新聞與研究工作，在撰寫《AI帝國》時，最讓你震驚的是什麼？我原本以為我會寫一本聚焦於AI供應鏈垂直危害的書籍，這裡展示了AI產業中的勞動力剝削如何發生，這裡展示了AI產業如何產生環境危害，而在我的報導結束時，我意識到這裡存在著一種橫向危害，我採訪過的每個社區，無論是藝術家
52:49.599–52:56.800
write a book focused on vertical harmsof the AI supply chain. Here's how laborexploitation happens in the AI industry.
撰寫一本聚焦於人工智慧供應鏈垂直危害的書籍。以下是人工智慧產業中勞動力剝削的運作方式。
52:56.800–53:04.559
Here's how the environmental harms arearising out of the AI industry. And athe end of my reporting, I realized that
這就是環境危害如何從人工智慧產業中產生的方式。而在我的報導結束時，我意識到
53:04.559–53:10.319
here is a horizontal harm that'shappening here. Every single communitythat I spoke to, whether it was artists
這裡存在一種橫向的傷害，在我採訪過的每個社區中，無論是藝術家
53:10.319–53:16.800
having their intellectual property takenor Chilean water water activists havingtheir fresh water taken, they all said
知識產權被剝奪的藝術家，或是新鮮水源被奪取的智利水資源活動人士，他們都說
53:16.800–53:24.640
that when they encountered the empire,they initially felt exactly the sameway. A complete loss of agency to
當他們遇到這個帝國時，最初感受到的都是同樣的情緒，一種完全喪失自主權的無力感
53:24.640–53:33.839
self-determine their future. And that iswhen I realized the horizontal harm hereis AI is threatening democracy. If the
自主決定自己的未來。就在這時我意識到這裡存在的橫向傷害是人工智慧正在威脅民主。如果
53:33.839–53:46.240
majority of the world is going to feelthis loss of agency overself-determining their future, democracycannot survive and again specifically
世界大多數的人都會感受到這種喪失自主決定未來能力的無力感，而民主將無法存续，特別是針對
53:46.240–53:55.599
Silicon Valley's approach scale at allcosts AI development.>> But you also chronicle the resistance.You talk about how the Chilean water
矽谷的開發策略是無所不用其極地追求規模。>> 但你同時也記錄了抵抗的行動。你談到智利的水資源
53:55.599–54:05.680
actors felt at first, how the artistsfeel at first. So talk about thestrategies that these people havemployed and if they've been effective.
他們最初有何感受，藝術家們最初又有何感受，請談談這些人採用的策略，以及這些策略是否有效。
54:05.680–54:12.400
>> So the amazing thing is that there hasince been so much push back. Theartists have then said wait a minute we
>> 所以令人驚訝的是，此後出現了如此強烈的反彈。藝術家們接著說，等等，我們
54:12.400–54:18.640
can sue these companies. The Chileanwater activists said wait a minute wecan fight back and protect these water
可以起訴這些公司。智利的水資源活動人士表示，我們可以反擊並保護這些水資源。
54:18.640–54:25.200
esources. The Kenyan workers that Ispoke to who are contracted by OpenAIthey said we can unionize and escalate
資源。我採訪過的肯亞工人由 OpenAI 僱用，他們表示我們可以組織工會並升級
54:25.200–54:35.680
our story to international mediattention. And so even in these evenwhen I thought that these communitiesyou could argue are the most vulnerable
將我們的故事推向國際媒體關注。因此，即使在我認為這些社區可能是最脆弱的時候，
54:35.680–54:45.040
in the world have the least amount ofagency, they were the ones thatremembered that they do have agency andthat they can seize that agency and
在全球擁有最少自主權的人群中，正是他們記起自己確實擁有自主權，並且能夠掌握這份自主權
54:45.040–54:54.960
fight back. And I think it it was it wasremarkably heartening to encounter thosepeople to remind me that actually the
反擊。我認為，遇到這些人並讓我想起其實這是非常令人振奮的
54:54.960–55:02.960
first step to reclaiming democracy isremembering that no one can take youragency away.>> Karen How, author of the new book Empire
奪回民主的第一步，就是記住沒有人能奪走你的自主權。>> 卡倫·霍，新書《帝國：人工智能的夢想與夢魘》作者
55:02.960–55:11.680
of AI: Dreams and Nightmares and SamAlman's Open AI. Karen How is a formereporter at the Wall Street Journal and
人工智慧的夢與夢魘，以及山姆·阿爾特曼的開放人工智慧。凱倫·豪曾是華爾街日報的記者，並
55:11.680–55:19.760
MIT Technology Review. And that does itfor this special broadcast. [music]Democracy Now is produced with Mike
麻省理工科技評論。本期特別節目到此結束。[音樂]民主新聞由麥克·伯克、蕾妮·費爾、迪娜·古斯特、梅西婭·羅德斯、努爾明·沙克、瑪麗亞·特雷塞納、妮可·薩拉薩、[音樂]薩林·納賽爾、特里娜·納杜拉、薩姆·阿洛夫、T·瑪麗亞·喬、約翰·漢密爾頓、羅比、[音樂]卡倫、霍尼·馬蘇德和薩夫韋特·納茲共同製作。我們的執行總監是茱莉·克羅斯比。特別感謝貝卡·斯蒂利、約翰·蘭道夫、保羅·鮑威爾、[音樂]麥克
55:19.760–55:25.040
Burke, Renee Fel, Dina Guster, MessiahRhodess, Nurmine Shake, Maria Terasena,Nicole Salazar, [music] Sarin Nasser,
伯克、蕾妮·費爾、迪娜·古斯特、梅西婭·羅德斯、努爾明·沙克、瑪麗亞·特拉塞納、妮可·薩拉薩、[音樂] 薩林·納賽爾、
55:25.040–55:31.599
Trina Nadura, Sam Alov, T Maria Joe,John Hamilton, Robbie [music] Karen,Honey Massud, and Safwet Naz. Our
崔娜·納杜拉、薩姆·阿洛夫、T·瑪麗亞·喬、約翰·漢密爾頓、羅比 [音樂] 凱倫、哈尼·馬蘇德和薩夫韋特·納茲。我們的
55:31.599–55:37.280
executive directors, Julie Crosby.Special thanks to Becca Stelli, JohnRandolph, Paul Powell, [music] Mike
執行董事茱莉·克羅斯比。特別感謝貝卡·斯特利、約翰·蘭道夫、保羅·鮑威爾、[音樂] 麥克
55:37.280–55:44.160
DeFippo, Miguel Nggera, Hugh Grant, CarlMarxer, Dennis Moahan, David Prud,Dennis McCormack, Matt Elely, Anna
德菲波、米格爾·恩格拉、休·葛蘭、卡爾·馬克瑟、丹尼斯·莫漢、大衛·普魯德、丹尼斯·麥考馬克、馬特·埃萊利、安娜
55:44.160–55:52.799
Osbeck, Emily Anderson, Dante Toriieri,Buffy [music] St. Marie Hernandez withJuan Gonzalez. I'm Amy Goodman. Happy
奧斯貝克、艾米麗·安德森、但丁·托里耶里、巴菲 [音樂] 聖·瑪麗·埃爾南德斯與胡安·岡薩雷斯。我是艾米·古德曼。新年快樂
55:52.799–56:01.920
New Year.Thanks for watching Democracy Now onYouTube. Subscribe to the channel andturn on notifications to make sure you
新年快樂。感謝收看 Democracy Now 的 YouTube 頻道。訂閱頻道並開啟通知，確保您
56:01.920–56:08.559
never miss a video. And for more of ouraudience supported journalism, go todemocracynow.org or where you can
絕不錯過任何影片。若想支持更多由觀眾資助的我們的新聞報導，請前往democracynow.org或
56:08.559–56:15.000
download our news app, sign up for ournewsletter, subscribe to the dailypodcast, and so much
下載我們的新聞應用程式、訂閱我們的電子報、訂閱每日播客，還有更多

影片筆記：“Empire of AI”: Karen Hao on How AI Is Threatening Democracy & Creating a New Colonial World

<!-- 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

記者 Karen Ho 在《Democracy Now!》專訪中，將其新書《The Empire of AI》的核心論點闡述為：AI 產業正透過掠奪資源（數據、能源、水）、剝削全球南方勞工、與軍事及政治權力結盟，複製過去殖民帝國的擴張邏輯，並透過「規模至上」的開發模式威脅民主與社區主權。

核心重點

  • AI 產業的帝國主義特徵：
  • 將 AI 產業發展類比為殖民帝國擴張，雖無公開暴力，但透過掠奪資源（數據、土地、能源、淡水）實現願景。
  • 批判矽谷「不惜代價追求規模」（scale at all costs）的發展軌跡。
  • 資源消耗巨大：未來 5 年 AI 基礎設施擴張所需的能源相當於加州年消費量的 2 到 6 倍，且多依賴化石燃料；全球 2/3 的數據中心建於缺水地區，直接消耗公共飲用水。
  • 政治與軍事化的緊密結合：
  • 特朗普政府放寬 AI 監管，並與 AI 產業緊密勾結（如 Trump Media 與核融合公司的合併）。
  • 推出價值 5000 億美元的「Stargate 計畫」（實為私人投資），旨在強化美國 AI 基礎設施。
  • AI 公司為回收高昂研發成本，轉向國防產業尋求合約，將技術整合進軍事基礎設施（「與國防產業親近」）。
  • OpenAI 推出「國家版 OpenAI」（OpenAI for countries）計畫，被批評為缺乏公眾諮詢的「技術威權主義」。
  • 全球南方的勞動剝削與環境正義問題：
  • 肯亞案例：OpenAI 僱用肯亞合約勞工進行數據標註（Data Annotation）與內容審核，勞工閱讀有害內容面臨心理創傷與低薪，對比矽谷高管高薪。
  • 智利案例：Google 試圖在聖地亞哥附近建立數據中心，引發當地社區對淡水資源被奪取的抗議，最終迫使政府介入協商，項目延遲 4-5 年。
  • 孟菲斯案例：Elon Musk 的 Colossus 超級電腦使用未獲許可的甲烷燃氣渦輪機，排放有毒污染物。
  • AGI 的「準宗教」性質與內部衝突：
  • AGI（人工通用智慧）缺乏科學共識，更像是一種基於信念的「準宗教運動」。
  • 內部存在「烏托邦派」（Boomers，認為 AGI 帶來巨大經濟價值）與「毀滅派」（Doomers，認為 AI 將統治或消滅人類）的衝突。
  • Sam Altman 擅長根據聽眾調整說法，利用非營利身分吸引人才與公眾善意，隨後透過複雜結構籌措資金，成為矽谷最具資本主義色彩的企業之一。
  • 中美競爭與 DeepSeek 的挑戰：
  • 美國透過出口管制限制中國獲取先進晶片，但中國公司 DeepSeek 以極低算力成本（約 600 萬美元）開發出與美國公司相當能力的模型。
  • 證明「堆積算力」並非唯一路徑，中國在資源受限下展現了工程創新能力。
  • 美國政策（如「中國行動計畫」）導致人才回流中國，反而促進了中國 AI 生態系的崛起。
  • 民主化 AI 治理與社區代理權：
  • 強調社區對數據、土地、水資源的主權，拒絕無條件讓企業獲取資源。
  • 紐西蘭毛利人案例：Maui Radio 透過社區同意、使用小數據集（幾百小時音頻）開發特定用途 AI 以復興語言，證明小數據、特定任務的 AI 比「規模至上」模式更有效且資源需求低。
  • 受影響群體（藝術家、智利活動家、肯亞工人）展現了「代理權」（agency），透過起訴、組織工會等方式反擊。

詳細大綱

1. 書籍介紹與核心論點

  • Karen Ho 介紹新書《The Empire of AI: Dreams and Nightmares in Sam Alman's Open AI》。
  • 將 AI 產業比作「帝國」，透過掠奪資源實現願景。
  • 批判矽谷「scale at all costs」的發展軌跡。

2. 能源與水資源消耗

  • 能源：
  • 未來 5 年 AI 基礎設施擴張所需的能源，相當於加州年消費量的 2 到 6 倍。
  • 主要依賴化石燃料，導致煤電廠壽命延長、未獲許可的天然氣燃氣輪機出現。
  • 淡水：
  • 數據中心需使用淡水以防設備腐蝕與細菌生長，通常直接連接公共飲用水供應。
  • Bloomberg 分析顯示，全球 2/3 的數據中心建於缺水地區。

3. 軍事化與政治背景

  • 特朗普政府政策：簽署行政命令禁止州與地方政府制定自己的 AI 監管法規。
  • 商業合併：Trump Media and Technology 宣布與一家旨在建造全球首個可行核融合廠的公司進行 60 億美元合併，以供應 AI 項目電力。
  • 國防合約：AI 公司為回收數千億美元的研發成本，與國防產業緊密結合（cozying up to the defense industry），將技術整合進軍事基礎設施。
  • Stargate 計畫：特朗普宣布 5000 億美元 AI 基礎設施計畫，資金來源為私人投資（如軟銀集團）。

4. 全球南方的勞動剝削

  • 數據標註（Data Annotation）：
  • 從互聯網抓取的海量數據需經過清洗、標註才能訓練模型。
  • 中間商公司（middleman firms）僱用合約勞工進行內容準備。
  • 肯亞案例：
  • Open AI 僱用肯亞勞工閱讀最糟糕的互聯網文本及 AI 生成的有害內容。
  • 勞工需將內容分類（性內容、暴力內容等）以訓練過濾系統。
  • 低薪（幾美元一小時）與心理創傷，對比矽谷研究人員的高額薪酬，體現「帝國的邏輯」。

5. 智利的水資源爭議

  • 背景：智利曾長期處於獨裁統治，公共資源（包括水）私有化。
  • 衝突：Google 試圖在聖地亞哥大都會區的一個社區建立數據中心，該社區擁有公共淡水資源。
  • 社區抗爭：
  • 居民未從 Google 獲得稅收利益，且擔心淡水被無限制抽取。
  • 發起地面行動（boots on the ground activism），最終迫使 Google 代表（僅會英語）及智利政府介入。
  • 結果：項目被延遲 4-5 年，社區獲得談判席位，但抗爭仍未結束。

6. Sam Alman 與 Open AI 的軌跡

  • Sam Alman 的背景：矽谷創始人、Y Combinator 總裁、Open AI CEO。
  • 其發展模式反映矽谷「規模至上」的文化。
  • Alman 擅長講述未來故事，吸引投資者與員工。
  • 早期在 YC 識別 AI 為趨勢，並建立相關投資組合。

7. OpenAI 的資本主義演變

  • Sam Altman 的投資策略：量子計算、核融合、自動駕駛車、基礎 AI 實驗室。
  • 從非營利轉向營利：為追求規模與速度，建立包含營利部門的結構以籌措數百億資金。
  • 結果：成為矽谷歷史中最資本主義的公司之一，持續籌資以生產目前經濟影響中等的技術。

8. AI 對就業的影響與定義

  • OpenAI 對 AGI 的定義：高度自主系統，在大多數經濟價值工作中表現優於人類。
  • 勞動自動化（Labor Automating）：旨在取代員工，降低企業成本。
  • 勞動輔助（Labor Assistive）：旨在協助員工（如醫生、教師），提升服務品質與結果。
  • 現狀：企業因誤解技術能力而裁員，導致白領階級職涯階梯斷裂，初級職位減少。

9. AGI 的科學與社會屬性

  • AGI 缺乏科學基礎：目前對人類智慧尚無科學共識，AGI 更多是一種信念。
  • 準宗教運動（Quasi Religious Movements）：矽谷內部的 AGI 信仰者分為兩派，但都認為必須控制技術而非民主化。
  • Altman 的溝通策略：對不同派系說他們想聽的話，以維持領導地位。

10. 中美 AI 競爭與 DeepSeek 的挑戰

  • 美國透過出口管制限制中國獲取 Nvidia 等先進晶片。
  • 中國在算力受限下，展現出人才密度與資本投資的競爭力。
  • DeepSeek 案例：
  • 開發 Deep Seek 模型。
  • 訓練成本僅約 600 萬美元，遠低於美國公司的數億至數百億美元。
  • 證明「堆積算力」並非提升 AI 能力的唯一路徑。
  • 中國公司透過工程技巧與現有技術組合，達成同等能力。

11. 人才流動與地緣政治影響

  • 美國「中國行動計畫」（China Initiative）將學術人員刑事化，導致中國研究人員回流。
  • 中國 AI 人才聚集是 Deepseek 能繞過美國出口管制並創新的前提。
  • 歐洲透過高額資金（Million-dollar packages）吸引人才，加速美國的人才外流。
  • 歷史類比：錢學森（Chen Shuen）被驅逐回中國，成為火箭科學之父，並拒絕再踏足美國。

12. 民主化 AI 治理與資源控制

  • 民主化 AI 不僅是治理，更包含民主化開發。
  • AI 供應鏈需要數據、土地、能源、水及部署空間（學校、醫院、政府機構）。
  • 強調社區對數據、土地、水資源的主權，拒絕無條件讓企業獲取資源。
  • 參考智利水資源活動家的模式，透過集體協商保護資源。

13. 紐西蘭毛利人（Māori）的 AI 實踐案例

  • 背景：毛利語言瀕危，非營利組織 Maui Radio 希望復興語言。
  • 挑戰：缺乏高級別毛利語講者進行語音轉錄，需依賴 AI。
  • 方法：
  • 首先詢問社區是否需要該 AI 工具。
  • 獲得社區完全同意後，捐贈數據進行訓練。
  • 僅使用約幾百小時的音頻數據，而非矽谷所需的龐大數據集。
  • 結果：成功開發出高效能的語音識別模型，並開源教育資源。
  • 啟示：小數據集、特定任務（Task-specific）的 AI 模型比「規模至上」模式更有效，且計算資源需求較低。

14. 受影響群體的反擊與代理權展現

  • 藝術家：威脅起訴相關公司。
  • 智利水資源活動人士：反擊以保護水資源。
  • 肯亞合約工人：受 OpenAI 合約約束，組織工會並將故事推向國際媒體關注。
  • 核心觀點：即使是最脆弱、代理權最少的群體，也展現了奪回代理權並反擊的能力。
  • 民主與代理權：恢復民主的第一步是記住沒有人能奪走你的代理權。

工具 / 模型 / 名詞整理

  • 書籍/作品：
  • 《The Empire of AI: Dreams and Nightmares in Sam Alman's Open AI》（AI帝國：Sam Alman 的 Open AI 中的夢與 nightmare）
  • 《Democracy Now!》（節目名稱）
  • 《The War and Peace Report》（節目單元名稱）
  • 《Power in Progress》（書籍名稱）
  • 公司/組織：
  • Open AI
  • Trump Media and Technology
  • Wall Street Journal
  • MIT Technology Review
  • National Book Critic Circle
  • McKinzie（疑點，可能指 McKinsey）
  • Google / Google Chile / Google Mountain View
  • Y Combinator
  • Bloomberg
  • SoftBank
  • Meta
  • Greater Kudu LLC
  • Nvidia
  • Deep Seek
  • Maui Radio
  • Democracy Now
  • 產品/技術：
  • ChatGPT
  • AI models / AI models
  • Supercomputers
  • Data centers
  • Nuclear fusion plant
  • Data annotation firms
  • Content moderation / Filter
  • Stargate project
  • AGI (Artificial General Intelligence)
  • Grok
  • Colossus
  • OpenAI for countries
  • Deep Seek
  • GPT
  • EU AI Act
  • 人物：
  • Amy Goodman
  • Karen How（疑點，聽寫可能為 Karen Ho）
  • Sam Alman（疑點，聽寫可能為 Sam Altman）
  • Juan Gonzalez
  • Elon Musk
  • President Trump
  • Masayoshi Son（逐字稿提及為 Masay Yoshi Son）
  • Jeron Austin（疑點，MIT 經濟學家）
  • Simon Johnson
  • Alan Turing
  • Chen Shuen（疑點，可能指錢學森 Qian Xuesen）
  • Renee Fel（疑點）
  • Messiah Rhodess（疑點）
  • Nurmine Shake（疑點）
  • Maria Terasena（疑點）
  • Sarin Nasser（疑點）
  • Trina Nadura（疑點）
  • Sam Alov（疑點）
  • T Maria Joe（疑點）
  • Robbie [music] Karen（疑點）
  • Honey Massud（疑點）
  • Safwet Naz（疑點）
  • Julie Crosby（疑點）
  • Becca Stelli（疑點）
  • John Randolph（疑點）
  • Paul Powell（疑點）
  • Mike DeFippo（疑點）
  • Miguel Nggera（疑點）
  • Hugh Grant（疑點）
  • Carl Marxer（疑點）
  • Dennis Moahan（疑點）
  • David Prud（疑點）
  • Dennis McCormack（疑點）
  • Matt Elely（疑點）
  • Anna Osbeck（疑點）
  • Emily Anderson（疑點）
  • Dante Toriieri（疑點）
  • Buffy [music] St. Marie Hernandez（疑點）

操作流程整理

  • AI 開發與訓練流程：
  1. 從互聯網抓取海量數據。
  2. 數據清洗與標註（Data Annotation），由合約勞工（如肯亞勞工）進行內容分類（性內容、暴力內容等）。
  3. 訓練神經網絡進行統計計算，識別數據模式（如語言語法）。
  4. 從單晶片擴展到數十萬晶片的超級電腦進行訓練。
  5. 部署數據中心，消耗大量能源與淡水。
  • 社區抗爭與治理流程：
  1. 社區識別資源（如淡水）被企業（如 Google）威脅。
  2. 發起地面行動（boots on the ground activism）。
  3. 迫使企業代表及政府介入協商。
  4. 獲得談判席位，延遲或改變項目進程。
  5. 毛利人案例中，先詢問社區是否需要 AI 工具，獲得同意後捐贈小數據集進行訓練，開發特定用途模型。

值得注意的限制或風險

  • 環境風險：
  • AI 基礎設施擴張所需的能源巨大，依賴化石燃料導致煤電廠壽命延長、未獲許可的天然氣燃氣輪機出現。
  • 數據中心消耗大量淡水，全球 2/3 建於缺水地區，引發社區資源爭奪。
  • 排放有毒空氣污染物（如孟菲斯案例）。
  • 社會與勞動風險：
  • 全球南方勞工面臨低薪與心理創傷（如肯亞數據標註勞工）。
  • 企業誤解技術能力而裁員，導致白領階級職涯階梯斷裂，初級職位減少。
  • 「勞動自動化」取代員工，而非「勞動輔助」協助員工。
  • 政治與民主風險：
  • 特朗普政府放寬 AI 監管，與 AI 產業緊密勾結。
  • AI 公司與國防產業緊密結合，技術被用於軍事背景。
  • 「技術威權主義」：缺乏公眾諮詢的 AI 部署（如 OpenAI for countries）。
  • 出口管制與人才政策導致人才流動變化，影響全球 AI 產業格局。
  • 矽谷「規模至上」的 AI 開發模式威脅民主，因為多數人喪失了自我決定的能力（

生字列表

生字讀音類型中文
overt/ˈoʊ.vɚt/公開的、顯著的、不隱瞞的
seize/siːz/verb攫取、奪取、抓住
extract/ɪkˈstrækt/verb提取、採掘、取出
trajectory/trəˈdʒæk.tər.i/noun軌跡、發展軌道、途徑
bottleneck/ˈbɑː.t̬l.nek/noun瓶頸、阻礙
alarm/əˈlɑːrm/verb使驚慌、使擔憂
escalate/ˈesk.ə.leɪt/verb升級、擴大（衝突或情況）
vigilant/ˈvɪdʒ.ə.lənt/adjective警惕的、警戒的
inherently/ɪnˈhɪr.ənt.li/adverb本質上、固有地
curate/kjʊˈreɪt/verb策劃、精選、整理
agency/ˈeɪ.dʒən.si/noun自主權、行動能力、代理權
rally/ˈræl.i/verb團結、動員、恢復
repent/rɪˈpent/verb懺悔、悔悟
alienate/ˈeɪ.li.eɪt/verb疏遠、使疏離
inexplicably/ˌɪn.ekˈsplɪk.ə.bli/adverb無法解釋地、莫名其妙地
articulate/ɑːrˈtɪk.jə.leɪt/verb清楚表達、闡述

生字解說

overt /ˈoʊ.vɚt/

· C1

意思：公開的、顯著的、不隱瞞的

解說：指某種行為或特徵是明顯可見的，並非隱藏或秘密進行的。在原文中用來形容過去殖民歷史中那些「公開的暴力與殘酷」。

影片原句
The empires of AI are not engaged in the same overt violence and brutality that marked this history
AI帝國並未參與同樣標誌著這段歷史的公開暴力與殘酷行為
延伸例句
The company faced overt criticism for its environmental policies.
該公司因其環境政策而受到公開批評。

seize /siːz/

verb · B2

意思：攫取、奪取、抓住

解說：指用力抓住或迅速取得某物，常帶有強制性或侵略性的語氣。原文用來描述AI產業對資源的掠奪行為。

影片原句
but they too seize and extract precious resources to feed their vision of artificial intelligence
但它們同樣掠奪並提取珍貴資源以餵養他們對人工智慧的願景
延伸例句
The rebels seized control of the capital city overnight.
叛軍在一夜之間奪取了首都的控制權。

extract /ɪkˈstrækt/

verb · B2

意思：提取、採掘、取出

解說：指從某物中取出或獲得某物，通常經過努力或複雜的過程。原文指從自然環境中提取水資源和能源。

影片原句
but they too seize and extract precious resources to feed their vision of artificial intelligence
但它們同樣掠奪並提取珍貴資源以餵養他們對人工智慧的願景
延伸例句
The company extracts oil from deep beneath the ocean floor.
該公司從深海海底下方採掘石油。

trajectory /trəˈdʒæk.tər.i/

noun · C1

意思：軌跡、發展軌道、途徑

解說：指物體運動的路徑或事物發展的過程與方向。原文用來批判AI發展所採取的特定路線。

影片原句
it's actually a critique of the specific trajectory of AI development that led us to chatbt
實際上是对導致我們走向ChatGPT的特定AI發展軌跡的批判
延伸例句
The trajectory of the economy has shifted towards sustainability.
經濟的發展軌跡已轉向可持續性。

bottleneck /ˈbɑː.t̬l.nek/

noun · B2

意思：瓶頸、阻礙

解說：指限制整體進程或產出的狹窄部分或障礙。原文指出資本是限制AI發展規模的關鍵瓶頸。

影片原句
they realized the bottleneck was capital
就意識到瓶頸是資本
延伸例句
Lack of skilled workers is a bottleneck for the industry's growth.
缺乏熟練工人是該行業增長的瓶頸。

alarm /əˈlɑːrm/

verb · B2

意思：使驚慌、使擔憂

解說：指引起某人的恐懼或擔憂。原文形容將技術整合進軍事基礎設施的做法令人擔憂。

影片原句
integrate their technologies more and more to the into the infrastructure of the military is really alarming
將其技術越來越深入地整合到軍事基礎設施中，這確實令人擔憂
延伸例句
The rapid spread of the virus alarmed health officials.
病毒的迅速傳播讓衛生官員感到擔憂。

escalate /ˈesk.ə.leɪt/

verb · B2

意思：升級、擴大（衝突或情況）

解說：指使情況、衝突或行動變得更加嚴重或激烈。原文描述抗議活動從社區層面升級到政府層面。

影片原句
they escalated so dramatically that it escalated to Google Chile
因此他們大規模升級行動，升級到了智利谷歌
延伸例句
The dispute between the two countries escalated into a war.
兩國之間的爭端升級為戰爭。

vigilant /ˈvɪdʒ.ə.lənt/

adjective · C1

意思：警惕的、警戒的

解說：指保持警覺以發現潛在危險或問題。原文強調活動人士必須保持警惕以監督專案。

影片原句
They need to stay vigilant
他們需要保持警惕
延伸例句
Vigilant drivers are essential for road safety.
警惕的駕駛員對道路安全至關重要。

inherently /ɪnˈhɪr.ənt.li/

adverb · C1

意思：本質上、固有地

解說：指某事物內在的、不可分離的特性。原文用來描述人類智能本質上是計算性的這一信念。

影片原句
They believe human intelligence is inherently fully computational
他們相信人類智能本質上是完全計算性的
延伸例句
This problem is inherently difficult to solve.
這個問題本質上很難解決。

curate /kjʊˈreɪt/

verb · C1

意思：策劃、精選、整理

解說：指仔細選擇、組織和呈現事物（如藝術品或數據）。原文指精心策劃的小數據集。

影片原句
when you have highly curated small data sets
當你擁有高度策劃的小數據集時
延伸例句
She curated a collection of rare books for the library.
她為圖書館策劃了一套珍稀書籍收藏。

agency /ˈeɪ.dʒən.si/

noun · C1

意思：自主權、行動能力、代理權

解說：指個人或群體做出選擇並採取行動影響自身生活的能力。原文強調社區喪失自主決定未來的能力。

影片原句
A complete loss of agency to self-determine their future
一種完全喪失自主權的無力感自主決定自己的未來
延伸例句
Empowering women increases their agency in society.
賦權女性增加了她們在社會中的自主權。

rally /ˈræl.i/

verb · B2

意思：團結、動員、恢復

解說：指聚集人群以支持某項事業或目標。原文指社群為專案凝聚支持。

影片原句
because the community rallied support around this project
因為社群為這個專案凝聚了支持
延伸例句
The community rallied to help the family after the fire.
火災後，社群團結起來幫助這家人。

repent /rɪˈpent/

verb · C1

意思：懺悔、悔悟

解說：指對過去的錯誤或罪行感到後悔並試圖補償。原文指紐西蘭政府試圖通過語言復興來贖罪。

影片原句
the New Zealand government has tried to repent by by trying to encourage that the revival of that language
而紐西蘭政府也試圖透過努力促進該語言的復興來贖罪
延伸例句
He sought to repent for his past mistakes by volunteering.
他通過志願服務來為過去的錯誤懺悔。

alienate /ˈeɪ.li.eɪt/

verb · C1

意思：疏遠、使疏離

解說：指使某人感到不被歡迎或與群體隔閡。原文指美國政府疏遠關鍵人才庫。

影片原句
to really alienate a key talent pool for US innovation
旨在疏遠對美國創新至關重要的關鍵人才庫
延伸例句
Harsh policies can alienate potential customers.
嚴苛的政策可能會疏遠潛在客戶。

inexplicably /ˌɪn.ekˈsplɪk.ə.bli/

adverb · C1

意思：無法解釋地、莫名其妙地

解說：指某事發生卻沒有明顯或合理的原因。原文形容陳省身被軟禁的原因無法解釋。

影片原句
who in the 1950s was inexplicably held under house arrest for years
在1950年代被莫名其妙地軟禁多年
延伸例句
He disappeared inexplicably from the public eye.
他從公眾視野中莫名其妙地消失了。

articulate /ɑːrˈtɪk.jə.leɪt/

verb · C1

意思：清楚表達、闡述

解說：指清晰且有邏輯地表達想法或觀點。原文指無法清楚說明如何從當前狀態到達未來願景。

影片原句
Can you articulate to me exactly how we get here?
你能向我精確地闡述我們如何到達這裡嗎？
延伸例句
She struggled to articulate her feelings about the change.
她難以清楚表達她對變化的感受。

句型解說（含實例）

not necessarily because... but because...

意思：不一定是因為……而是因為……

接續：not necessarily because + [reason A] + but because + [reason B]

解說：用於否定一個表面或常見的原因，並指出真正的原因。強調後者才是導致結果的關鍵因素。

影片原句
Not necessarily because the technology itself is really capable of replacing jobs, but it is perceived as capable enough that executives are laying off workers.
這不一定是因為技術本身真的有能力取代工作，而是人們認為它有能力做到這一點，因此高層主管正在裁員。
實例
  1. The failure was not necessarily because of lack of effort, but because of poor planning.
    失敗不一定是因為缺乏努力，而是因為規劃不周。
  2. His success was not necessarily because of talent, but because of relentless practice.
    他的成功不一定是因為天賦，而是因為不懈的練習。

the same... that...

意思：同樣的……那……

接續：the same + [noun] + that + [clause]

解說：用於將當前情況與過去或另一種情況進行比較，指出兩者具有相同的特徵或性質。

影片原句
Karen how compares the actions of the AI industry to those of colonial powers of the past. She writes, quote, 'The empires of AI are not engaged in the same overt violence and brutality that marked this history'
卡倫·豪將這些行動與標誌著這段歷史的同樣公開暴力與殘酷相比較
實例
  1. It is the same problem that caused the previous system to crash.
    這是導致先前系統崩潰的同樣問題。
  2. They faced the same challenges that we encountered last year.
    他們面臨著我們去年遇到的同樣挑戰。

once... [result clause]

意思：一旦……就……

接續：once + [condition/event], + [result]

解說：表示某個條件或事件發生後，另一個結果隨即或必然發生。強調時間上的緊接性或因果關係。

影片原句
And he was like, 'Well, once these come once we have AGI and it can just create trillions of dollars of economic value, we can just give them cash payouts.'
他說：「嗯，一旦這些情況會在我們擁有通用人工智慧且它能創造數兆美元經濟價值時出現，我們可以就此給予他們現金補貼。」
實例
  1. Once you understand the basics, the rest will be easy.
    一旦你理解了基礎知識，其餘的就會很容易。
  2. Once the decision is made, there is no turning back.
    一旦決定做出，就無法回頭。

not only... but also...

意思：不僅……而且……

接續：not only + [element A] + but also + [element B]

解說：用於連接兩個並列的成分，強調兩者都具備，且後者往往更重要或更具遞進性。

影片原句
Karen, could you talk some more about not only the energy requirements but the water requirements of these huge data centers
卡倫，你能否再多談談這些龐大資料中心不僅是能源需求，還有水資源需求
實例
  1. The new policy not only reduces costs but also improves efficiency.
    新政策不僅降低成本，還提高了效率。
  2. She is not only a talented musician but also a skilled painter.
    她不僅是一位有才華的音樂家，還是一位熟練的畫家。

whether it was... or...

意思：無論是……還是……

接續：whether it was + [option A] + or + [option B]

解說：用於列舉不同的情況或對象，表示這些情況都適用於後續的描述或結論。

影片原句
Every single community that I spoke to, whether it was artists having their intellectual property taken or Chilean water activists having their fresh water taken
在我採訪過的每個社區中，無論是藝術家知識產權被剝奪的藝術家，或是新鮮水源被奪取的智利水資源活動人士
實例
  1. Whether it was rain or shine, the event continued.
    無論下雨還是晴天，活動都繼續進行。
  2. Whether it was a small gift or a large donation, the gesture was appreciated.
    無論是小禮物還是大捐贈，這份心意都受到感激。

try to [verb] by [gerund/phrase]

意思：試圖通過……來……

接續：try to + [verb] + by + [gerund or prepositional phrase]

解說：描述為了達到某個目標所採取的方法或手段。

影片原句
the New Zealand government has tried to repent by by trying to encourage that the revival of that language
而紐西蘭政府也試圖透過努力促進該語言的復興來贖罪
實例
  1. He tried to improve his health by exercising regularly.
    他試圖通過定期鍛鍊來改善健康。
  2. The company tried to reduce waste by recycling materials.
    該公司試圖通過回收材料來減少廢棄物。

just [verb] [object]

意思：只是……（強調簡單性或輕描淡寫）

接續：just + [verb] + [object]

解說：用於簡化描述一個動作，暗示其簡單性或無需過多解釋。原文中用於描述給予現金補貼的簡單性（反諷）。

影片原句
we can just give them cash payouts
我們可以就此給予他們現金補貼
實例
  1. You just need to press the button to start.
    你只需要按下按鈕即可啟動。
  2. Just ask for help if you are unsure.
    如果不確定，只需尋求幫助。

the [noun] of [noun]

意思：……的……

接續：the + [abstract noun] + of + [concrete noun/phrase]

解說：用於構成名詞短語，表達抽象概念與具體事物之間的關係，常作為主題或核心論點。

影片原句
That is the logic of Empire
這就是帝國的邏輯
實例
  1. The cost of living is rising rapidly.
    生活成本正在迅速上升。
  2. The future of education depends on technology.
    教育的未來取決於技術。