硅谷寡頭們擔(dān)憂他們的技術(shù)會給世界帶來風(fēng)險，卻忘記了普通民眾。

信用...

大衛(wèi)·華萊士-威爾斯

2026年5月8日

“我為生存做好準(zhǔn)備，”O(jiān)penAI 的 Sam Altman 在 2016 年坦言。“我有槍支、碘化鉀金、抗生素、電池、水、以色列國防軍的防毒面具，還有大蘇爾的一大片我可以飛過去的地方。”

十多年來，奧特曼和他的聯(lián)合創(chuàng)始人一直生活在對人工智能既焦慮又克制的焦慮之中。事實上，這正是人工智能軍備競賽五大巨頭中三家公司的起源故事，它們都源于對其他公司未能足夠重視這項技術(shù)帶來的生存危機(jī)的恐慌。他們似乎不太擔(dān)心來自人類的政治反彈，認(rèn)為這種反彈不會及時出現(xiàn)，或者會被機(jī)器智能迅速超越，又或者可以用全民基本收入或治愈癌癥之類的空洞承諾來收買。

但上個月，當(dāng)有人向奧特曼位于舊金山的住所投擲燃燒彈時，民眾的強(qiáng)烈反彈直接波及到了他家門口。幾天后，奧特曼的家再次遭到襲擊，這次是槍擊。人們很難不聯(lián)想到聯(lián)合健康保險公司首席執(zhí)行官布萊恩·湯普森被殺一案，路易吉·曼吉奧內(nèi)被控犯有此罪。作家賈斯敏·孫稱之為“人工智能民粹主義的警告”。

美國民眾仍然擔(dān)憂數(shù)據(jù)中心對當(dāng)?shù)氐挠绊懀娂娪肯蚴姓d進(jìn)行抗議。他們也同樣擔(dān)憂失業(yè)和經(jīng)濟(jì)動蕩，越來越多的政客也對此憂心忡忡，卻對未來走向漠不關(guān)心。但對許多人來說，如今最大的人工智能實驗室也如同美國寡頭政治的新面孔——一種令人畏懼的經(jīng)濟(jì)和社會權(quán)力集中，加劇了極端不平等，這種不平等已經(jīng)撕裂美國社會數(shù)十年之久。如果未來真如我們常被告知的那樣屬于人工智能，那么令許多人感到不安，甚至令一些人感到憤慨的是，似乎只有極少數(shù)人掌握著如此絕對的控制權(quán)。

從某種意義上說，人工智能公司兜售的愿景極其缺乏人性：我們把越來越多的責(zé)任和判斷權(quán)交給超級智能的黑匣子，它們迅速開始塑造人類未來的走向，而這些決策對我們其他人，包括它們的設(shè)計者來說，仍然難以理解。“外行人常常驚訝和擔(dān)憂地發(fā)現(xiàn)，我們竟然不了解自己創(chuàng)造的人工智能是如何運(yùn)作的，”人智研究所的達(dá)里奧·阿莫迪去年寫道。“他們的擔(dān)憂不無道理：這種理解上的缺失在技術(shù)史上幾乎是前所未有的。”

從另一個角度來看，人工智能或許代表著有史以來最個性化的推銷方式，它強(qiáng)加給被動接受信息的美國消費(fèi)者——一種愿景，即由五家公司開發(fā)的工具，在五位特定人士（其中幾位被廣泛認(rèn)為是反社會人格者）的掌控下，幾乎完全接管美國的經(jīng)濟(jì)、社會和認(rèn)知生活。這份名單如此之短，你可能幾乎都認(rèn)識他們：薩姆、達(dá)里奧、埃隆和馬克。（谷歌旗下DeepMind的負(fù)責(zé)人德米斯·哈薩比斯或許名氣稍遜。）

這些人如今都已是億萬富翁，或接近這個級別，而且照此發(fā)展下去，隨著反精英主義情緒在他們周圍日益高漲，他們的財富和影響力似乎還將呈指數(shù)級增長。或許正因如此，去年皮尤研究中心的一項調(diào)查顯示，50%的美國人對人工智能的未來感到擔(dān)憂而非興奮，只有10%的人表示興奮。這是一個巨大的鴻溝，而整個社會正被推向深淵。

圖像

2026年，人工智能的討論如同人工智能的能力一樣，幾乎每周都在飛速發(fā)展。但或許我讀過的關(guān)于未來發(fā)展趨勢最令人印象深刻的文章，是科幻作家特德·姜（Ted Chiang）2017年發(fā)表在BuzzFeed News上的一篇文章。當(dāng)時OpenAI成立才兩年；埃隆·馬斯克和阿莫迪都還沒有獨(dú)立創(chuàng)業(yè)；馬克·扎克伯格距離他瘋狂投資人工智能的時代還有近十年。但像馬斯克這樣的末日預(yù)言家已經(jīng)開始警告美國州長協(xié)會，“人工智能對人類文明的生存構(gòu)成根本性威脅”——他的意思是，擁有超強(qiáng)能力的人工智能可能會認(rèn)為存在的意義在于制造回形針或采摘草莓，從而使其他一切，包括人類自身，都變得無關(guān)緊要。

姜寫道：“大多數(shù)人覺得這種設(shè)想荒謬至極，但令人驚訝的是，很多技術(shù)專家認(rèn)為這揭示了一個真正的危險。硅谷試圖構(gòu)想超級智能，結(jié)果卻造就了肆無忌憚的資本主義。”

如今，美國正深陷一場臭名昭著的生活成本危機(jī)，而造成這場危機(jī)的主要原因是住房短缺，缺口可能高達(dá)1000萬套。去年，美國在人工智能基礎(chǔ)設(shè)施建設(shè)上的投入超過了獨(dú)棟住宅的建設(shè)投入。我們建造的數(shù)據(jù)中心數(shù)量是排名第二的建設(shè)國（德國）的十倍。我們在人工智能領(lǐng)域的投資額是全球第二大投資國（中國）的二十多倍。總而言之，人工智能對美國經(jīng)濟(jì)而言無疑是一場巨大的豪賭。

雖然不難想象投資獲得回報的故事，但也很容易看出它與那些熟悉的人工智能寓言之間的某種相似之處，在這些寓言中，神一般的超級智能選擇優(yōu)先考慮回形針制造或草莓采摘，而不是所有其他人類活動。

然而，如今人們很少再聽到關(guān)于近期生存風(fēng)險的討論，即便這仍然是某些研究人員最為關(guān)注的問題。不久前，一半的受訪者認(rèn)為人工智能至少有10%的風(fēng)險會導(dǎo)致人類滅絕。而現(xiàn)在，人們幾乎不再談?wù)撊斯ぶ悄苡糜谥圃焐镂淦鞯娘L(fēng)險，盡管大型語言模型如今已開始為設(shè)計超級細(xì)菌提供建議——這令流行病學(xué)家們憂心忡忡。

我們已經(jīng)度過了對人工智能垃圾和生成式虛假信息的恐慌時期，盡管社交媒體上仍然充斥著這些內(nèi)容，而且關(guān)于人工智能泡沫的資產(chǎn)負(fù)債表辯論目前也已經(jīng)平息下來。

盡管人們?nèi)匀黄毡閾?dān)憂大規(guī)模失業(yè)，但目前有關(guān)失業(yè)的數(shù)據(jù)相當(dāng)模糊，而且近來經(jīng)濟(jì)學(xué)家們傾向于對大規(guī)模就業(yè)中斷的可能性發(fā)表更為令人安心的言論。人工智能領(lǐng)域的領(lǐng)導(dǎo)者們也越來越多地呼應(yīng)這一觀點(diǎn)，他們在最近幾周改變了說法，試圖淡化大規(guī)模失業(yè)的風(fēng)險。

這看起來像是企業(yè)公關(guān)，試圖平息多年來過度宣傳后引發(fā)的民粹主義反彈。但由于此次宣傳活動主要由少數(shù)幾位熟悉的面孔主導(dǎo)，他們正向我們保證未來工作、未來戰(zhàn)爭，乃至醫(yī)療、陪伴和編程的未來都將一帆風(fēng)順，因此很難不讓人產(chǎn)生這樣的印象：這些人現(xiàn)在基本上掌控著一切。在最近由帕蘭提爾基金會在耶魯大學(xué)舉辦的一次會議上，曾參與特朗普政府最初人工智能政策制定的政策專家迪恩·鮑爾做出了一個令人不寒而栗的預(yù)言，他將人工智能描述為“一個巨大的酸池”，它將溶解大多數(shù)美國人眼中的“社會”等中介機(jī)構(gòu)。“人工智能不會在政府中發(fā)揮作用，”鮑爾預(yù)測道，“人工智能將取代政府。” 去年一項針對30個國家的調(diào)查發(fā)現(xiàn)，美國人對人工智能的擔(dān)憂程度位居前列，而且沒有哪個國家比美國人更不信任政府對人工智能的監(jiān)管能力。

本周，白宮發(fā)出信號，其人工智能政策可能出現(xiàn)戲劇性的180度大轉(zhuǎn)彎——此前，政府傾向于對行業(yè)發(fā)展采取放任自流的態(tài)度，而現(xiàn)在卻提出一項提案，要求所有新的專有模型在發(fā)布前都必須經(jīng)過聯(lián)邦審查。與此同時，美國民眾也在各自力所能及的范圍內(nèi)劃清界限。根據(jù)Heatmap民調(diào)顯示，2025年9月，美國民眾對在其社區(qū)建設(shè)新的數(shù)據(jù)中心似乎態(tài)度較為矛盾，支持新建數(shù)據(jù)中心的選民比反對者多2個百分點(diǎn)。然而，僅僅四個月后，到了2026年2月，反對者的比例就飆升了24個百分點(diǎn)。這無疑是民意的一次驚人轉(zhuǎn)變。

北弗吉尼亞是數(shù)據(jù)中心飛速發(fā)展的重災(zāi)區(qū)，在2023年至2025年間，當(dāng)?shù)剡x民投票反對在本地建設(shè)數(shù)據(jù)中心，反對票比反對票高出69個百分點(diǎn)——從45個百分點(diǎn)的贊成票變?yōu)?4個百分點(diǎn)的反對票。更令人驚訝的是，在真正的數(shù)據(jù)中心中心——勞登縣，數(shù)據(jù)中心預(yù)計將在2027年貢獻(xiàn)近一半的本地稅收——該縣預(yù)計當(dāng)年29億美元的稅收中，有13億美元來自數(shù)據(jù)中心。縣政委員克里斯汀·烏姆施塔特(Kristen Umstattd)最近告訴《城市雜志》(City Journal)的格洛克法官(Judge Glock)：“多年來，評論家們一直感嘆美國不再進(jìn)行建設(shè)了。” “然而，在我們眼皮底下，美國歷史上最偉大的建筑熱潮之一已經(jīng)展開”——這是一場投機(jī)性的基礎(chǔ)設(shè)施擴(kuò)張，堪比 20 世紀(jì) 60 年代和 70 年代的州際公路建設(shè)熱潮，盡管它不像 19 世紀(jì)定義美國的鐵路熱潮那樣令人興奮和魯莽，后者造就了我們熟悉的、充斥著我們對第一個鍍金時代卡通記憶的強(qiáng)盜大亨。

昆尼皮亞克大學(xué)最近的一項民意調(diào)查顯示，唯一對這項技術(shù)在日常生活中的作用持樂觀態(tài)度的收入群體是年收入超過 20 萬美元的人群，這或許并不令人驚訝。

圖像

過去幾年，人工智能的發(fā)展就像一場驚心動魄的競賽。或者更確切地說，是一場接一場的競賽：領(lǐng)先企業(yè)之間的競賽，這些企業(yè)與監(jiān)管機(jī)構(gòu)和游說團(tuán)體之間的競賽，知識工作者與可能取代他們的機(jī)器代理之間的競賽，美國產(chǎn)業(yè)與中國產(chǎn)業(yè)之間的競賽。所有這些競賽都預(yù)設(shè)了某種終點(diǎn)線，一旦過了這個點(diǎn)，發(fā)展速度將如此之快，以至于任何優(yōu)勢——無論是某個模型、某家公司還是某個國家——都會隨著時間的推移而不斷擴(kuò)大。

這一最終目標(biāo)被稱為“通用人工智能”或“超級人工智能”。業(yè)內(nèi)許多人士現(xiàn)在都在談?wù)摗斑f歸式自我改進(jìn)”的過渡階段，在這個階段，人工智能開始獨(dú)立改進(jìn)自身的源代碼。許多人工智能研究人員認(rèn)為這一階段即將到來；Anthropic公司的杰克·克拉克本周預(yù)測，完全獨(dú)立的遞歸式自我改進(jìn)可能在不到兩年的時間內(nèi)實現(xiàn)。但就在不到兩年前，舊金山灣區(qū)的風(fēng)險投資家們還在興奮地互相詢問是否能“感受到通用人工智能的到來”。

或許我們?nèi)匀辉诔@個方向前進(jìn)。與此同時，你更有可能聽到關(guān)于“擴(kuò)散”這一棘手問題的務(wù)實討論：新模型走出實驗室，走向世界，找到用戶和用途，遇到人為瓶頸和現(xiàn)實世界的障礙，需要新的策略或更精準(zhǔn)的模型來應(yīng)對或繞過這些障礙，從而影響公眾接受的速度和方式。

這是一種截然不同的愿景：人工智能或許會繼續(xù)快速發(fā)展，甚至改變我們生活的方方面面，但并非所有權(quán)力都會集中在頂尖實驗室或掌管這些實驗室的五位負(fù)責(zé)人手中。在這種觀點(diǎn)看來，世界一流模型的尖端成就遠(yuǎn)不如誰在使用人工智能以及用于什么目的重要。

今年四月，Anthropic公司高調(diào)宣布拒絕發(fā)布其新模型Claude Mythos。該公司聲稱，該模型能夠發(fā)現(xiàn)并利用所有測試軟件中的安全漏洞，包括全球關(guān)鍵IT基礎(chǔ)設(shè)施中使用的軟件。目前尚不清楚Mythos在這些基準(zhǔn)測試中究竟領(lǐng)先多少，但它似乎促使白宮的態(tài)度發(fā)生了180度大轉(zhuǎn)彎。然而，六個月后，Mythos的開源版本勢必會出現(xiàn)，或許性能略遜一籌，但開發(fā)成本要低得多，屆時全球更多用戶將能夠訪問并根據(jù)自身需求進(jìn)行定制。或許在這樣的競爭中，最先進(jìn)的模型最終會勝出，領(lǐng)先的實驗室將始終保持領(lǐng)先于那些不知名的后起之秀，從而保護(hù)自身利益。但如果這是一場競賽，它并沒有明顯的終點(diǎn)線，也并非贏家通吃。政治學(xué)家杰弗里·丁將其稱為一場“擴(kuò)散馬拉松”。

這就是人工智能領(lǐng)域人士有時將人工智能描述為“通用技術(shù)”時的含義，就像蒸汽機(jī)、電力，或者更近期的計算機(jī)和互聯(lián)網(wǎng)一樣。一些發(fā)明家和企業(yè)家開發(fā)并完善了這些技術(shù)，在當(dāng)時積累了巨額財富，顛覆了他們所繼承的世界，并最終為我們創(chuàng)造了如今我們所生活的世界。但他們中沒有一個人能夠長期絕對掌控這些技術(shù)，更遑論掌控他們所開啟的長遠(yuǎn)未來。我們?nèi)匀挥浀媚切皬?qiáng)盜大亨”的名字，并且在某種程度上仍然生活在他們的陰影之下。但我們并非他們的奴隸。我們真的確定人工智能會有所不同嗎？

A.I. Populism Is Here. And No One Is Ready.

Silicon Valley oligarchs worried about the risks their technology posed to the world. They forgot about people.

Listen · 15:47 min

• Share full article

• 322

Credit...Photo illustration by Chantal Jahchan

By David Wallace-Wells

May 8, 2026

“I prep for survival,” OpenAI’s Sam Altman confessed back in 2016. “I have guns, gold potassium iodide, antibiotics, batteries, water, gas masks from the Israeli Defense Force and a big patch of land in Big Sur I can fly to.”

For more than a decade now, Altman and his fellow founders have lived in a state of controlled anxiety about artificial intelligence. This is, in fact, the origin story for three of the five major players in the A.I. arms race, each of which was formed out of panic that the other players weren’t taking existential fears about the technology seriously enough. They seem to have worried less about the risk of political backlash from actual humans, assuming that it wouldn’t materialize in time, would be quickly outmaneuvered by machine intelligence or could be bought off, perhaps, by talk of basic-income payments or thin promises of curing cancer.

But last month, when a Molotov cocktail was thrown at Altman’s San Francisco property, the human backlash landed literally on his doorstep. A few days later, Altman’s home endured another attack, this time by gunfire. It was hard not to think of the killing of the UnitedHealthcare chief executive Brian Thompson, which Luigi Mangione is charged with. The writer Jasmine Sun called this “A.I. populism’s warning shots.”

Americans still worry about the local impacts of data centers, storming to town halls en masse to protest them. They still worry about job loss and economic turmoil too, as do a growing number of politicians with their fingers lifted to the wind. But to many, the biggest A.I. labs now loom like the new faces of American oligarchy, as well — a fearsome concentration of economic and social power producing a self-compounding pattern of extreme inequality of the kind that has lacerated American life for decades. If the future lies with A.I., as we are so often told, it is unsettling to many and outrageous to some that so few people seem to stand in such absolute control of it.

In one sense, the vision peddled by A.I. companies is remarkably depersonalized: We hand more and more responsibility and judgment off to superintelligent black boxes, which rapidly begin shaping the course of the human future with decisions that remain illegible to the rest of us, including their designers. “People outside the field are often surprised and alarmed to learn that we do not understand how our own A.I. creations work,” Anthropic’s Dario Amodei wrote last year. “They are right to be concerned: This lack of understanding is essentially unprecedented in the history of technology.”

In another sense, and in the meantime, A.I. represents perhaps the most personalized sales pitch ever foisted on the passive American consumer — a vision of a near-total takeover of the country’s economic, social and cognitive lives by tools engineered by just five companies, run by five particular people, several of whom are widely described as sociopaths. The list is so short that you may know most of them by first name: Sam, Dario, Elon and Mark. (Demis Hassabis, who runs Google’s DeepMind, is perhaps less famous.)

These men are all already billionaires, or close to it, and on their current trajectories their wealth and influence look set to expand exponentially as, around them, anti-elitism multiplies, too. Perhaps this is one reason 50 percent of Americans told the Pew Research Center last year they were more concerned than excited about what’s to come from A.I. Only 10 percent said they were more excited. That is a yawning gap into which an entire society is being asked to tumble.

Image

The San Francisco home of Sam Altman, OpenAI’s chief executive, was attacked twice in April.Credit...Lea Suzuki/San Francisco Chronicle, via Associated Press

In 2026, A.I. discourse, like A.I. capability, jumps forward almost by the week. But perhaps the most memorable thing I’ve read about the shape of what’s to come remains an essay published back in 2017, by the science-fiction writer Ted Chiang, on BuzzFeed News. OpenAI had been founded just two years before; neither Elon Musk nor Amodei had yet split off on their own, and Mark Zuckerberg was almost a decade from his desperate A.I. spending spree. But already apocalyptic evangelists like Musk were warning the National Governors Association that “A.I. is a fundamental risk to the existence of human civilization” — by which he meant the possibility that superpowered A.I. might decide that the purpose of existence was the manufacturing of paper clips or the harvesting of strawberries, making everything else, including humans, irrelevant.

“This scenario sounds absurd to most people, yet there are a surprising number of technologists who think it illustrates a real danger,” Chiang wrote. “When Silicon Valley tries to imagine superintelligence, what it comes up with is no-holds-barred capitalism.”

Today, the United States is in the middle of a notorious cost-of-living crisis fueled in large part by a housing shortage of perhaps 10 million units, and last year, the country spent more money building A.I. infrastructure than single-family homes. We built 10 times as many data centers as the next biggest builder (Germany). We invested more than 20 times as much money into A.I. as the world’s next biggest investor (China). Among other things, artificial intelligence is an enormously big bet for the American economy to have made.

And though it isn’t all that hard to imagine a story in which the investment pays off, it’s also not hard to see a kind of parallel with those familiar A.I. parables, in which a godlike superintelligence chooses to prioritize paper-clip manufacturing or strawberry picking over all other human endeavors.

Yet you hear considerably less about near-term existential risk these days, even if it remains a pre-eminent concern among some researchers. Not that long ago, half of those surveyed said that there was at least a 10 percent risk that artificial intelligence would bring about human extinction. And you no longer hear much at all about the risk posed by A.I. helping to create biological weapons though large language models are now routinely dispensing advice for designing superbugs — much to the alarm of epidemiologists.

We’ve passed through a panic about A.I. slop and generative disinformation, though social media is inarguably awash in them still, and balance-sheet debates about the A.I. bubble have subsided for now, too.

And while there is still widespread worry about mass unemployment, for now the data on job loss is pretty ambiguous, and these days, economists tend to strike more reassuring notes about the possibility of large-scale job disruption. Increasingly, that line is echoed by A.I. leaders themselves, who in recent weeks have executed a rhetorical about-face to downplay the risk of mass unemployment.

This may look like corporate P.R., an effort to tamp down populist backlash after years of hyping up investors. But because the outreach features the small set of familiar faces now reassuring us about the future of work and the future of war, not to mention the futures of medicine and companionship and coding, it’s hard to escape the impression that these same people are basically in charge of everything now. At a recent conference staged by the Palantir Foundation at Yale, Dean Ball, the policy wonk who was an architect of the Trump administration’s original A.I. policy, offered a chilling prophecy, describing A.I. as “this giant acid vat” which would dissolve the mediating institutions most Americans see as “society.” “It will not be A.I. in government,” Ball predicted. “It’s going to be A.I. as governments.” A survey of people in 30 countries last year found that Americans were among the most nervous about A.I., and that nobody trusted their government to regulate A.I. less than we did.

This week, the White House signaled that it may make a sudden and dramatic U-turn on A.I. policy — once inclined toward hands-off support industry growth, the administration is now floating a proposal to force federal review of all new proprietary models before release. And Americans are drawing lines in the sand where they can, too. In September 2025, Americans seemed roughly ambivalent about the construction of new data centers in their communities, according to Heatmap polling, with voters 2 points more likely to support new construction than to oppose it. Four months later, in February of 2026, they were 24 points more likely to oppose it. That is a shockingly large swing in public opinion.

Northern Virginia is ground zero for the breakneck build-out of data centers, and between 2023 and 2025, voters there swung 69 points against building them in their own communities — from 45 points in favor to 24 points against. This is all the more remarkable given that in Loudon County, the true center of activity, data centers are expected to generate nearly half of all local tax revenue in 2027 — $1.3 billion of the $2.9 billion the county expects to bring in that year, County Supervisor Kristen Umstattd recently told the magazine City Journal’s Judge Glock. “For years, pundits have lamented that America no longer builds things,” Glock observed. “Yet under our noses, one of the great building booms in American history has unfolded” — a speculative infrastructure expansion comparable to the interstate-highway push of the 1960s and 1970s, if not quite as heady and heedless as the railroad boom that defined 19th-century America and that gave us the familiar robber barons that populate our cartoon memories of the first Gilded Age.

Perhaps it should not be surprising that according to a recent Quinnipiac poll, the only income bracket with hopeful views of the technology for their day-to-day lives was those making over $200,000 per year.

Image

Former Gov. Brian Sandoval of Nevada, left, with Elon Musk at a National Governors Association meeting in 2017.Credit...Stephan Savoia/Associated Press

For the last few years, A.I. has felt like a breakneck race. Or, perhaps, one race after another: races between the leading companies, between those companies and regulators and lobbyists, between knowledge workers and the machine agents which may replace them, between the American industry and the Chinese one. All in their way presume a certain kind of finish line, a point past which progress proceeds so rapidly that any advantage — by a model, a company, a country — will be extended over time.

This end point has been called “artificial general intelligence” or “artificial superintelligence.” Many in the industry now talk about an interim phase of “recursive self-improvement,” in which A.I. starts to independently improve its own source code. Plenty of A.I. researchers believe it is right around the corner; Anthropic’s Jack Clark predicted this week that fully independent recursive self-improvement might be less than two years away. But less than two years ago, Bay Area venture capitalists were already giddily asking one another about whether they could “feel the A.G.I.”

And maybe we’re still on track for that. In the meantime, you’re more likely to hear pragmatic conversations about the thorny problem of what is called “diffusion”: the speed and shape of public uptake as new models spread out into the world beyond the lab, finding users and uses, hitting human bottlenecks and real-world obstacles and requiring new strategies or more narrowly trained models to navigate through or around them.

This is a pretty different vision, in which A.I. may continue to rapidly progress, even transform much of our lives, but without all the power necessarily accruing to the leading labs or the five individuals in charge of them. In this view, the state-of-the-art achievements of world-class models matter less than who is using A.I. and for what.

In April, to great fanfare, Anthropic refused to release a new model, Claude Mythos, which the company said could find and exploit security vulnerabilities in every tested piece of software, including those used in critical pieces of global I.T. infrastructure. It’s not exactly clear how far ahead Mythos really was on these benchmarks, but it appears to have inspired the White House’s apparent about-face. Yet six months from now, there will inevitably be an open-source version of Mythos, perhaps not quite as good but much cheaper to produce, that many more users around the world will have access to — and be able to customize to their own purposes. Maybe the state-of-the-art models will win out, in competitions like this, with the leading labs staying far enough ahead of ragtag upstarts to protect themselves. But if this is a race, it doesn’t have an obvious finish line, and it doesn’t really seem winner-take-all. The political scientist Jeffrey Ding calls it a “diffusion marathon.”

This is what A.I. people mean when they sometimes describe A.I. as a “general-purpose technology,” like steam engines, electricity or, more recently, computers and the internet. Some of the inventors and entrepreneurs who developed and refined those technologies made huge fortunes in their time, disrupting quite a lot of the world they inherited and giving us, ultimately, much of the world we ourselves live in. But none of them retained absolute control over those technologies for very long, let alone over the long-run future they unleashed. We still know the names of the robber barons, and live still somewhat in their shadows. But we are not their serfs. Are we sure A.I. will be different?