❌

普通视图

  • ✇The New Yorker
  • Is the F.A.A. Grounding the First of Many?
    2026-11-20T21:23:04.199ZSave this storySave this storyListen and subscribe: Apple | Spotify | Wherever You ListenSign up to receive our twice-weekly News & Politics newsletter.The veteran transportation-safety reporter Matthew L. Wald joins Tyler Foggatt to discuss the Federal Aviation Administration’s recent nationwide grounding order, which disrupted more than ten thousand scheduled flights, and what it revealed about the vulnerabilities of the country’s aviation system. They
     

Is the F.A.A. Grounding the First of Many?

2026年11月21日 06:06
2026-11-20T21:23:04.199Z

Listen and subscribe: Apple | Spotify | Wherever You Listen

Sign up to receive our twice-weekly News & Politics newsletter.


The veteran transportation-safety reporter Matthew L. Wald joins Tyler Foggatt to discuss the Federal Aviation Administration’s recent nationwide grounding order, which disrupted more than ten thousand scheduled flights, and what it revealed about the vulnerabilities of the country’s aviation system. They discuss what caused the ground stop, how a seemingly minor equipment upgrade could have such far-reaching consequences, and whether responsibility lies with the Trump Administration’s cuts and disruption of the F.A.A.’s workforce or with previous Administrations’ failures to modernize the aging infrastructure. They also explore the growing role of automation in air travel, including an eight-hundred-and-seventy-five-million-dollar A.I. system being introduced by the F.A.A., and whether such innovations stand to improve aviation or introduce new risks, from user error to cyberattacks. Does this incident, along with a recent series of high-profile outages affecting technology, banking, and national-security systems, point to a growing vulnerability in the infrastructure that people rely on every day?

Recommended reading:

Tune in to The Political Scene wherever you get your podcasts.

  • ✇The New Yorker
  • The Case for A.I. Optimism
    2026-11-20T19:00:00.000ZSave this storySave this storyListen and subscribe: Apple | Spotify | Wherever You ListenSign up for our daily newsletter to get the best of The New Yorker in your inbox.In Washington, it is essential to read, or at least display, the collected works of Robert Caro. (The spines of “The Power Broker” and the multivolume biography in progress of Lyndon Johnson are compulsory in every Zoom background on cable news.) In Silicon Valley, the required reading list s
     

The Case for A.I. Optimism

2026年11月21日 04:06
2026-11-20T19:00:00.000Z

Listen and subscribe: Apple | Spotify | Wherever You Listen

Sign up for our daily newsletter to get the best of The New Yorker in your inbox.


In Washington, it is essential to read, or at least display, the collected works of Robert Caro. (The spines of “The Power Broker” and the multivolume biography in progress of Lyndon Johnson are compulsory in every Zoom background on cable news.) In Silicon Valley, the required reading list surely includes Richard Rhodes’s “The Making of the Atomic Bomb,” a magisterial history of the discovery of nuclear fission and its epochal resolution with the Manhattan Project and the start of the nuclear age. Not a few of the tech barons and innovators of our time see themselves as neo-Oppenheimers. When A.I. reaches maturity, it may get its Caro or Rhodes. In the meantime, Kevin Roose’s lucid journalistic account “The AGI Chronicles: The Inside Story of the Race to Create an Artificial Superintelligence” does the trick.

Roose is a classic beat reporter, having covered technology for outlets including the Times, The Atlantic, New York, and NPR. Recently, he bolted from the Times to work independently, and, these days, he is co-hosting a podcast, with Casey Newton, called “Machine Gods.” Among the most interesting A.I. podcasters around—on “Dwarkesh Podcast,” “No Priors,” and “Latent Space,” to name a few—Roose stands out for his ability to tell a story with narrative drive. He breaks things down neatly and clearly for the non-expert.

I spoke with Roose for The New Yorker Radio Hour just before Anthropic, Dario Amodei’s A.I. outfit, made what has turned out to be the largest initial public offering of all time. And, of course, while Amodei and his colleagues were coming into their many billions, Jensen Huang, the former king of the Nvidia empire, which has by many metrics powered the A.I. revolution, was appealing to his old board of directors to let him back in after his humbling. On the utopian-to-doomer scale, Roose is . . . hopeful. Or, as they say on the Sabbath Gasbag shows, cautiously optimistic. His interlocutor, by nature, is less so. We had a lot to talk about.

Anthropic is just going public. In their filing paperwork for the first half of 2026, they reported a revenue of $16.3 billion and a net loss of $80.7 billion. The company is also on the hook for $900 billion in future obligations for chips and data-center leases. How do they make the math make sense?

Basically, the bet that investors are making on Anthropic is that it will continue to grow at something close to the rate that it has been growing for the last several years. This is a company that has been growing faster than any company at its scale in the history of capitalism. Four or five years ago, they had essentially no revenue. Three years ago, they had tens of millions in revenue, and it’s been growing tenfold every year for the last three or four years. Not ten per cent; ten times.

That is due, in large part, to the popularity of their models for coding and other white-collar work. And the bet is that this growth just continues—that as these tools get better and more helpful, that more white-collar knowledge workers will incorporate them into their work, and developers will be paying much more, because usage will continue to go up.

The math doesn’t make sense if you think that A.I. is already as popular and widely adopted as it will ever be. It starts to make more sense if you anticipate that what has been happening for the past few years will keep happening—which is that the models get better and more people use them.

Some companies out here in Silicon Valley are already spending more on A.I. tokens than they are on payroll. So the long-term bet is that the economy essentially switches over from running primarily on human labor and paying human salaries to running on A.I. labor and paying token prices.

If that in fact is the right bet to make, give me a picture of the economy on an A.I. basis, as opposed to human labor. What’s it going to look like?

I think it looks a lot like my company. I’m starting a startup, Machine Gods Media. We are a six-person company right now. In our Slack, where we communicate all day, we have six humans and probably a dozen agents. And the agents are doing things we don’t know how to do ourselves. They’re writing code. I just got back from what I jokingly call our engineering department, but it’s really just a bunch of Claudes running around. They’re building tools, helping with scheduling and coördination, with all of the pieces that go into making podcasts, aside from the actual talking into a microphone.

What would those be? Is it so complex that you need Claude to do it?

It turns out to be pretty complex. For example, we have all these deliverables. When you have a show, it needs to get delivered in all these different formats. Apple needs one format, Spotify needs one format, and YouTube needs another format. And it’s all got captions, and titles, and descriptions, and metadata. And all that used to have to be done by hand. Now we have this tool that does it all for us.

Tell me how that works for a big law firm, as opposed to a six-person podcasting outfit.

It might not be a big law firm. It might actually be a small law firm, in terms of the human payroll and footprint. It might be a law firm that only has ten people, but, using A.I., they have access to essentially unlimited amounts of paralegal, junior-associate-quality labor. And so they can do, with ten people, what a more traditional firm might need a thousand people for. That may or may not be possible—there might be gaps in what these models do that might make them unsuited to that level of integration. But that’s the bet that investors are making. It’s not just that Anthropic will continue to grow or that its models will get better, but that A.I. will become the foundation of a new kind of company that will be built on top of A.I.

How many of these companies can prevail in the new world of A.I.? We have Anthropic and OpenAI, among others. Who’s in the lead and why?

Right now there are two clear frontier labs: OpenAI and Anthropic. I would say that, of the two, Anthropic may be slightly ahead; its models are considered better by some metrics. But those two are definitely the front-runners. Then there’s a small-to-medium drop-off before Google DeepMind, which to my mind is the third frontier lab. And then there’s a pretty steep drop after that—to Meta, SpaceX, the Chinese open-model companies, and some other players.

What do I get out of using Anthropic’s Claude, as opposed to OpenAI’s ChatGPT?

I use both of them—I’m pretty promiscuous. I find that some are better for some tasks than for other tasks. I use Claude to write code and to get editorial feedback on my writing. I don’t use A.I. to write, but I do use it to help me revise and edit—

As an editor, I have to stop you there. Do you get editorial feedback that is any good, as compared to when you wrote for the New York Times, or when you wrote your book?

I’ve been blessed and fortunate to have some of the most amazing human editors in my career, including at the Times and at F.S.G., where I published my book. So I hope they’re plugging their ears when they hear this next part: A.I. has gotten really good at editing.

You’re killing me here, Kevin.

I know. Look, I’m not here to tell reassuring fictions. I’m here to tell uncomfortable truths. I had a council of Claudes that would look at my chapters before I submitted them to F.S.G., and give me feedback from various perspectives. I would assign one to be the dyed-in-the-wool A.I. skeptic, and one to be an A.I. historian that knew all of the technical details about the history of these labs. And I would get the feedback in the form of these long dossiers. Maybe ninety per cent of it was crap, but ten per cent was great. My council of Claudes actually caught a few mistakes that would’ve otherwise made it into—

Whoa, whoa, whoa. Ninety per cent of the comments were crap, and you think that’s good editorial counsel?

Well, how much of what your editors give you is useful?

One hundred per cent!

[Laughs.] Well, maybe it’s just my vanity that makes me think it’s only ten per cent useful. But I do feel like, if you’re willing to wade through a bunch of subpar feedback, you can actually find good stuff in there.

The P.R. that I get is that Anthropic is the good company, as you suggest. The job interviews even have a moral dimension. Whereas OpenAI is quite a different thing. Is that fair?

That’s the image that has been portrayed.

What’s the reality?

The reality is that Anthropic and OpenAI are locked in this race together. I think we have been focussing on the wrong question with respect to these companies. I’ve spent time with Sam Altman, with Dario Amodei, and with many leaders and researchers at these companies. We in the tech press are used to asking the question of “Are the people running these companies good and trustworthy?” To my mind, that’s less useful than looking at the technology itself.

A good, moral, ethical person could still build an A.I. model that goes rogue and kills millions of people. On the other side of the coin, someone who is totally craven and corrupt could manage to produce an incredibly ethical A.I. model with all of the proper guardrails. I don’t know that it makes sense to fit the image of the founder or the C.E.O. onto the over-all impression we have of these companies.

In your book, which is fascinating, you write that a surprising number of people who work in these A.I. companies joined because they were . . . terrified. They decided that, if A.I. was going to be built, it would be better for it to be built by people who were scared, who were alarmed, who were cautious, so that the appropriate measures against disaster would be taken. Is there any evidence that these people, who are terrified, are having a significant impact within the companies—to hold them back, restrain them, to modify?

Absolutely. It’s almost impossible to imagine a counterfactual, because the companies were started by people who were terrified of A.I. Sam Altman and Elon Musk started OpenAI, in 2015, because they were scared that Demis Hassabis and Google DeepMind were going to win the race to A.G.I., artificial general intelligence. And then, several years later, Dario Amodei and six of his co-workers at OpenAI left to start Anthropic because they were terrified of the technology, and they didn’t trust Sam Altman to build it safely.

This is a recurring motif in the modern A.I. industry: people who are scared of A.I., and who don’t believe their bosses or rivals are taking enough steps to contain it, break off and start their own companies. They say, “We’re going to be the safe A.G.I. company. Those other guys are going to blow up the world if they do it, but we can be trusted.”

Some of this, it seems to me, has to do with personality, and how that personality is projected to the wider world through the press. Sam Altman projects one kind of personality, and Dario Amodei perhaps another. Tell me a little about those differences and what effect they have.

Sam Altman is a master salesman and fund-raiser. That is his superpower. From a very young age, he was a networker; he was close to major dealmakers in Silicon Valley, and extremely talented at raising money. He ran Y Combinator, the startup incubator, before leaving to become the C.E.O. of OpenAI. His reputation is that of an extremely ambitious, high-E.Q. people pleaser. As someone once told me, his superpower is being able to zero in on exactly what you want to hear and give it to you. That has served him well in some instances and poorly in others, such as when his board decided that he was lying to them and tried to fire him. It has also contributed to his reputation as somewhat two-faced or duplicitous.

What about Dario?

Dario is very high-I.Q., stubborn, and neurotic. He’s a moral absolutist, with a strong belief in what is good and right and righteous. He’s also a brilliant scientist who foresaw many of today’s A.I. trends and dynamics years before anyone else. His superpower—and his kryptonite—is that he cannot lie. Someone close to him once told me, “Dario is a guy who cannot tell a lie to save his life, which is an asset for us on even-numbered days and a liability on odd-numbered days.”

How does that work?

Sometimes it leads to him standing his ground and standing up for his principles. When the Pentagon says, “We want to use Claude for whatever we want to do,” he’s the guy who has the cojones to stand up to them, while everyone else is busy trying to snivel and get on their good side.

The Pentagon banned the use of Claude.

Yes. And a lot of people at Anthropic saw that and admired Dario for standing his ground, even when it would have been expedient for him to say whatever it took to make Pete Hegseth happy. But it also leads to these incredible moments. Remember that press briefing outside the White House, where all these A.I. C.E.O.s were standing in a semicircle? Dario was practically jumping out of his skin, unable to conceal his displeasure at having to stand alongside these cowards. He couldn’t hide it. It was written all over his face. So, I think that’s what they meant by an asset on even-numbered days and a liability on odd-numbered days.

What has stood out most for you in Anthropic’s I.P.O. and the way this has all gone?

I think about the absolute rocket-ship trajectory this company has been on, more than any other company I’ve ever covered, and perhaps more than any company that has ever existed.

The first time I visited Anthropic to report on them was three years ago. Three years—not that long ago. They had a hundred and fifty employees, working out of a cramped office in Jackson Square, in San Francisco. They didn’t even have a sign-in desk. You just took the elevator up and started walking around. They had no plans to make money. They weren’t a commercial enterprise and had barely released a product. They were a bunch of A.I.-safety-obsessed nerds who seemed genuinely worried about the technology.

If you had told me on that first visit that, three years later, this company would go public at a valuation of more than a trillion dollars and become one of the largest companies in the world, I would have laughed in your face. It seemed so unlikely that they would go from a little hundred-and-fifty-person startup to a company with, I think, more than five thousand employees, and one of the biggest in the world.

There’s perhaps a too easy confidence in the world of long-form journalism that what we do is special—that gathering information, analyzing it, and presenting it in narrative form is something A.I. simply won’t be able to do. Whereas asking for the ten best restaurants in Seattle is easy enough: you can scrape that information off the internet and present it. Are we living in a dream?

I think we are. I think there will still be a market for human-authored work—there’s just something about A.I. writing that is hollow and generic. But what’s interesting is that researchers have conducted blind tests in which people are shown two snippets of writing and asked which they prefer. Most of the time, when they don’t know which is which, people prefer the A.I.-generated work.

Really?

But when people learn which one was A.I. and which one was human-written, the preference swaps.

So, what does that tell you?

Almost everyone likes the human-written one.

They like the idea of the human-written one, but they might not innately like the human one at all.

I think the future of at least some writing will involve humans sticking their names on A.I.-generated literature because they don’t want to break that connection, that spell, with the audience.

That is not entirely reassuring, Kevin.

I’m sorry. I wish I had a rosier view.

On the continuum from doomers, people who believe A.I. will kill us all, to those who are absolutely sanguine about this whole thing—who think it’s going to be great and we’re going to work one day a week and spend the rest of our time sunning ourselves and, I don’t know, whistling—you are pretty sanguine about this subject, and the opposite of a doomer. Am I right?

I am an optimist by temperament. I believe that we can and should thread this needle, but we’ve got a lot of work to do.

What concerns you the most?

Right now, what concerns me is an intelligence explosion. That’s a term in the A.I. industry for what could happen after A.I. becomes capable of building A.I. It’s sometimes called “recursive self-improvement.” It’s the idea that an A.I. model becomes superhuman at developing other A.I. models, capable of doing the entire workload of a senior research engineer at OpenAI or Anthropic. That hasn’t happened yet, but some people, including at METR, a credible third-party organization, believe it could happen within the next year. At that point, you could use A.I., in a closed, no-human environment, to train more capable A.I. models.

What are the potential disasters, in concrete terms, that could come out of that manifestation?

If these people are right—and let’s caveat that they might not be—but if we were to see, say, a decade’s worth of improvements in A.I. capabilities compressed into a single year, because all of this could happen at machine speed rather than at human speed, I think we’d be looking at systems that start breaking a lot of things all at once. We’re already seeing some of this with cyberattacks, the incident involving Hugging Face, and the U.S. government scrambling to contain threats to critical infrastructure from these A.I. models. Imagine that, but times ten or times a hundred. There’s just a base rate of chaos and instability—

And, at the same time, a very low propensity toward regulation.

For now. But I think that could change very quickly. In fact, I believe part of the reason the Democrats won the midterms so decisively was—

Anxiety about A.I.

People had anxiety about A.I. So things could change pretty quickly on this. As unpopular as A.I. might be right now, I expect it to be even more unpopular. Right now we’re seeing this massive backlash against A.I. Unemployment is still under five per cent. It was not at all intuitive to me that we would see this level of backlash before any of the widespread job losses arrived.

Well, because you’re reading about it every two seconds, and you’re seeing reports on it on television and YouTube. There’s an anticipatory anxiety about it.

But think about how much more anxiety there would be if people were actually losing their jobs en masse. I desperately want that not to happen, but, to me, the timing of this backlash has been a little surprising, because it seems like there’s still a lot more risk out there to be priced in.

Now, I think you have to be impressed that we’ve gotten this far in our conversation without having mentioned Jensen Huang. What do you make of this? What did we just see? My God, it’s like the worst James Bond plot you have ever experienced.

It really is. It’s dime-store stuff. This is a guy who has been so central to the A.I. boom. Nvidia’s G.P.U.s are the reason we have any of this today.

Tell us what a G.P.U. is.

G.P.U. is a graphics-processing unit. Basically, a modern A.I. chip is descended from these things that gamers used to stick in their PCs to render backgrounds and polygons better. And Jensen Huang is the reason that those chips are now the most valuable commodity in the world—so valued that we’ve had to pass special export controls to prevent China from getting access to them, because, if compute is the determining factor between success and failure, we don’t want them to get our most powerful chips.

Do you figure they have it anyway?

It sure seems like they’ve gotten a bunch of them anyway. And the question is—

Thanks to Alice Huang?

Well, there’s this whole gray market of cutouts in Malaysia, Singapore, and other countries. They’ll buy a bunch of Nvidia chips because they’re allowed to buy them—they’re not China—and then they’ll either set up a data center and give Chinese companies remote access to it, or, in some cases, they’ll just smuggle the chips into China. And one of the major questions going forward is how much of this Jensen Huang knew about. Was he aware that Nvidia chips were being smuggled into China? And how much did his relationship with Alice Huang and Megaspeed factor into that?

So, what are the ramifications of this?

I mean, we’ll see where Jensen Huang ends up. He’s now saying he would like to return to Nvidia. I suspect that won’t work.

I’d like to win the Super Bowl, too, but I don’t think that’s happening.

In the short term, it means that the entire industry will have to approach things slightly differently. For many years, Jensen Huang has been the sole guy with control of the commodity that everyone in A.I. needs—the key ingredient. His relationships really determined who got how many of Nvidia’s latest chips. It was not a pure auction market. It was, like, “Oh, Alice Huang and I are friends, so I’ll save some of these chips for Megaspeed.” Or, “Elon Musk and I go way back. I’ll give him a big allotment.”

Who benefitted most from the relationship?

Elon Musk has gotten a very large allotment of chips that was maybe out of step with his company’s importance in A.I. because he and Jensen Huang are old friends. I think Sam Altman has benefitted from a close relationship with Jensen Huang.

Dario Amodei and Anthropic have actually not had a close relationship with Jensen Huang. As I reported in my book, there was a very tense meeting between Dario Amodei and Jensen Huang—it was the first time they’d met, and Jensen was basically grandstanding about how big the data center they were going to build was, while Dario was peppering him with somewhat rude questions. “Oh, how big is it going to be? What order of magnitude? Where are you going to get the power?” And eventually Dario mutters to someone sitting next to him that Jensen is acting Trump-like, and the two of them didn’t like each other for a long time.

If Dario didn’t have that kind of access for Anthropic, how has he prevailed anyway?

They have adopted a very unusual compute strategy—they have essentially gotten a little bit from everyone. They have a bunch of Amazon chips called Trainium. They rent a bunch of chips from Google, and they have a bunch of Nvidia chips, too. Rather than tethering themselves to one chip company or one hyperscaler, they have drawn from all of them to varying degrees. But they’re compute-constrained just like everyone else in the industry. They wish they had more data centers. They wish they had more chips. No one’s got enough.

I think it’s fair to say that when the internet and the iPhone first appeared we had no real concept of what an unregulated experiment on human consciousness and behavior these things would represent, on all kinds of levels—whether it was about reading, concentration, mental health, or how we related to each other. Here we are again, on the brink of another revolution that is arguably much greater—something vastly more complicated. Despite your optimism, in terms of human relations and the way we live our lives, what are you expecting in the next decade or so?

A decade is so long from now, David.

That’s fair.

I struggle to think about what the world will look like in a year.

Really?

I do. I think parts of it will look largely the same. I will still probably drop my kid off at school, go to the grocery store. There will be parts of my life that are much the same.

But when you and I talk next Thanksgiving, what will we have experienced? What changes will have been brought into our lives over the course of a year?

If the next year plays out in a way that is thoughtful, I think we will have put in place vastly more guardrails.

Where do you get your confidence that thoughtfulness is anywhere in abundance in the U.S. government, or anywhere else?

Thoughtful, or maybe just self-interested. I think we have a deep desire to remain in control of our society, of our government, of our surroundings, of our lives. We still want to be in the driver’s seat. And as long as we recognize that A.I. is not just a tool or a hype bubble but is actually, in some sense, a type of entity that could displace us at the top of the intelligence ladder . . . We could be the second-smartest thing on earth a year from now, and I think we have a lot to do between now and then. But my hope would be that we look back and say, “Wow, that was a crazy year, and a bunch of cybersecurity incidents happened, and we fixed a bunch of stuff just in time, and it was like Y2K. It could have been this major crisis, but we averted it through heroic effort and late nights at the computer terminal.” That, I think, is the good outcome.

That’s a lot of close calls to endure and still be an optimist.

We’ve done it before.

When?

Y2K, the Cold War. The Industrial Revolution played out over seventy-five or eighty years, but we muddled through. It did not lead to human extinction.

A lot of people talk about a kill switch for A.I., as if, when things get out of hand, suddenly somebody goes to a special room and turns a switch and A.I. is turned off and things can settle down, and whatever is broken can be fixed. What is a kill switch? Is it an imaginary construct or is it a real thing?

It’s an imaginary construct. I think it comes from science-fiction movies or thrillers where—

That’s where I get all my best stuff.

The red lights in the control room start blinking, and the engineer sprints over to the server rack and yanks out the cord—

Clips the blue wire instead of the red wire. Bruce Willis always chooses correctly.

Unfortunately, that is not how modern A.I. systems are running. They are distributed across huge numbers of servers and data centers. And there’s been some research into how these things can actually replicate themselves on smart devices, like smart ovens and toasters in your home. We can’t just pull the plug if it might mean taking down some huge chunk of computers across the world.

In a few years, are we going to be at twenty-five-per-cent unemployment?

I hope not. I hope we will continue to find useful and valuable things to do, but I think a lot of that depends on the decisions we make not just as a public sector, as the government, as regulators, but as employers and business leaders. I mean, do you fire your human workers as soon as they’re no longer technically adding anything to the organization, or do you retrain them for some other part of the job? I hope that we have enlightened leaders out there who will see it as part of their responsibility not to conduct layoffs as soon as it’s technically feasible.

Where does your sense of confidence, whether in the government or the private sector, reside?

My confidence resides in the public, actually. As recipients, users, and, in some cases, victims of this technology, we are not powerless or impotent. We have seen large, society-altering technological transformations turn out great, because people made it turn out great. My optimism also comes from the general belief that technology usually makes our lives better in the long run. ♦

The New Yorker Radio Hour is a co-production of WNYC Studios and The New Yorker.

  • ✇The New Yorker
  • Is Science Accelerating?
    2026-11-20T11:00:00.000ZSave this storySave this storyYou’re reading Open Questions, Joshua Rothman’s weekly column exploring what it means to be human.What was American life like a hundred years ago? The economist Robert J. Gordon offers a vivid picture in his book “The Rise and Fall of American Growth,” from 2016. In 1926, Gordon writes, there were almost no paved roads west of the Mississippi. Fewer than one in ten homes had an electric refrigerator. (A typical one co
     

Is Science Accelerating?

2026年11月20日 20:06
2026-11-20T11:00:00.000Z

What was American life like a hundred years ago? The economist Robert J. Gordon offers a vivid picture in his book “The Rise and Fall of American Growth,” from 2016. In 1926, Gordon writes, there were almost no paved roads west of the Mississippi. Fewer than one in ten homes had an electric refrigerator. (A typical one cost the equivalent of ten thousand dollars today.) If you lived in Muncie, Indiana, the odds were one in four that your house had no running water, and one in three that it had no sewer connection. Life was rough by our standards. And yet, Gordon shows, scientific and technological progress was obvious everywhere. Thousands of new hospitals were being constructed, many with cutting-edge X-ray departments. Ninety-three per cent of Iowa farmers owned automobiles, even though the automobile itself had begun to appear only three decades before. 1926 was the year it became possible for an ordinary person to travel on an airplane; it was also when an inventor named Clarence Birdseye perfected frozen food, making the consumption of fish affordable in landlocked places. The way people lived was clearly better than how their parents had lived.

This rapid “epochal transformation,” Gordon writes, actually continued for a whole century, stretching from 1870 to 1970. It was powered by five “Great Inventions”—electricity, urban sanitation, chemicals and pharmaceuticals, internal-combustion engines, and signal transmission. The Great Inventions, in turn, were based on foundational discoveries in physics, biology, medicine, chemistry, math, and other fields. “What makes the period 1870–1970 so special is that these inventions cannot be repeated,” Gordon writes. The germ theory of disease, which Louis Pasteur advanced in the eighteen-eighties, led to sweeping changes in pretty much every aspect of life, and yet it could be discovered only once. The principle of electromagnetism unlocks everything from engines to radios, but there aren’t infinite, similarly powerful principles to uncover and apply. The staircase of science—or at least this flight of it—appears to have only so many steps.

What happens after the biggest discoveries have been made, explored, and put to use? Arguably, the rate of progress slows, and dramatic transformation becomes harder to achieve. The practice of science itself also changes. In recent decades, economists and other researchers have explored a field that they call “metascience”—the study of why the pace of scientific discovery increases or decreases. They point out that, although more scientific papers are being published, fewer are truly disruptive; more patents are granted, but a smaller number are for “breakthrough” inventions. Increasingly, science is conducted inside enormous bureaucracies that favor cautious, incremental research. Grants are awarded mainly to sure bets; power concentrates among august institutions and scientists. In a 2020 paper, the economists Nicholas Bloom, Charles I. Jones, John Van Reenen, and Michael Webb asked, “Are ideas getting harder to find?” They observed that the number of researchers required to double the speed of computer chips was eighteen times larger than it had been in the early nineteen-seventies. For a long time, they concluded, progress had emerged from “a constant flow of new ideas.” Now that flow has slowed.

Is this how life feels in 2026? Does it seem to you that scientific and technological progress has plateaued? It’s true that the iPhone Duo is only somewhat exciting. But, on the other hand, there is now a pancreatic-cancer pill and a melanoma vaccine. In Massachusetts, a man named Tim Andrews lived for the better part of a year while relying on a gene-edited kidney grown inside a pig. Norwegian salmon farms now use underwater A.I.-controlled laser robots to kill sea lice; in France, a startup has made tiny drones that can do the same for mosquitoes in your back yard. “Technological evolution is still accelerating,” the computer scientist John von Neumann wrote, in a 1955 essay, toward the end of Gordon’s spectacular century. In that particular essay, von Neumann was thinking mainly of nuclear power, machine automation, and climate control—technologies that have progressed slowly or fizzled out. And yet computer science, a field to which he made foundational contributions, has made steady—and now startling—new advances. The question of whether science is slowing or accelerating is complicated by its essential unpredictability. And because science is an interpersonal, interwoven enterprise, progress in one area can enable it in others, in ways that might not be easy to foresee.

A cartoon equation for science might go like this: B (brainiacs) + M (money) x T (technology) = D (discovery). Needless to say, it isn’t that simple. The brainiacs can be arranged in all sorts of organizations and hierarchies; the money can flow to the right ones or the wrong ones. The technology can be employed correctly or incorrectly; the discoveries can result in smarter brainiacs, more money, and better technology, or they can die on the vine. Fields and methods, meanwhile, can be young or grown-up, flexible or ossified, fashionable or passé. “At the time I was starting, lots and lots of things were just opening up,” Eve Marder, a neuroscientist who began her career in the nineteen-seventies, told me. “I started just after the time when someone could put an electrode anywhere in any nervous system and see something that no one had ever seen before—which is very, very different from the much more mature science that we have today.” For neuroscientists now, she went on, “what constitutes ‘success’ or ‘new knowledge’ has changed, and the incentive structures have also changed.” Marder recalled going to her first neuroscience conference, which had around five hundred attendees, allowing her to meet many of the most important figures in the field. Today, the big annual meeting put on by the Society for Neuroscience typically hosts around twenty thousand researchers.

Marder is now an eminent scientist: her lab, at Brandeis University, studies how individual neurons connect to create stable circuits that can survive and function over time. She has won major awards, supervised large numbers of scientists and graduate students, and participated in research projects of varying sizes. “We have a very complicated challenge in science today, which is that there are some things that can only be done by ‘big science,’ ” she said. “If you’re doing ‘small science,’ with small groups of people, then the ones who make mistakes can be very easily corrected or can self-correct. But, if you do big science poorly, somebody can make a bad decision and take everyone down a bad path.”

In a big-science world, researchers often work in large consortia that cross institutional and disciplinary lines. They collaborate to gather vast quantities of data, hoping that it will be amenable to analysis now and in the future. Young scientists can find themselves assimilated, Borg-style, into these expansive projects and the institutions that make them possible. “A student who goes into a field is captivated by ideas,” Marder said. “They want to discover something, to create something, just like a philosopher wants to think something no one’s ever thought. And then they are socialized by having to get to the next step, or get the next fellowship. They start worrying about who they have to please—and that can flow down and change the science that they’re willing or able to do.” Some young scientists “are very stubborn, and will continue to be independent and innovative, and become leaders in their fields,” Marder went on. But “a lot of very creative people get driven out.”

If the scale of modern science can dwarf individual scientists, it can also sideline their contributions. “Right now, on my desk in my office, I have a huge pile of really important papers by brilliant people that I don’t have time to read—there’s too many of them,” Michael Levin, a biologist at Tufts, told me. “I conclude, by symmetry, that nobody has time to read my papers, either. That raises the question: Who the hell am I writing all this for?” Levin, who studies bioelectric computation—the process through which cells communicate and make collective decisions—believes that science is, in fact, accelerating: one sign is that “it’s becoming more and more difficult to stay apprised of the important things.” In a 2005 paper titled “The Burden of Knowledge and the ‘Death of the Renaissance Man,’ ” the economist Benjamin F. Jones explored this paradox. Since “knowledge accumulates as technology progresses,” he wrote, there is ever more to know, and so the “educational burden” on researchers increases; eventually, it becomes harder not just to know everything but to know what’s needed for you, in particular, to push knowledge forward. Ironically, science may be slower precisely because it’s moving quickly in too many directions.

Organizational and communicative challenges exist alongside what might be called problems of ambition: science can be slow simply because, at this point, its aims are Olympian. Kristin Persson, a materials scientist at the University of California, Berkeley, explained one overarching goal in her field: predicting how different arrangements of atoms will behave under different conditions, so that they can be combined into new and useful materials. “For a very long time, we’ve been trying to do what some people call multi-scale modelling,” she said. “You start out with atoms and electrons, and you have an equation for behavior on that length scale and time scale. Then you take those properties and you feed them into higher-scale equations.” The approach, she said, has “never really taken off, because of what’s lost in translation every time you move from one set of equations to the next.”

It’s hard, in other words, to precisely predict the behavior of matter at every scale. The theoretical basis for such predictions was established a century ago—in 1926, as it happens—by the physicist Erwin Schrödinger, who devised a single equation describing how electrons interact under the laws of quantum mechanics; it’s relatively simple to solve for small numbers of electrons, but becomes unwieldy as atoms add up. If you want to use it at a large scale, you can solve it only approximately, unless you want your computer to calculate for billions of years. Science is full of such bottlenecks: places where dramatic insights, some long-standing, can’t yet yield practical results.

Today’s computers may not be fast enough to effortlessly apply Schrödinger’s Equation—but they are, apparently, fast enough to read and think. Accordingly, Levin told me, scientists are using artificial intelligence to respond to the burden of knowledge. “They take that stack of papers and put it through an A.I., and say, ‘Summarize this for me,’ ” he said. Levin regards it as a useful innovation, wielded by experts, that might itself be accelerated. “If all my work is going to get filtered through an A.I., then why am I putting out a traditional paper, with all of this stuff we do for each other—with the freaking grammar, and the PDFs, and the figures, which the A.I. has to take apart?” he asked. He anticipates a near future in which the form of the scientific paper becomes A.I.-ready, with old knowledge ably surfaced by artificial intelligence alongside new work. “There are amazing authors with amazing sets of papers from twenty, thirty years ago that nobody knows today,” he said. “Theoretically, that information is there somewhere, but, if nobody knows it and nobody uses it, then do we really have it?”

It could be that some version of this transformation is already under way. Scientists are increasingly using large language models not just to read papers but to write them: studies from last year, when A.I. was less capable, found that many researchers had experienced writing-productivity boosts of between thirty-three and fifty per cent, and submissions to journals and to conferences are up, in some cases dramatically. Meanwhile, in a study titled “AI Article Reviews Dominate Human Peer Reviews at Their Best,” two Harvard researchers, Gary King and Steven Worthington, show that peer reviews performed by L.L.M.s “are now better, by every measure we could construct, than reviews written by well-compensated human experts”; they’re even “judged to be of unusually high quality by journal editors and authors.” It’s possible to discern, in these developments, a new rhythm of scientific communication.

L.L.M.s work with language: it’s not surprising that they can be used to process information. But what about the actual labor of science, which is concrete, empirical, and often outrageously time-consuming? “There’s a lot of ‘donkey work,’ ” Persson, the materials scientist at Berkeley, said. Experiments must be done painstakingly, over and over, for years, and even then they can produce unpredictable results; a brilliant young researcher might spend her whole postdoc on the creation of a statistical model that’s immediately bested by someone else’s. Scientific progress is slow, in part, because there aren’t that many passionate, hyper-educated people willing to devote themselves to hands-on, detail-oriented, ruthlessly competitive, and often fruitless labor.

It’s clear, at this point, that A.I. can automate some of the donkey work. “We have a bunch of computational neuroscientists in my lab, and every last one of them says, ‘I can now do in a week what would have taken me six months,’ ” Marder, the neuroscientist at Brandeis, told me. Materials scientists use computer simulations that can easily run off the rails. “When I was a graduate student, I logged in late at night to see if my simulation had worked out,” Persson said; she’d then spend untold hours troubleshooting. Now, she explained, “L.L.M.s are learning from errors in the simulations and applying their corrections autonomously to keep them on track.” This leaves time-constrained researchers free to focus on higher-level questions.

The advantages may go beyond vibe coding. A.I. systems “are so helpful at solving problems because they can try lots of things,” Michael Brenner, a professor of applied math and physics at Harvard who also helps direct A.I. science efforts for Google, told me. At Google, Brenner led the development of a system called Empirical Research Assistance (ERA), which scientists can use to create data-analysis tools. In many ways, ERA also addresses the burden of knowledge: researchers pose questions to it, and provide it with data, and it responds by trying and mixing different extant methods of analysis in numberless combinations, discovering new paths to progress. (Researchers have used ERA to create improved disease-forecasting models, for example.) “It’s like swapping and flipping, flipping and swapping the pieces, and trying to figure out something that actually works,” Brenner said. The power of the system, he went on, was “not Einsteinian but Edisonian”: it consisted in the novel application of known ideas.

Could A.I. offer even more in the way of novelty, contributing to the automated production of new data, new knowledge, and new theories? Both Levin and Persson described automated laboratories as a logical next step. Currently, Persson said, the creation of new experimental data is extraordinarily expensive and time-intensive—but robots, she argued, could be used “to produce data at scale,” which could then be analyzed by existing A.I. models or used to train new ones. (This data, she noted, might also have the advantage of being unusually consistent and free from human error.) “We’re going to get lab automation,” Levin said. “Then there’s going to be no more concern with hallucination or confabulation—the A.I.s are going to try stuff, and it either works or it doesn’t. At that point, once they have the ability to actually test out the consequences of their hypotheses, the loop is going to accelerate, and all this is going to go exponential.” This vision—in which A.I. speeds up not just the collection and analysis of data but the generation of hypotheses and the design of experiments—represents one of the more radical acceleratory possibilities for science. Levin suggested one major obstacle: “A.I. has been obviously trained on human output where conventional theories are overrepresented,” he said. To arrive at “unconventional ideas”—the kinds that lead to full-throttle acceleration, and to world-changing invention—it would have to become less consensus-driven.

If dramatic scientific acceleration occurs, it will be significant on at least three levels. First, there’s the question posed by John von Neumann, the computer scientist, in his 1955 essay: “Can we survive technology?” Today, the concept of a technological “singularity” is easily taken to refer to a single Promethean moment, such as the invention of superintelligent A.I. But the concept, which is often credited to von Neumann, can have a simpler meaning. New technologies are frequently hard to manage; as more of them pop up, their management gets more and more difficult, until they become, in aggregate, unmanageable. How many simultaneous “upgrades” can society withstand? This problem concerned von Neumann precisely because he was writing toward the tail end of the century of invention described in “The Rise and Fall of American Growth.” If we are entering such a period again, we need to be as concerned as he was with the problem of technology management. We need to shift from a fetishization of disruption to a fascination with control.

Second, there’s the question of growth. In his book, Gordon shows how much life can improve when Great Inventions are changing the world. During the hundred years he describes, people got healthier and ate better; worked less and travelled farther; moved into bigger, more comfortable houses; and grew increasingly educated while exercising more personal freedom. The economy expanded because lives were expanding. But today, the United States, along with most of the Western world, is facing a crisis of growth. Populations are aging, and the ratio of workers to pensioners is falling. The costs of housing, food, education, health care, and other necessities are rising, as are our debts. “What are we going to do?” the economist Tyler Cowen asked, recently. “Plan A is that A.I. will help us grow more quickly, and we will manage to pay our bills. . . . There is no Plan B.” We need some new Great Inventions; an acceleration in science, enabled partly by A.I., is our most promising path to getting them.

Finally, there’s the not-unimportant matter of human knowledge. Progress in science is good—but it doesn’t always feel good. Often, it centers not on the arrival of answers but on the proliferation of difficult questions. The same may be true for progress in the practice of science. Sudden advances in mathematics have led some mathematicians to say that their field has been “broken” by artificial intelligence. “I don’t think math is broken,” Brenner told me. “It’s hard for the community, and some of my math colleagues might disagree with me, but I think it’s a very exciting time in mathematics, a time to really rethink things and break new ground.” Persson agreed: “Maybe the mathematicians are having a rough time, but honestly I would congratulate them,” she said. “You should talk to some of my students: They see a lot of their mundane work going away. They can focus on fun things; they can do simulations they could never do before.” It was, she suggested, a great time to become a scientist. ♦

❌
Subscriptions