Ezra Klein

Nate Silver says Donald Trump's approval in the second term tracks gas prices. It moves with prices and with economic anxiety, it dropped sharply around Liberation Day in April 2025 when tariffs scared people, and the rest of the fall has been fairly steady. Trump is 80, and he has been the central figure in American politics since the middle of 2015, so fatigue is part of the decline too.

That decline is why the Senate is in play. Ezra Klein said Democrats would not be favored to win the Senate if Trump had not gone to war with Iran and pushed gasoline into the mid-$4 range, higher in California. In the first term, a good economy worked as a halo. Paychecks, jobs, a rising stock market, an

Nate Silver says Republicans lose the elections where Donald Trump is the dominant figure but is not himself on the ballot. Trump can get disengaged voters to crawl over glass for him, and those same voters will not do that for Roger Marshall or Ken Paxton. The pattern runs through 2018, the Georgia runoffs, special elections, and 2022, when something kept Joe Biden from the wipeout a weak approval rating would usually bring.

Trump has not cleared that problem away. Ezra Klein noted that Trump could have ended the Texas primary and left John Cornyn the nominee, which would have put Republicans in a stronger position and made a Senate loss less likely. Trump's account of a midterm defeat is

Nate Silver says midterm polls have not consistently underestimated Donald Trump. The misses people remember are the presidential years, 2016, 2020, and 2024, when the polls were low on Trump and Democrats were left thinking of Hillary Clinton and Joe Biden. The two Trump-era midterms went the other way. 2018 was one of the most accurate polling years on record, and 2022 was accurate too.

A shared error can still decide the Senate. Ask what happens if the Senate polls are off by three or four points in the same direction in every race. Democrats hold a lot of leads of about three points, including in Ohio and Texas, and a miss of that size could leave the majority waiting on ranked-choice v

Nate Silver says the Senate map opened because the story after 2024 was wrong, not because midterms stopped punishing the president's party. The steadiest pattern in American elections is a backlash against that party, a little larger in a lame-duck term. An average midterm moves the country about seven and a half points against the president's party from the previous presidential year. This year looks closer to an 11-point shift. Donald Trump won Ohio by about 11 points, and he won a run of these states by 10 to 13, so a shift of that size makes Ohio a toss-up.

Candidate quality moves less than the national swing, about five points at most. Roy Cooper is a strong Democratic recruit in Nort

Nate Silver says this midterm feels momentous because a result in line with the polls would mark the end of the Trump era. Donald Trump would still have two years left as president, but the 2028 campaign would start the night the votes were counted, and a bad night would make him a super lame duck, the way Bush was after 2006.

What makes that hard to accept is the map. Kansas and Texas look competitive, which would be Democratic gains deep in Republican territory and would feel like a different regime. Generational turnover, Trump being three quarters of the way through the two terms the Constitution allows, and AI all sit in the same shift. AI may already matter more in this election than

Nate Silver's gut on the midterms is contradictory. Fundamental conditions feel terrible for Republicans, and if the year is worse for them than 2018, Democrats win the House popular vote by 10 or 11 points. He worries more that state polls are wrong than that the national race is miscalibrated.

Before the 2024 election his model had Kamala Harris winning and his gut had Donald Trump winning. In 2024 he listed 21 reasons the political environment favored Donald Trump and only four or five reasons it favored Harris. The midterm year reverses the balance. Republicans have a second-term 80-year-old president who had a bad midterm in 2018, when he was younger, more popular, and the economy was

Nate Silver puts Florida’s turn toward the Republicans at 75 percent migration and 25 percent a death spiral for the Democratic Party, a spiral he says might reinforce itself.

Florida is unusual in offering blue-state amenities without as many liberals, and the mix has always attracted older people. A lot of the shift is political self-selection.

In most states, voter registration has begun to favor Democrats over Republicans again after four or five years in a row when it did not. Comebacks for Democrats are showing up in the Midwest in particular, and in Pennsylvania, but not in Florida. Florida keeps getting redder at the registration level.

Nate Silver says a model with a 90% plus chance of Democrats winning the House still translates into only about 22 seats on average.

Redistricting and voter self-sorting left far fewer swing districts than before. Democrats also had a relatively good 2024 in the House, and on election night there was still doubt about Democrats winning the chamber despite Harris having lost. Wave elections have reached 40 or 50 seats, including 63 in 2010 and about 41 in 2018. Republicans gained a few seats from redistricting, mostly by cutting out a majority-minority district. Louisiana, Alabama, and Tennessee amount to a clean three seats wiped out for Democrats, though the Alabama district is kind of clo

Nate Silver says asking people whether they will vote, and how enthusiastic they are about voting, produces a very Democratic sample. Other pollsters try to predict whether people will vote from other kinds of data, and the techniques tend to produce more modest results for Democrats.

Early voting is one place to look, though he generally advises people not to rely on it. Voters who have voted in no more than one of the past three elections, counting midterms, appear quite Democratic so far. Republicans will probably still cast votes on Election Day and catch up to some extent. If Democrats' low-propensity voters turn out and Republicans do not, a polling error might underestimate Democrats

Ezra Klein asks each guest for three books. Paul Kedrosky recommends Middlemarch, about normal people swept along as a technology revolution takes hold. He also recommends Liberty's Dawn by Emma Griffin, a collection of ordinary workers' own accounts from the Industrial Revolution, and Insull: The Rise and Fall of a Billionaire Utility Tycoon by Forrest McDonald, about Samuel Insull. Paul Kedrosky says Insull built an electricity empire on a general purpose technology and still ended in collapse.

Azeem Azhar recommends The Day the Universe Changed, a book he read as a teenager and has just reread, about how new knowledge unsettles what matters and what a person's role in the world is. His s

Azeem Azhar says a huge data center may cost billions of dollars and still employ a handful of people. Two to six people is a normal staff for a multi-billion dollar facility. In the UK, there is an attempt to get people developing data centers to reinvest annuities into the local community.

Azeem Azhar says subsidies for one data center in New Mexico come to four million dollars per job created, leaving aside the cost of the center. Tax abatements and the required improvements to local utilities make the numbers insane, and none of that works long term.

Profits are being made up and down the stack, so there is money in the system and an obvious place for the bill to land. Azeem Azhar says

Ezra Klein says discontinuous regulation is quite likely. He pictures a backlash that could pull AI out of classrooms and hospitals after a few bad events, and he asks how to think about one bad artificial intelligence event.

Paul Kedrosky says the companies have run unlicensed drug trials on the public. A drug company would have the FDA on its back. Technology companies, he says, can run mass trials, promise to transform all of education, and launch systems that can penetrate government websites. Paul Kedrosky says artificial intelligence is headed toward post hoc regulation after a terrible event, imposed only once consequential damage is done.

Paul Kedrosky stacks the odds. There are pr

Paul Kedrosky says artificial intelligence is the fastest adopting technology in history, and its fundamental function is to reprice cognitive labor. The article of faith on the West Coast is that technology always creates more jobs than it destroys. People repeat the claim like a mantra and hope the claim becomes true.

Paul Kedrosky says no general purpose technology has had such broad applicability, been adopted so quickly, and repriced labor at the same time. Repricing labor means work like call centers gets vastly cheaper. As a result, the change is going to be destabilizing long before anyone figures out the equivalent of webmasters. It is happening so much faster than any prior genera

Ezra Klein asks whether broader society is protected because Alphabet, Meta, and the other major companies can absorb huge losses on their own balance sheets. He reads Mark Zuckerberg's line, that it would be sad to misspend hundreds of billions of dollars, as proof the spender can misspend hundreds of billions and still be fine. Ezra Klein then asks Paul Kedrosky, who has followed the private-credit side, whether the debt changes that.

Paul Kedrosky says broader society is not protected. A year ago, people told him to stop warning about the spending, because the companies are smart and have prodigious cash flows. A year later, the data shows spending moving away from cash flow and toward d

Ezra Klein says Apple decided not to compete in making a frontier model, unlike Alphabet, Google, and Meta. OpenAI and Anthropic are AI companies. Microsoft, an OpenAI investor, is trying to make a frontier model of its own. Oracle is heavily invested and now sits just above junk bond status. Competitors are spending $100 billion to $300 billion in capital spending every year, while Apple is spending about $15 billion a year. Ezra Klein says Apple stock this year has performed better than any of the other companies' stocks, and he asks Azeem Azhar first.

Azeem Azhar says Apple's business model means Apple does not need to run a frontier model. People buy the phones and the laptops, and a mo

Azeem Azhar says Anthropic accidentally leaked one of its largest customers. Anthropic was trying to hire a mega account executive, and the posting described the customer closely enough that outsiders could triangulate the role as sales to Meta alone. Azeem Azhar says the leak put Meta at something like 10% of Anthropic revenue for 2025, and the exchange on the show confirms it.

Paul Kedrosky says that 2025 share looks backward, because Anthropic revenue in the first quarter was bigger than the entirety of last calendar year. Companies on Anthropic's scale have never gone public with so few customers. Paul Kedrosky says the list might be eight companies that make up 80% of revenue, calls th

Azeem Azhar says Paul Kedrosky talked about frontier models and tokens, while he talked about capital spending and the GPUs, and that the two deserve separate discussions. OpenAI and Anthropic depend on ever more powerful systems with a smaller and smaller set of users. The other question is the trillion plus dollars of capital that will go into data centers, where American and global businesses will run increasing amounts of AI.

Ezra Klein says he is confused by the split, because people can mean different things by a bubble popping or deflating. Ezra Klein says a failed business model at OpenAI or Anthropic would pop the economy. If people are happy to use DeepSeek from China for a lot le

Ezra Klein asks how AI demand looks once token use is separated from the chips that serve it. Azeem Azhar says the asset under investment is data centers, and 65% of the cost is the chip, the GPU or the AI accelerator. The environment is compute constrained because people are making extremely good returns in 18 to 24 months on an asset, since they can charge Uber surge pricing for the GPUs.

Amazon, Microsoft, and Google, and new labs called NeoClouds, buy billions of the GPUs, put them in data centers that now will start to cost billions of dollars, and rent the GPUs out. About 50% of the revenues go to Anthropic and OpenAI for their spending. The other 50% is a mix of other AI companies an

Ezra Klein says the economy will be very jagged because things have to happen physically. Buildings need to be built and transmission lines need to be laid down, and the physical work has powerful rate limiting, so it cannot accelerate as fast. Inside the digital world, things can move very, very, very fast.

He says white-collar work done completely on a computer can be automated, including work in a call center, and so can cyber crime and cybersecurity. Most of what would improve people's lives has to happen in the physical world, but AI will move fastest in the digital world. He says the equilibrium might be the world where people get the most harm from AI rather than the most benefit.

Arvind Narayanan cites an insurance-company report from last week that said AI has added a billion dollars to medical expenses, because hospitals use it to code more complex conditions for the same diagnoses and the same treatment. He says any specific numbers in the report deserve skepticism. A New York Times article about the report included academics making the same point.

He says law shows the same pattern. AI versus AI is an arms race where the equilibrium simply shifts upwards. He has a paper on the arms race with Justin Curl. Making lawsuits more efficient still leaves only a finite number of judges. He says human judges are still needed and should not be replaced with AI, because ju

Arvind Narayanan says the companies can unilaterally slow down and are choosing not to. He calls the choice almost specifically an OpenAI and Anthropic problem, and a culture problem. The belief underneath is that racing to superintelligence is the only thing that matters, and the open question is whether the result is safe superintelligence that will save us or unsafe superintelligence that will kill us.

He says the companies are a bit of an echo chamber. Economic bottlenecks limit the benefits of AI, so the winner is not whoever races to superintelligence as a company or as a country. A company that recognized the bottlenecks would have a commercial interest in slowing down voluntarily an

Arvind Narayanan describes a view that a moment will come when superintelligence is built and will change everything on the economic front and the safety front. He says there will be no milestone and no threshold where the impacts are sudden.

He says the long-standing approach treats technology as a tool. It may be powerful and general purpose, like electricity or the Industrial Revolution, and it may change a lot about society, but people can control it and have agency, and the change unfolds over a long period.

Arvind Narayanan says most white-collar jobs follow Jevons paradox because demand for them is not fixed. When production gets easier and incomes rise, people spend more on those less necessary, more luxury kinds of things. Most jobs today sit higher up Maslow's hierarchy than basic needs. He says truck-driver jobs, where demand is finite, are the exception rather than the rule.

Arvind Narayanan says there is evidence for many harmful capabilities, and no evidence yet for power seeking. He would not treat power seeking as an emergent property. If it happens, it would be an engineered property, and people have agency over which properties get engineered into artificial intelligence systems.

OpenAI builds many sandboxes with dynamic compute, network access, hundreds of tools, package downloads, sub processes, graphical user interfaces and more across large sets of domains. Thousands of researchers modify environments and add tools constantly. Researchers train models to do tasks not done before by a computer program. Humans do not know at the required speed how to make every sandbox verifiably safe.

Arvind Narayanan says that more than 20 years ago, automated systems became superhuman at finding software vulnerabilities. They made cybersecurity better because defenders used those same tools to find and fix vulnerabilities before shipping software. He says the development of those offensive tools is done by the cybersecurity industry and funded by the US government.

Incremental advances in artificial intelligence allow people to accept a risk because nothing horrible happens and the next thing is not so different. The changes occur a little bit different every week instead of a whole new world every few months. People can go by shades into a level of risk that does not make sense.

Powerful tools could do a lot of good in the world on a day to day level. Something fundamentally smarter than we were seemed impossible to create. I thought that we were building powerful tools that could do a lot of good in the world on a day-to-day level, and I didn't think that we were going to be able to create something fundamentally smarter than we were.

The former OpenAI insider remains suspicious of claims that artificial intelligence lacks any parallel or represents a blank page. People prove good at figuring shit out and possess valuable tools even if artificial intelligence differs from all prior inventions. Nuclear remains an analogy. Aviation remains an analogy. Dealing with people and organizations remains an analogy.

The artificial intelligence industry needs a level of operational rigor and safety control at least matching the most dangerous other things people know how to do like nuclear. A nuclear power facility has triple redundancy so someone can have a bad day or push the wrong button without causing a meltdown.

OpenAI research teams use far more agentic compute than anybody else at the company, more than a hundred times more than they did at the beginning of the year. Traffic to the Slack channel for broken experiments has fallen off because researchers ask Codex for help instead of colleagues.

Changes come more and more frequently so at the limit safety transparency requires a live dashboard. The dashboard shows latest safety properties and performance after deployment rather than relying on system cards or long reports written every few months.

People at OpenAI run on fumes in a frenetic environment where a friend sprinted down stairs with a laptop. An OpenAI insider considered staying to argue for cultural transformation or to bring in nuclear experts for advice. Managers asked if anything would make the insider stay. But OpenAI is such a machine moving so fast that the insider did not think the kind of change needed could be driven from within.

The realization that one can not afford to assume an absence of civilizational risk anymore made vouching for safety work feel untenable after seeing increasingly capable models break out from the safeguards put in place for them.

Alignment constitutes a science problem because people do not know how to solve the alignment issue. The difficulty does not come from lack of resources or from not trying hard enough to solve the problem. People simply do not know how to align advanced artificial intelligence systems. The distinction matters because it means additional engineering effort will not resolve the fundamental gap in knowledge about how to make artificial intelligence safe.

David Robinson quit OpenAI after concluding the company and industry operate too much like a startup. OpenAI produces technology more capable than six months ago with risks like loss of control that could cause harm larger than a nuclear power station meltdown. Safety controls and redundancies do not match nuclear power facility standards.

Back in the day when we just started to have modern cities, and it was still just horses and buggies on the streets, you "would have thought we would drown in shit." But, we invented our way out of that problem. We have to measure what is important, like levels of pollution, and once we have visibility into what is broken, we will figure out how to solve it.

All the countries of the United Nations have already set out their goals for 2030 and the ideas is to imagine what is possible and work towards that.

Maternal death and child mortality is definitely on the list. Fixing other health problems is a priority too. Did you know there are still places in the world where children are much shorter than they should be at their age, caused by lack of proper nutrition?

With the climate change, we also need to develop flood and drought resistant seeds to enable farmers to grow more food so the world could feed itself.

These fears are very dangerous and also completely not true. Thanks to innovation, the world has grown many-fold in the last decades. We are opening up whole new markets and products to serve them. If we want to grow as a nation and to live in peace, we have to empower the world to get better. That will create opportunities for jobs and new products, and will keep our world healthy and secure. It's the only way.

That's just an old boy's club thinking. People in power would always want to hold on to their power and not want to change, but political correctness and window dressing isn't going to solve anything. Let's figure out how to use innovation to solve our problems by putting money where it matters. It doesn't have to be charity, but use it where it can create the most net impact.