Generative-AI companies have been selling a narrative of unprecedented, endless progress. Just last week, OpenAI introduced GPT-4.5 as its “largest and best model for chat yet.” Earlier in February, Google called its latest version of Gemini “the world’s best AI model.” And in January, the Chinese company DeekSeek touted its R1 model as being just as powerful as OpenAI’s o1 model—which Sam Altman had called “the smartest model in the world” the previous month.

Yet there is growing evidence that progress is slowing down and that the LLM-powered chatbot may already be near its peak. This is troubling, given that the promise of advancement has become a political issue; massive amounts of land, power, and money have been earmarked to drive the technology forward. How much is it actually improving? How much better can it get? These are important questions, and they’re nearly impossible to answer because the tests that measure AI progress are not working. (The Atlantic entered into a corporate partnership with OpenAI in 2024. The editorial division of The Atlantic operates independently from the business division.)

Unlike conventional computer programs, generative AI is designed not to produce precise answers to certain questions, but to generalize. A chatbot needs to be able to answer questions that it hasn’t been specifically trained to answer, like a human student who learns not only the fact that 2 x 3 = 6 but also how to multiply any two numbers. A model that can’t do this wouldn’t be capable of “reasoning” or making meaningful contributions to science, as AI companies promise. Generalization can be tricky to measure, and trickier still is proving that a model is getting better at it. To measure the success of their work, companies cite industry-standard benchmark tests whenever they release a new model. The tests supposedly contain questions the models haven’t seen, showing that they’re not simply memorizing facts.

[Read: The words that stop ChatGPT in its tracks]

Yet over the past two years, researchers have published studies and experiments showing that ChatGPT, DeepSeek, Llama, Mistral, Google’s Gemma (the “open-access” cousin of its Gemini product), Microsoft’s Phi, and Alibaba’s Qwen have been trained on the text of popular benchmark tests, tainting the legitimacy of their scores. Think of it like a human student who steals and memorizes a math test, fooling his teacher into thinking he’s learned how to do long division.

The problem is known as benchmark contamination. It’s so widespread that one industry newsletter concluded in October that “Benchmark Tests Are Meaningless.” Yet despite how established the problem is, AI companies keep citing these tests as the primary indicators of progress. (A spokesperson for Google DeepMind told me that the company takes the problem seriously and is constantly looking for new ways to evaluate its models. No other company mentioned in this article commented on the issue.)

Benchmark contamination is not necessarily intentional. Most benchmarks are published on the internet, and models are trained on large swaths of text harvested from the internet. Training data sets contain so much text, in fact, that finding and filtering out the benchmarks is extremely difficult. When Microsoft launched a new language model in December, a researcher on the team bragged about “aggressively” rooting out benchmarks in its training data—yet the model’s accompanying technical report admitted that the team’s methods were “not effective against all scenarios.”

One of the most commonly cited benchmarks is called Massive Multitask Language Understanding. It consists of roughly 16,000 multiple-choice questions covering 57 subjects, including anatomy, philosophy, marketing, nutrition, religion, math, and programming. Over the past year, OpenAI, Google, Microsoft, Meta, and DeepSeek have all advertised their models’ scores on MMLU, and yet researchers have shown that models from all of these companies have been trained on its questions.

How do researchers know that “closed” models, such as OpenAI’s, have been trained on benchmarks? Their techniques are clever, and reveal interesting things about how large language models work.

[Read: The GPT era is already ending]

One research team took questions from MMLU and asked ChatGPT not for the correct answers but for a specific incorrect multiple-choice option. ChatGPT was able to provide the exact text of incorrect answers on MMLU 57 percent of the time, something it likely couldn’t do unless it was trained on the test, because the options are selected from an infinite number of wrong answers.

Another team of researchers from Microsoft and Xiamen University, in China, investigated GPT-4’s performance on questions from programming competitions hosted on the Codeforces website. The competitions are widely regarded as a way for programmers to sharpen their skills. How did GPT-4 do? Quite well on questions that were published online before September 2021. On questions published after that date, its performance tanked. That version of GPT-4 was trained only on data from before September 2021, leading the researchers to suggest that it had memorized the questions and “casting doubt on its actual reasoning abilities,” according to the researchers. Giving more support to this hypothesis, other researchers have shown that GPT-4’s performance on coding questions is better for questions that appear more frequently on the internet. (The more often a model sees the same text, the more likely it is to memorize it.)

Can the benchmark-contamination problem be solved? A few suggestions have been made by AI companies and independent researchers. One is to update benchmarks constantly with questions based on new information sources. This might prevent answers from appearing in training data, but it also breaks the concept of a benchmark: a standard test that gives consistent, stable results for purposes of comparison. Another approach is taken by a website called Chatbot Arena, which pits LLMs against one another, gladiator style, and lets users choose which model gives the better answers to their questions. This approach is immune to contamination concerns, but it is subjective and similarly unstable. Others have suggested the use of one LLM to judge the performance of another, a process that is not entirely reliable. None of these methods delivers confident measurements of LLMs’ ability to generalize.

Although AI companies have started talking about “reasoning models,” the technology is largely the same as it was when ChatGPT was released in November 2022. LLMs are still word-prediction algorithms: They piece together responses based on works written by authors, scholars, and bloggers. With casual use, ChatGPT does appear to be “figuring out” the answers to your queries. But is that what’s happening, or is it just very hard to come up with questions that aren’t in its unfathomably massive training corpora?

Meanwhile, the AI industry is running ostentatiously into the red. AI companies have yet to discover how to make a profit from building foundation models. They could use a good story about progress.

20 August 1991: Soviet armored vehicles roll down the main highway toward the Tallinn television tower. A coup attempt is under way in Moscow and at Mikhail Gorbachev’s dacha in Crimea. The Baltic States were moving toward independence and needed to be brought back under control. Seizing the television tower to control communications was critical.

But Estonian civilians fought back and protected the tower. The Soviet army stood down.

In the late 20^th century, control of television and radio stations was an early step in a coup. Sometimes it worked. The point was to broadcast that “We are now in control of the government.”

Elon Musk is now attempting an analogous move with computers. “We control all your data.” But he’s got a problem.

A couple of problems, really. In 1991, one television tower controlled television across Estonia. A number of computers in Washington are central to the way the government works, but no one fully controls. That leads to the second problem: You have to understand how the computers work together.

Government is at its best when most people don’t notice what it’s doing. Flying is safe, food is safe, income tax refunds and other checks arrive on time. Most people don’t have to understand the details, and most don’t.

Elon Musk doesn’t. He grew up in another country. He believes that he is the master of everything, can learn anything in an hour or two. The boys working for him have been encouraged by the Silicon Valley ethos not to spend time in school, where they might have been exposed to a cursory idea of the structure and function of the government. The adults working for Musk have less excuse, but all of them live in the Silicon Valley bubble.

The thing about a coup, though, is that you have to do it fast. The Estonian tower operators sat in the top floor, the story goes, and wedged a matchbook into the elevator door to disable it so that the Soviet soldiers had to climb up the stairs. One thousand of them. Meanwhile, the operators broadcast for help. Civilians gathered outside, outnumbering the soldiers.

Musk, or one of his lieutenants, figured out that the Office of Personnel Management was a central repository of personnel information and that Treasury is where the checks are printed. But this is a government for a country with 340 million people. The government employs more people than the population of Estonia.

What Musk didn’t figure out is that the decisions associated with hiring and firing people and with sending out checks resided elsewhere. So they hit OPM and Treasury first. It’s a strong move, but not as strong as they may have thought.

Musk promised to remove $2 trillion dollars from the $6 trillion national budget. That’s a lot. The big money is not in personnel, it’s in Social Security, Medicare, Medicaid, and big contracts in the Defense Department and NASA. Musk’s website currently claims, with a great many errors, to have removed $2 billion dollars. They only need to find 999 times that more to make up the $2 trillion.

Personnel seemed the easiest, after some forays into ending contracts and overhead for NSF and NIH contracts, much smaller than the defense and space contracts, but damaging to those nasty people who thought they knew so much about COVID and insisted we wear masks. A large motivation for this effort seems to have been to remove civil rights protections from the workforce and generally cause chaos, along with whatever dollar amounts could be claimed.

After mucking around in the OPM and Treasury computers, installing an “email all” function at OPM to every government employee, they realized that they would have to go to the agencies. Presumably Musk was whipping them on for speed, but there was no low-hanging $2 trillion to be found. Aha! Employees on probationary status could be fired more easily than others. Probationary status included recent hires and those recently promoted or moved to new jobs.

Not knowing the functions of the agencies, they fired people who were watching over nuclear weapons and bird flu and running the national parks. Some are being hired back, but the Musk operation didn’t retain contact information. Some agencies, like the FBI and CIA, seem to have rebuffed the cuts..

Over the weekend, Musk sent out an email, as he did when he took over Twitter, demanding progress reports from the entire workforce. Not all agencies are complying. It’s another pretext for firing workers. They didn’t get enough with the provisional workers.

The pace is slowing, and resistance is rising. Citizens at congressional town halls are demanding that Musk and his boys be removed from the computers. Lawsuits are in progress against their actions.

Matt Bai has interviewed people run over by the Musk operation (gift link), with more detail. I’ve developed this scenario on the basis of information available publicly. I don’t have any inside information. The structure of the federal government has stymied some of Musk’s destruction, and the slowing pace is allowing resistance to form, just as that matchbook held the Soviet soldiers at bay. Twenty of the civil service employees in the office that was made into DOGE have resigned in protest against the actions, and lawsuits questioning the legality of DOGE are in progress.

Tallinn Television Tower photo by Dmitry G – Own work, Public Domain

Cross-posted to Lawyers, Guns & Money

Here's a chart showing airline fatalities over the past century. In 1980, at about the same time that commercial airlines began hiring Black and women pilots, fatalities began a 30-year decline to zero:

This is not a smoking gun. Fatalities likely went down for reasons of their own that were unrelated to civil-rights era hiring practices. Nonetheless, it almost certainly shows that "DEI hires" didn't produce any increase in accidents or fatalities. The evidence is flatly inconsistent with that.

So don't worry about the color or gender of your pilot the next time you board an airplane. It doesn't matter.

The nostalgia behind the slogan “Make America great again” has always provoked the obvious questions of just when America was great, and for whom. Early in the second Trump administration, we are getting the answer.

In August, speaking with someone he believed to be a sympathetic donor, one of the Project 2025 architects, Russell Vought, said that a goal of the next Trump administration would be to “get us off of multiculturalism” in America. Now Vought is running Donald Trump’s Office of Management and Budget, and the plan to end multiculturalism is proceeding apace. Much of the chaos, lawlessness, and destruction of the past few weeks can be understood as part of the administration’s central ideological project: restoring America’s traditional hierarchies of race and gender. Call it the “Great Resegregation.”  

[From the January/February 2024 issue: Civil rights undone]

Since taking office, Trump has rescinded decades-old orders ensuring equal opportunity in government contracts and vowed to purge DEI from the federal government, intending to lay off any federal worker whose job they associate with DEI. Yesterday evening, Trump fired the chairman of the Joint Chiefs of Staff, General Charles Q Brown, and replaced him with a lower ranking white official, a retired three-star Air Force officer named Dan Caine. Secretary of Defense Pete Hegseth had previously attacked Brown as an unqualified diversity hire based on the fact that he is Black. Trump’s Department of Justice has implied that it will prosecute or sue companies that engage in diversity outreach. Elon Musk’s DOGE is attempting to purge federal workers “that protect employees’ civil rights and others that investigate complaints of employment discrimination in the federal workplace,” the Washington Post reported. Colleges and universities are being threatened with defunding for any programming related to DEI, which the free-speech organization PEN America has noted could include “everything from a panel on the Civil Rights Movement to a Lunar New Year celebration.”

Trump has also signed executive orders that threaten government funding for scientific research on inequality or on health issues that disproportionately affect nonwhite ethnic groups, and has imposed censorious gag orders that could block discussion of race or sex discrimination in American classrooms. During her confirmation hearing, Trump’s education-secretary nominee, Linda McMahon, said she did not know if schools could lose funding for teaching Black-history classes under the order. The legality of the order over K–12 curricula is unclear, but the chilling effects are real nonetheless.

Under the Trump administration, schools within the Department of Defense system that serve military families—American service members are disproportionately Black and Hispanic—have torn down pictures of Black historical figures and removed books from their libraries on subjects such as race and gender. This record, within a school system entirely under the administration’s control, offers an alarming preview—one in which a historical figure like Harriet Tubman is no longer a welcome subject in educational settings because she was a Black woman.

An OMB memo ordering a federal-funding freeze illustrates the ideological vision behind these decisions. The memo states that the administration seeks to prevent the use of “federal resources to advance Marxist equity, transgenderism, and green new deal social engineering policies,” Acting Director Matthew Vaeth wrote. Equal opportunity in employment is described here as “Marxist,” because it affirms what the desegregators see as an unnatural principle: that nonwhite people are equal to white people, that women are equal to men, and that LGBTQ people deserve the same rights as everyone else.

If the Great Resegregation proves successful, it will restore an America past where racial and ethnic minorities were the occasional token presence in an otherwise white-dominated landscape. It would repeal the gains of the civil-rights era in their entirety. What its advocates want is not a restoration of explicit Jim Crow segregation—that would shatter the illusion that their own achievements are based in a color-blind meritocracy. They want an arrangement that perpetuates racial inequality indefinitely while retaining some plausible deniability, a rigged system that maintains a mirage of equal opportunity while maintaining an unofficial racial hierarchy. Like elections in authoritarian countries where the autocrat is always reelected in a landslide, they want a system in which they never risk losing but can still pretend they won fairly.

The battles of the Great Resegregation are now taking place in at least three overlapping arenas. The first is politics, where right-wing legal organizations have succeeded in rolling back many civil-rights-era voting protections; they want to now fully destroy the remaining shreds. The second is education and employment, particularly at elite institutions, such as the media and academia; right-wing legal strategies have been similarly fruitful here in attacking diversity, thanks to the conservative capture of the Supreme Court. The third is popular culture, where conservatives have sought to leverage anger and nostalgia against movies, television, books, and other creative media brought to life by artists of color.

The term DEI, frequently invoked by the Trump administration, functions as a smoke screen. It allows people to think that the Trump administration’s anti-DEI purge is about removing pointless corporate symbolism or sensitivity trainings. Although it is easy to find examples of DEI efforts that are ill-conceived or ill-applied, some conservatives have leveraged those criticisms to pursue a much broader agenda that is really about tearing anti-discrimination laws out at the roots, so that businesses and governments are free to extend or deny opportunities based on race, gender, and sexual orientation if they so choose.

“This is really taking us back to a kind of pre-civil-rights-movement vision of America,” Sherrilyn Ifill, the former head of the NAACP Legal Defense and Educational Fund, told me in an interview last year, before Trump won the 2024 election. “A backlash is a pushback. This is really much more of a demolition effort.”

As the Trump State Department official Darren Beattie wrote, “Competent white men must be put in charge if you want things to work. Unfortunately, our entire national ideology is predicated on coddling the feelings of women and minorities, and demoralizing competent white men.” This analysis is perceptive in the sense that the exact reverse is true—we are now in the second decade of a years-long temper tantrum sparked by the election of Barack Obama—not to mention the failed attempts to elect a woman to succeed him—and the effect it had on the fragile self-esteem of people like Beattie.

[Read: Is there anything Trump won’t blame on DEI?]

Other MAGA figureheads have promoted similar ideas. In 2020, the conservative writer Christopher Caldwell published a book arguing that the Civil Rights Act of 1964 had revoked “the de jure constitution of 1788, with all the traditional forms of jurisprudential legitimacy and centuries of American culture behind it.” Because of the Civil Rights Act, white people had fallen “asleep thinking of themselves as the people who had built this country and woke up to find themselves occupying the bottom rung of an official hierarchy of races.”

Caldwell’s assessment has grown in popularity among prominent conservatives. The right-wing activist Charlie Kirk has described the Civil Rights Act as having “created a beast, and that beast has now turned into an anti-white weapon,” and he has attacked Martin Luther King Jr., Wired reported, as part of a “broader strategy to discredit” King and “the Civil Rights Act.” On his social network, X, the South African–born Musk, who is playing a key role in the Trump administration, regularly promotes scientific racism, the pseudoscientific ideology that holds that race determines individual potential. Some of the staffers Musk has hired to dismantle the enforcement of anti-discrimination laws seem to share those ideological predilections. One DOGE staffer resigned after the Wall Street Journal revealed he maintained a pro-eugenics social media account where “he appeared to have a special dislike for Indian software engineers.” He was reinstated after receiving public support from Trump and Vance.  The problem conservatives trying to undermine anti-discrimination law seem to have with an “official hierarchy of races” is not that one exists but that, in their warped conception, white people are not on top, as they should be.

This ideology is apparent in the rote blaming of diversity by some conservatives for every catastrophic event—as they did following a midair collision over the Potomac River. Or a freighter crashing into a bridge in Baltimore. Or doors flying off Boeing planes.The contention, overt or implied, is always that unlike white men, whose competence can be assumed, the non-white people with desirable jobs are undeserving. The irony, of course, is that many of the white men making these assumptions are themselves unqualified. Transportation Secretary Sean Duffy is best-known for being a reality-television star.  

Even so, the Great Resegregation seeks not a return to the explicit racial separation of Jim Crow, but rather an embrace of ostensibly “color-blind” policies intended to sustain a de facto segregation that is more durable and less overt, one in which Black access to the middle and elite strata of American life will be ever more rare and fleeting. The numbers of Black doctors, lawyers, scientists, architects, showrunners, and generals would no longer rise. And there would be no more Black presidents. The real but fragile advancement of the Black poor into the Black middle class would be stalled or reversed. Most Black people would be confined to, as Trump memorably put it, the menial “Black jobs” they were meant for, save for those willing to sustain the self-serving fiction that they are among the good ones.

The demolition of multiracial democracy began a dozen years ago, when the Supreme Court’s conservative majority rolled back voting-rights protections adopted in the 1960s to enforce the rights enshrined in the Fifteenth Amendment. Those protections made America, for the first time, a democracy for all its citizens. They diversified Congress, and led to the election of the first Black president. The Roberts Court has steadily eroded those protections, insisting that they are no longer necessary, even as racist ideas once considered beyond the pale return to the mainstream. These changes have had the predictable outcome of increasing racial disparities in voting.

The Roberts Court has treated policies meant to rectify racial discrimination as themselves racist. The Court shut down what remained of public-school integration efforts. It overturned affirmative action in higher education. These decisions have eroded diversity in the classroom. But they’re just the beginning for the resegregators, who intend to ensure that America’s traditional racial hierarchies are persistent and stable.

One clear example comes in the world of higher education. Because giving all Americans equal access to elite higher education is a step toward broader societal integration, such efforts must be shut down. To this end, conservative groups are suing colleges even in states such as California, where affirmative action in public universities has long been banned, claiming that the fact that their incoming classes have become more diverse rather than less is evidence of reverse discrimination. At least two conservative justices have objected to color-blind, class-based affirmative-action programs. This approach suggests a topsy-turvy understanding of racial discrimination, in which a diverse classroom is one in which white men have been discriminated against, based on the conviction that white men are by definition the most competent possible candidates.

[Read: Donald Trump is very busy]

When Trump officials speak of a society that is color-blind and merit-based, they do not appear to mean meritocracy or color-blindness in the traditional sense. Instead of individual meritocracy, they seem to be advocating a racial meritocracy, in which the merit of an individual hire or admission can be assessed not by their individual accomplishments but by how well the group they are associated with fits a particular role. In this way, the Great Resegregation seeks firmer moral ground than the racial apartheid of the past. Racial disparities can be framed not as the result of discrimination, but as a fact: that white people are just better and more qualified. And by withholding federal funding from places that engage in scientific inquiry on social inequalities or offer historical instruction that could be seen as portraying America as “fundamentally racist, sexist, or otherwise discriminatory,” the Trump administration can make the causes of those inequalities illegible.

What the proponents of the Great Resegregation seek is a counterrevolution not merely in law, but also in culture. The civil-rights revolution of the 1960s changed hearts and minds as well as laws, and one of those changes was that racially exclusive institutions became morally suspect. Notably, Trump officials are not willing to state their aims explicitly; they feel obligated to pay lip service to ideals of color-blind meritocracy and mislead about their intentions.

“My view is that the diversity ethos has really sunk deep roots,” the Harvard Law professor Randall Kennedy told me. “There are a lot of people across a wide variety of ideological positions who would not like a racially homogeneous, all-white outfit. Even people who say they’re against affirmative action, they would feel somewhat nervous or somewhat embarrassed or somewhat guilty about that.” Trumpists seek to not just repeal protections against discrimination, but reverse the “diversity ethos” that has enabled America’s tenuous strides toward equality.

And that progress is not only fragile but remarkably incomplete. Neither schools nor workplaces have ever been particularly integrated. Public-school integration stalled long ago. Even prior to the Supreme Court’s decision outlawing affirmative action in admissions, enrollment of Black and Hispanic students at elite universities had stalled at percentages far below their share of the student-age population. Occupational segregation has remained stagnant since the ’90s. Black workers with or without college degrees are concentrated in professions that pay less than those of their white counterparts, despite a rise in Black people obtaining college degrees. Corporate DEI efforts never made much progress on integration to begin with, in part because many of these efforts were more about branding and limiting liability than equal opportunity, and now the federal government will be dead set on reversing whatever headway was made.

“The segregation we see in the labor market right now is three to five times worse than we would expect if race wasn’t a core factor,” Justin Heck of Opportunity@Work, an organization that advocates for workers without college degrees, told me. “We’ve seen it go down a little bit in the years leading up to 1990. But the current world looks the same as it did in 1990. It’s been stagnant or worse, or slightly worse today.” Heck is one of the authors of a 2023 study on occupational segregation published by the National Bureau of Economic Research.

There are perhaps two exceptions. One is the federal government, where until now, anti-discrimination laws have been strictly enforced. Trump’s cronies have tried to discredit the federal workforce precisely because it is often more meritocratic, and therefore more integrated than the private sector. “It’s harder in a federal-government job to get a position simply through an informal network,” the political scientist Ashley Jardina, who also worked on the NBER study, told me. “Whereas in the private sector, especially in building trades, for example, a lot of people are getting their jobs through their social networks, which are incredibly segregated.”

That is why Trumpists are so focused on “ending DEI” in the federal workforce. They see anti-discrimination and inclusion as a ladder of upward mobility for people they do not believe should have one. Under Trump, a workplace or college that is perceived as too diverse might come under legal scrutiny, effectively enforcing racial quotas. For example, Andrew Bailey, the attorney general of Missouri, is suing the coffee chain Starbucks on the basis that after adopting DEI programs its workforce has become “more female and less white.”

The second place where America has grown more integrated is media and entertainment, arenas highly visible to the public. This has depreciated the value of what W. E. B. Du Bois called the “psychological wage” of white racial identity—making those who once held an unquestioned hegemony over American culture feel like something has been stolen from them. And this shift helped fuel the nationwide backlash to diversity efforts that Trump rode to office.

[W. E. B. Du Bois: Strivings of the Negro people]

The slight but substantive integration of characters in film, television, and other forms of entertainment has itself led to a visible backlash, subjecting actors, writers, and other creative workers of color to harassment whenever they participate in a high-profile project, especially in the genres of science fiction or fantasy. An integrated cast, writers’ room, or development team is deemed “woke,” by which critics simply mean integrated, and therefore suspect. A woman, LGBTQ person, or person of color in a leading role is deemed unqualified, or worthy of rejection just because of who they are. What may seem like silly internet controversies are in fact demands for a resegregation of creative workplaces.

“I think probably part of why we observe more integration in some spaces and others also just has a lot to do with the demands that capitalism places on having a market,” Jardina told me. “It earns money for media organizations and studios to diversify their shows and their casts, because there’s a market for that, in the same way that there isn’t in a lot of industries.”

In other words, the exceptions to America’s persistent segregation have taken place in America’s most public-facing professions, among those assigned to interpret the world around them. What people consuming American media see, for the most part, is a mirage of a more integrated America that has yet to come into being. In virtually every other arena—the private-sector workplace, housing, schooling—America remains profoundly segregated, with opportunities limited by class and race.

This is why Trump’s funding freeze has targeted DEI despite no evidence that the government has lowered its standards on behalf of women and minorities. Asked to provide a real example of lowered standards in the military during his confirmation hearings, Defense Secretary Pete Hegseth was unable to. The U.S. military has long led the way in demonstrating how a diverse workforce yields American excellence—one reason some conservatives are fixated on its relative egalitarianism, which they deride as “wokeness.” Hegseth recently said he believed that “the single dumbest phrase in military history is our diversity is our strength." The Nazis and Confederates learned otherwise.

Of course he himself is an illustration of lowered standards—Hegseth has no demonstrable expertise for the job he was given—but because he is a white man, his qualifications for the job are assumed, as a result of the Trumpist concept of racial meritocracy. This is why the funding freeze is targeting research on inequality. It is why private companies are threatened with government lawsuits and prosecutions if they seek a broader pool of applicants. It is why the Trump administration’s deportations do not target merely undocumented criminals but also immigrants on Temporary Protective Status. It is why Trump’s loyalists are dismantling any and all government programs that might conceivably even the playing field between those born with plenty and those born with little.

For all the big talk about putting an end to “social engineering,” the Great Resegregation is itself a radical attempt to socially engineer America to be poorer, whiter, less equal, and less democratic. Much as the old Jim Crow measures kept many southern white people impoverished and disenfranchised alongside the Black southerners they targeted, the Great Resegregation will leave wealthy white elites with a firmer grip on power and the working classes with fewer opportunities and a weakened social safety net. The only people left with more will be those who already had more than they needed to begin with.

“They get the one starving kid in Sudan that isn’t going to have a USAID bottle, and they make everything DOGE has done about the starving kid in Sudan.” — a White House official.

I’ve been a USAID contractor for most of the last 20 years. Not a federal employee; a contractor. USAID does most of its work through contractors. I’ve been a field guy, working in different locations around the world.

If you’ve been following the news at all, you probably know that Trump and Musk have decided to destroy USAID.  There’s been a firehose of disinformation and lies.  It’s pretty depressing.  

So here are a couple of true USAID stories — one political, one personal.


The political one first.  I worked for years in the small former Soviet republic of Moldova.



Moldova happened to be one of the few parts of the old USSR suitable for producing wine.  The other was Georgia, in the Caucasus.

The Soviets, in their central planning way, decided that both Moldova and Georgia would produce wine — but Georgia would produce the good stuff, intended for export and for consumption by Soviet elites.  Moldova would produce cheap sweet reds, which is what most Russians think wine is.

So for decades, Moldova produced bad wine and nothing but bad wine.  But Russians liked it, so that was okay.

Then the USSR collapsed.  And, well, Moldova continued to produce nasty cheap sweet reds, because that was all they could do.   By the turn of the century, wine was Moldova’s single biggest cash export.  And about 80% of that wine went straight to Russia.

This continued through the 1990s and into the early 2000s.  Meanwhile, Vladimir Putin came to power in Russia.  Back in 2003 or so, he wasn’t invading Russia’s neighbors… but he was already swinging a big stick in Russia’s “near abroad”, the former Soviet republics that he thought should still be under Russia’s thumb.  Which absolutely included Moldova.

So whenever the Moldovan government annoyed or offended Putin… or whenever he just wanted to yank their chain… the Russian Ministry of Health would suddenly discover that there was a “problem” with Moldovan wine.  And imports would be frozen until the “problem” could be resolved.  Since wine was Moldova’s biggest export, and most wine went to Russia, this meant that Russia could inflict crippling damage on Moldova’s economy literally at will.  

This went on for over a decade, with multiple Moldovan governments having to defer to Moscow rather than face crippling economic damage.

Enter USAID.  Over a period of a dozen years or so, USAID funded several projects to restructure the Moldovan wine industry. 

They brought in foreign instructors to teach modern methods.  They worked with the wine-growers to develop training courses.  They provided guarantees for loans so that farmers could buy new equipment.  They helped Moldovan farmers get access to new varieties of grapes… you get the idea.

(By the by, the wine project was not my project. But it was literally up the street from my project.  It was run by two people I know and deeply respect — one American, one Moldovan — so I had a ring-side seat for much of this.)

The big one was, they worked with the Moldovans on what we call market linkages.  That is, they helped them connect to buyers and distributors in Europe, and figure out ways to sell into the EU.  I say this was the big one, because on one hand the EU is the world’s largest market for wine!  But on the other hand, exporting wine into the EU is really hard.  There are a bunch of what we call NTBTs — “non-tariff barriers to trade”.  For starters, your wine has to be guaranteed clean and safe according to the EU’s very high standards.  That means it has to consistently pass a bunch of sanitary and health tests, and also your production methods have to be certified.  Then there are a bunch more requirements about bottling, labelling and packaging. 

The EU regulates the hell out of all that stuff.  Like, the “TAVA” number?  There’s a minimum font size for that.  If you print it too small, it’ll be bounced right back to you.  The glass of the bottle?   Has to be a sort that EU recycling systems can deal with.  The adhesive behind the label?  It can be rejected for being too weak (labels fall off) or too strong (recycling system can’t remove it).  There are dozens of things like that.

And then of course they had to do marketing.  Nobody in Europe had heard of Moldovan wines!  Buyers and distributors had to be talked into taking a chance on these new products.  This meant the Moldovan exporters needed lines of credit to stay afloat.  This in turn meant that Moldovan banks had to be talked into… you get the idea.

This whole effort took over a decade, from the early 2000s into the teens.

And in the end it was a huge damn success.  With USAID help, the Moldovan wine industry was completely restructured.  Moldova now exports about $150 million of wine per year, which is a lot for a small country — it’s over $50 per Moldovan.  And it went from exporting around 80% of its wine to Russia, to around 15%.  Most Moldovan wine (around 60%) now goes to the EU, with an increasing share going to Turkey and the Middle East.  



(If you’re curious: their market niche is medium to high end vins du table.  Not plonk, not fancy, just good midlist wines.  I can personally recommend the dryer reds, which are often much better than you’d expect at their price point.)

Russia tried the “ooh we found a sanitary problem” trick one last time a few years ago.  It fell completely flat.  Putting aside that it was an obvious lie — if something is safe for the EU, believe me, it is safe for Russia — Moldovan wine exporters had now diversified their markets to the point that losing Russian sales was merely a nuisance.  In fact, the attempt backfired: it encouraged the Moldovans to shift their exports even further away from Russia and towards the EU.

So that’s the political story.  Russia had Moldova on a choke chain.  Over a dozen years or so, USAID patiently filed through that chain and broke Moldova loose.  Soft power in action.  It worked.

Nobody knows this story outside Moldova, of course. 

Okay, that’s the political story.  Here’s the personal one.

Some years ago, I moved with my family to a small country that was recovering from some very unpleasant history.  They’d been under a brutal ethnically-based dictatorship for a while, and then there was a war.  So, this was a poor country where many things didn’t work very well.

While we were there, my son suddenly fell ill.  Very ill.  Later we found out it was the very rapid onset of a severe bacterial infection.  At the time all we knew was that in an hour or two he went from fine to running a super high fever and being unable to stand up. Basically he just… fell over. 

Wham, emergency room.  They diagnosed him correctly, thank God, and gave correct treatment: massive and ongoing doses of antibiotics.  But he couldn’t move — he was desperately weak and barely conscious — and there was no question of taking him out of the country.  We had to put him in the local hospital for a week, on an IV drip, until he was strong enough to come home.

If you’ve ever been in a hospital in a poor, post-war country… yeah at this point someone makes a dumb joke about the NHS or something.  No.  We’re talking regular blackouts, the electricity just randomly switching off.  Rusting equipment, crumbling concrete, cracked windows.  A dozen beds crammed into a room that should hold four or five. Everything worn and patched and held together with baling wire and hope.   

We’re talking so poor that the hospital didn’t have basic supplies.  Like, you would go into town and buy the kid’s medication, and then you’d also buy syringes for injections — because the hospital didn’t have syringes — and then you’d come back and give those thing to the nurse so that your kid could get his medication. 

In the pediatric ward, they were packing the kids in two to a bed. Because they didn’t have a lot of rooms, and they didn’t have a lot of beds. And kids are small, yeah?  

But there we were.  So into the hospital he went.  Here’s a photo:

— Take a moment and zoom in there.  Red-white-and-blue sticker, there on the bed?  It says “USAID:  From The American People”.

Every hospital bed in that emergency room had been donated by USAID.  I believe they were purchased secondhand in the United States, where they were old and obsolete.  But in this country… well, they didn’t have enough beds, and the beds that they had were fifty years old.  Except for those USAID beds.  Those were (relatively) modern, light and adjustable but sturdy, and easily mobile.  The hospital staff were using them to move kids around, and they were getting a lot of mileage from them.

And of course, every USAID bed had that sticker on it.  And so did some other stuff.  There was an oxygen system that a sick toddler was breathing from.  USAID sticker.  Couple of child-sized wheelchairs.  USAID stickers.  Secondhand American stuff — USAID was under orders to Buy American whenever possible — but just making a huge, huge difference here.

As I said, it was crowded in there.  Lots of beds, lots of kids, lots of anxious parents.  So we got to talking with the other parents, as one does.  A couple of people had a little English.  And so my wife mentioned that we were here working on a USAID project…

…and god damn that place lit up like an old time juke box.  “USAID!”  “USAID!”  People were pointing at the stickers, smiling.  “USAID!”   “America, very good!”  “Thank you!”  “USA!  USA!”  “Thank you!”

This went on longer than most of us would find comfortable.  When it finally settled down… actually, it never really did entirely settle down.  For the whole time our son was there, we had people — parents, nurses, even the hospital janitor — smiling at us and saying “USAID!”  “Very good!”  “Thank you!”

I’m not prone to fits of patriotic fervor.  But I’m not going to lie: right then it felt good to be American.

Anyway, USAID stories.  I could go on at considerable length.  This is my career, after all!  I could tell more stories, or comment and gloss at greater length on these.

But this is long enough already.  More some other time, perhaps.

This portrait of Bacchus contains the images of two lovers. Where are they?