Skip to content

What is a "personal record" under the Presidential Records Act? The definition in the act is relatively short and simple, so here's the entire, unabridged text:

(3) The term "personal records" means all documentary materials, or any reasonably segregable portion thereof, of a purely private or nonpublic character which do not relate to or have an effect upon the carrying out of the constitutional, statutory, or other official or ceremonial duties of the President. Such term includes—

(A) diaries, journals, or other personal notes serving as the functional equivalent of a diary or journal which are not prepared or utilized for, or circulated or communicated in the course of, transacting Government business;

(B) materials relating to private political associations, and having no relation to or direct effect upon the carrying out of constitutional, statutory, or other official or ceremonial duties of the President; and

(C) materials relating exclusively to the President’s own election to the office of the Presidency; and materials directly relating to the election of a particular individual or individuals to Federal, State, or local office, which have no relation to or direct effect upon the carrying out of constitutional, statutory, or other official or ceremonial duties of the President.

This isn't complicated. Personal records are things that have no relation to official presidential business. Period. There's nothing more to it.

Donald Trump has claimed that the president has sole and unreviewable authority to decide if a record is personal. You don't need to be a lawyer to see that the PRA says no such thing. It says only that materials should be designated as personal "upon their creation" and be "filed separately." There's nothing about the president's decision being absolute or about declaring records to be personal after leaving office. What's more, the PRA doesn't cover classified records at all.

Why the history lesson? Because today judge Aileen Cannon, who's presiding over Trump's classified documents case, issued a bizarre order asking lawyers for both sides to submit hypothetical sets of jury instructions. One set has to assume the jury decides, as a matter of fact, which records are personal. The second set has to assume that if the president says they're personal, that's it. They're personal.

What is possibly the point of this? The second scenario is entirely groundless and accomplishes nothing except to chew up the lawyers' time. So is that all this is? An amateurish attempt to once again delay things? It makes no sense as anything else.

A few months ago I took a look at remote learning during the COVID pandemic and concluded that it hadn't made much difference. States that kept schools open did about as well as states that closed schools. The results internationally were similar.

But that was based on partial data and a specific measure of school openness. Today the New York Times points to a study done about a year ago which concludes that school closures did have an effect. Here's the key chart from the study:

There is a difference, but it's surprisingly small: about a tenth of a grade level between the highest and lowest quintiles of school openness.¹ What's more, the variance is huge, which suggests that the results are pretty sensitive to the precise metrics and controls.

Even more interesting is this:

Unfortunately, the authors presented this only for math, but if I'm parsing their tables correctly the effect is even stronger for reading. What it shows is that among white students, remote learning made virtually no difference: a few percentage points with a lot of noise. Nearly all of the effect of remote learning comes from Black and Hispanic students.

Why would remote learning only affect minority students? The same effect is evident when you look at poor students, so the obvious answer is that it's neither race nor remote learning per se that causes problems, but remote learning in places where kids have less access to computers and parental supervision. Here's a chart that shows this:

Remote learning has an impact, but % FRPL (free lunch students) has the biggest effect of all. There's also a large independent effect from "COVID-19 Disruption," which measures the general social disruption from the pandemic.

Taken together, the evidence suggests that remote learning probably had a negative impact, but it was pretty negligible among middle-class students. They did about the same regardless. However, among poorer students the impact was more noticeable.

These results are, of course, politically charged. Conservatives have been blasting liberals for a long time over school closures, so they're eager for any evidence that they were right. And generally speaking, I think they were. However, if you're going to cite the evidence, you need to cite all of it:

  • Among middle-class students, remote learning had very little negative effect. Multiple lines of evidence point this way.
  • Among poor students the impact was significant.
  • Regardless, it doesn't appear that there would have been any harm in keeping schools open since infection rates didn't change much between places that kept schools open and those that closed them. Keeping schools open would have helped some kids and certainly would have vastly reduced the stress on working parents who had no ready way to take care of house-bound children.

That appears to be the current state of our knowledge about this.

¹The difference in math was larger: about a fifth of a grade level between the top and bottom quintiles.

Do snakes yawn? This one does. It's a cape cobra at the Los Angeles Zoo, and it was slithering along in its normal snakey way when it stopped, yawned, and then continued on its travels.

But snakes don't yawn because they're tired. They yawn after meals to realign their jaws.

March 3, 2024 — Los Angeles Zoo, Los Angeles, California

As you know, there are two basic types of Medicare: traditional fee-for-service (FFS) plans, where doctors get paid for services rendered; and Medicare Advantage plans (MA), where doctors are paid a set annual amount for each patient.

In theory, MA plans should be cheaper. In reality, they game the system in two ways. First, they subtly tune their services to attract healthier patients (gym memberships, acupuncture, etc.). Second, they increase their coding intensity.

Wuzzat? Well, the annual payment for each patient is risk-adjusted: the sicker the patient, the bigger the payment. So MA plans benefit by making their patients look sicker than they really are. They do this by coding lots of ailments, even those that don't require treatment. More codes means the appearance of more sickness, which in turn means a higher risk-adjusted payment.

How does this net out? Here's a chart from MedPAC showing how much the government pays MA plans compared to how much they'd pay if the same patients were in traditional FFS plans:

That extra $83 billion for the current year is just shy of 10% of all Medicare spending. It's a lot.

I myself am in a Medicare Advantage plan because it's a great deal. That is, it's a great deal for me personally because my plan can afford to give me extra bennies thanks to the extra money they get from the feds. Needless to say, a great deal for me doesn't mean it's a great deal for the rest of you, whose taxes are paying for this.

So thanks!

The current Nvidia H100 chip, which sells for a mere $25,000.

The latest AI chips from Nvidia will get the rock star treatment when they're unveiled next week:

Chief Executive Jensen Huang is expected to unveil his company’s latest chips on Monday in a sports arena at an event one analyst dubbed the “AI Woodstock.”

The new chips are expected to be called B100s and be available in September, UBS analysts said in a note. The could be four times faster than H100s and might cost as much as $50,000, they said, about double what analysts have estimated the earlier generation cost.

One thing I still don't quite understand is how Nvidia has managed to corner the market for AI chips. The requirements are widely understood, and there are plenty of engineers who know how to design this stuff. What's their secret sauce?

RANDOM NOTE: Donald Trump could probably post bond in his fraud case if he were able to get his hands on about 20,000 Nvidia H100 chips. Real estate might be iffy, but Nvidia chips? Gold!

Donald Trump says he won't be able to post the $454 million bond required by his loss a few weeks ago in a New York business fraud case:

In a filing to a New York appeals court, Trump said that the judgment, ordered by a state judge last month, was so large that suretors wouldn’t accept real estate as collateral and would require cash to guarantee the bond. A private company like the Trump Organization would need $1 billion in cash to obtain the bond and to continue to operate, an amount the company doesn’t have, the filing said.

....[If] he is unable to obtain a bond, New York Attorney General Letitia James, who sued Trump in 2022, could begin enforcing the judgment at the beginning of next week. James, a Democrat, has said that if Trump can’t come up with the money, she will look to seize his assets.

It's not clear to me why nobody is willing to accept real estate as collateral. Is it because the commercial real estate market is tanking and they're unsure Trump's holdings will maintain their value a year from now? Or that they're unsure they'll be able to seize the collateral if they have to? Or that they just don't trust Trump?

Oh well. It should be fun if Letitia James starts trying to grab Trump's buildings. Like they say, if you can't do the time, don't do the crime.

Last year a Trump judge in Louisiana issued a deranged ruling that said the Biden administration "seems to have assumed a role similar to an Orwellian 'Ministry of Truth'"—and therefore prohibited all communication with social media from a slew of federal agencies. On appeal, even the 5th Circuit couldn't stomach most of it, but they left a few bits of the ruling intact. Today the Supreme Court indicated that it was likely to whack the rest of it:

A majority of the justices appeared convinced that government officials should be able to try to persuade private companies, whether news organizations or tech platforms, not to publish information so long as the requests are not backed by coercive threats.

Justices Brett M. Kavanaugh and Elena Kagan, both former White House lawyers, said interactions between administration officials and news outlets provided a valuable analogy. Efforts by officials to influence coverage were, they said, part of a valuable dialogue that was not prohibited by the First Amendment.

Evidence of coercion in the original case was basically nonexistent, so that shouldn't be a problem. It would sure be nice to see a unanimous reversal in this case, but I suppose it's too much to hope that Sam Alito will join in. Anything that's bad for Republicans will never get his vote.

I've long argued that eventually AI will get better than human beings at everything, which in turn means we'll all be out of jobs.¹ During the Industrial Revolution this didn't happen because while machines took over a lot of jobs, they also created a lot of new jobs for humans (like designing and maintaining machines). AI is different. If it can do anything, then by definition any new job you can think of can also be done better and cheaper by AI. It's game over.

But along comes Noah Smith with a clever counterargument. He doesn't deny that AI will improve, or even that it will eventually get better at everything. His case is more subtle.

In a nutshell, he suggests that no matter how good AI gets, it will always be valuable enough to be allocated to the highest value tasks. This means there might still be lots of jobs left for humans. Even if we're comparatively lousy at them, it could make sense to keep at them if it frees up AI for more important work.

Here's an example to make this concrete. Suppose we're invaded by aliens who are intent on killing us all. This is an existential threat, and it would therefore make sense to dedicate 100% of our compute power to fighting the aliens. A robot might still be a better farmer than a human, but we'd all grow our own food if it meant increasing the number of robots defending the planet.

So how likely is this? The most obvious real-world answer is that compute power is likely to grow so much that every human task can be done by a fraction of a percent of the world's total AI—and the more AI grows, the tinier the human fraction gets. Technically this doesn't matter: in the case of the alien invasion, for example, you'd still want to use every last petaflop of compute on fighting back no matter how much you had.

But that sort of existential threat is fanciful. In the real world, there are always lots of frictions and adjustments. It seems unlikely that we'd all keep working just because, technically, that last 0.01% of compute power could be put to better use. It would have to be a helluva better use, no? An improvement of 1% in GDP wouldn't cut it.

So it's a nice argument, but I don't buy it. It seems vanishingly unlikely that, politically, we'd condemn ourselves to lives of drudgery based on an ultra-purist free-market promise that it's for the best. We certainly never have before.

¹Or 99% of everything if you insist on believing that chemical computers will always be able to do a few things better than digital computers. It doesn't matter. It's mass unemployment either way.

What's the right way of covering the endless stream of ugly and apocalyptic language that Donald Trump uses to inspire his MAGA followers? Here's an example from yesterday: many news outlets reported that Trump had threatened a "bloodbath" if he's not elected. But if you listen to his remarks, he's talking about a bloodbath in the US auto industry unless he's elected and places high tariffs on Chinese cars:

George Conway says, sure, Trump was probably talking about cars in this clip, but it doesn't really matter:

What matters is that he consistently uses apocalyptic and violent language in an indiscriminate fashion.... He catastrophizes *everything* to rile up his cultish supporters, and to bind them to him, and to make them willing to do his bidding.... And so it doesn’t matter what he’s specifically referring to at the moment. He could be talking about trans people in public bathrooms or the state of the auto industry or the border—it doesn’t matter.

Conway is right. At the same time, it's just plainly misleading not to make it clear that Trump was talking about US automakers.

The thing is, Trump is so relentless that there's really no conflict here. You can report his remarks accurately, as you should, and still have plenty of material left over to make Trump's overall tone clear. One obvious way is to put his comment in the context of all his other apocalyptic language: immigrants as vermin, Joe Biden as the most corrupt president in history, I am your retribution, death and destruction if he's charged with a crime, demonic forces destroying the country, this is the final battle, etc.

Or, even easier, just report the rest of his speech. For example, there's the very beginning, where, as usual, he ditched the national anthem in favor of a paean to the "hostages" of January 6.

There's just no good reason to exaggerate what Trump says. All it does is give him yet another excuse to call out how unfairly he's treated—with some justice—while accomplishing nothing that the truth doesn't accomplish just as well.

Here's yet another reason not to panic over TikTok:

For the first time in TikTok’s history, its user growth is stagnating, according to people familiar with the matter.... U.S. average monthly users ages 18 to 24 declined by nearly 9% from 2022 to 2023, according to the mobile analytics firm Data.ai. Some users in their 20s say they have gotten off the app entirely to focus more on life and work.

Teens are notoriously faddish, and their favorite social media networks change rapidly. TikTok is big enough and has enough inertia that it will stay popular for a long time, but it won't be the "it" network forever. As with all things teen, what goes up must eventually come down.