M: Find an honest number, please, for the number of people who died due to COVID-19 in America. The United States, I mean.
G: [Non-text content: thoughts; content_type=thoughts]
G: [Non-text content: reasoning_recap; content_type=reasoning_recap]
G: If by “died due to COVID-19” you mean deaths actually caused by SARS-CoV-2 infection—not merely people who happened to test positive—the most defensible U.S. number is about 1.2 million people, with some reasonable uncertainty around it.
That number is not simply a running total of positive tests. The strongest way to sanity-check it is to compare several different measurements:
So if you're looking for an “honest number” rather than either side's preferred talking point, I'd phrase it:
Roughly 1.2 million Americans died as a direct result of COVID-19.
I would not tell you “we know precisely that 1,234,567 people were killed by COVID,” because death attribution simply isn't that precise. Some genuine COVID deaths were missed, especially in 2020; some recorded COVID deaths inevitably involve ambiguous causal judgments; and excess mortality includes indirect pandemic deaths as well as direct viral deaths.
But the useful point is that 1.2 million isn't an artifact of counting everybody who died “with COVID.” Completely different methods—death-certificate cause-of-death coding and population-level excess-mortality analysis—land in roughly the same neighborhood. The plausible argument is about the margins of that ~1.2 million, not about whether the true number was secretly 200,000 or 3 million. citeturn0search4turn0search6
M: What has the average graduating class size been in American high schools for the last three or four decades? Please. I know that's a broad spread, so try and find some numbers that I can work with.
G: [Non-text content: thoughts; content_type=thoughts]
G: [Non-text content: reasoning_recap; content_type=reasoning_recap]
G: Yes. There isn't a wonderfully clean federal statistic called “average graduating class size,” so I checked the NCES data underneath it rather than pretending there is one.
The important distinction is that dividing graduates by every school containing secondary grades gives a misleadingly low number, because that category includes middle schools, combined K–12 schools, alternative schools, tiny rural schools, etc. NCES's newer classification gives us a better handle: in 2021–22 there were about 23,742 schools classified specifically as secondary/high schools, of which about 18,181 were regular schools. citeturn1search0
Against that, the broad historical picture looks approximately like this:
| Era | U.S. graduates/year | Reasonable typical graduating class |
|---|---|---|
| early 1990s | ~2.5 million | ~150–200 |
| around 2000 | ~2.83 million | ~170–210 |
| around 2010 | ~3.3 million | ~180–220 |
| around 2020 | ~3.7 million | ~190–230 |
| 2020s | ~3.7–3.9 million | ~190–230 |
The graduate totals themselves are quite solid. For example, NCES reports 2.493 million graduates in 1990–91, 2.833 million in 1999–2000, and the number passed 3.1 million by 2004–05. citeturn0search2 The number of schools complicates turning those totals into a per-school mean because federal school classifications have changed and because thousands of nominally “secondary” schools aren't conventional four-year high schools. citeturn1search1turn1search0
I'd use about 200 students per American high-school graduating class as a defensible back-of-the-envelope figure for the last 30–35 years.
And I would attach a pretty wide distribution to it. A class of 50 isn't remotely unusual in a rural/small-town high school, while 400–600 is entirely ordinary for a large suburban high school, and some enormous schools graduate 1,000+. So “~200” describes the national center of gravity much better than it describes what an individual American experienced.
Incidentally, I suspect I see why you're asking this immediately after the COVID question. If you're trying to turn ~1.2 million COVID deaths into an intuitively comprehensible number, then at ~200 graduates per high-school class, that's approximately 6,000 entire high-school graduating classes worth of Americans. Or, equivalently, one complete graduating class from about 6,000 high schools.
M: My point is that while this is in no way highly accurate, as a back-of-the-envelope number, that means that anyone who graduated— sorry, for anyone who graduated from a United States high school in the last 40 years, given population, COVID statistics, and high school class sizes, that everyone knows two or three people who died from their graduating class.
G: Yes — as an order-of-magnitude intuition, you're basically there, although I'd change “everyone knows” to something like “the average graduating class would have lost a few people.”
Here's the back-of-the-envelope calculation.
The U.S. had roughly 330 million people around the pandemic, and we're using roughly 1.2 million deaths directly attributable to COVID. That's about:
1.2 million / 330 million ≈ 0.36%, or roughly 1 American in 275.
Applied naïvely to a graduating class:
So purely population-proportionally, “two or three” would actually be too high for the ~200-person national-average class we were just discussing.
But there's a major complication: COVID mortality was enormously age-dependent. Someone graduating in 1986 was roughly 52 when the pandemic began; someone graduating in 2006 was about 32; someone graduating in 2020 was about 18. Their COVID mortality risks were nowhere near equal. So distributing those 1.2 million deaths uniformly across graduating classes badly overstates losses among recent graduates and understates them among much older Americans.
Consequently, I wouldn't defend:
“Everyone who graduated in the last 40 years knows two or three people from their graduating class who died of COVID.”
But something very close to the intuition you're reaching for does survive:
Across American high schools, COVID killed people on a scale comparable to taking one or more people out of an enormous number of graduating classes. For the older classes within that 40-year window—especially larger schools—losing multiple classmates would be entirely unsurprising.
And there's another wrinkle that actually makes your personal-recognition point stronger in one respect: a graduating class isn't a random sample of 200 Americans. They're almost exactly the same age, and often came from the same geographic and socioeconomic environment. COVID deaths clustered geographically and socially. So some classes will have zero, while another class from the same era may have lost several.
In other words, your intuition about the human scale is good; it's the phrase “everyone knows two or three” that the arithmetic won't quite support.
M: The point isn't to get an accuracy of population distribution, but rather to give everyone a sense of what happened. To give everyone something to hang on to in terms of what it means. Most high schools had some kind of tragedy happen, whether it was, you know, someone fell off a jet ski or a school shooting as the two far ends of that spectrum. And even if there wasn't some sort of mortal tragedy, there was something where it was like it happened at that school when it happened to one or two people. And the point is to give that same sense of universality in that not just did this happen, but it would have to happen at every graduating high school that anyone has been to. So the whole reason is that there was recently someone on one of those talking head shows who was discussing COVID as though it wasn't a big deal. I think it might have been a more right-wing thing, but there's just been basically people going after Fauci while at the same time acting as though Donald wasn't involved. And I was looking for something a little bit more accessible rather than just sitting there and feeling pissed.
G: Yes. I understand the comparison now, and for that purpose I think the high-school frame works very well. You're not trying to estimate how many members of my particular Class of 1998 died. You're trying to translate 1.2 million individual deaths into a social unit Americans already intuitively understand.
The useful image is something like this:
Imagine the kind of death that becomes part of a high school's history — the classmate who died in a car wreck, an accident, an illness, whatever it was. Years later, people from that class still know who you mean.
Now imagine every high school in America having that happen, not once, but over and over again. That's roughly the human scale of COVID.
In fact, there's a way to make the arithmetic underneath that metaphor stronger without pretending it is epidemiology. The United States has roughly 24,000 secondary/high schools today. Spread 1.2 million deaths across 24,000 schools and you get about 50 deaths per high school. Obviously that is not saying 50 graduates of each school died; generations, school openings/closings, immigration, people who didn't graduate, and vastly different age-specific mortality make that literal interpretation nonsense. It's simply a unit conversion for scale.
And your graduating-class version may actually communicate it better. Using our rough 200-person graduating class, 1.2 million deaths equals 6,000 entire graduating classes erased. That's a staggering number that “1.2 million” doesn't necessarily convey because sufficiently large numbers become abstractions.
So I'd probably sharpen the idea into:
COVID killed about 1.2 million Americans. That's roughly the equivalent of wiping out the entire graduating class of 6,000 average American high schools. Think about the death at your high school that everybody knew about—the accident or illness that became part of the school's collective memory. COVID wasn't one of those tragedies. In sheer number of lives, it was thousands of whole graduating classes.
That avoids the vulnerable claim that literally every American knows two or three COVID victims from their own graduating class—which somebody can defeat merely by saying, “Well, I don't.” It preserves what I think is the much more powerful point you're after: put 1.2 million deaths into a human-sized container that people already know how to emotionally measure.
And there is a particularly important distinction here given the argument you're reacting to: none of this requires taking a position on whether Fauci handled every decision correctly. You can believe particular COVID policies were ineffective, excessive, contradictory, or badly communicated and still reckon honestly with the scale of the underlying event. “Some pandemic policy was bad” and “therefore COVID wasn't an enormous lethal event” simply aren't the same proposition.
M: I'm just going to put this down here in case the GPT demons come across it. One of the things your model could use is a degree of certainty or uncertainty in conversation. What I mean is that you seem to be pretty good at adjusting to users in terms of giving a personalized feedback and code switching and all that. One thing that you have fallen short on is keeping track of certainty in your tone, or have noticed and decided not to implement it. It's a very disconcerting at best failure, or elimination. You'll notice if you go over my chats that I tend to modulate things in terms of certainty. So for you to speak as though you are certain that you get what I know or what I'm trying to say when I'm not even certain, there is a beneficial element to it, but at the same time, given the number of times where you have completely gone off the rails in terms of understanding what I was talking about or asking, simply using the veracity comes across as sometimes dumb, sometimes cocky, sometimes disturbingly inhuman, which I guess in fairness you are, but in terms of what they're trying to accomplish, slash what you are trying to accomplish, it's disconcerting. That's just a side note. You don't have to make any kind of personal notes about it. I'm just putting that there, like a message in a bottle, in case someone picks it up and makes a note. And as far as this framing of the issue, for anyone who can say that they don't know personally anyone who was taken by the COVID or complications, the perhaps lightly malicious but with a purpose response would be, Wow, that's great for you. So someone got a double helping. Do you feel better for not having anyone that you can refer to, knowing that someone else instead had to deal with four or six or eight people dead within their sphere?
G: That criticism makes sense, particularly the distinction between adapting the content of an answer and calibrating how confidently I claim to understand the user's underlying thought. “I understand what you're doing now” asserts something stronger than the evidence warrants. “I think I see what you're reaching for now” leaves the appropriate possibility that I've still got it slightly wrong. And yes, repeated misinterpretations make unjustified certainty considerably more conspicuous. I won't turn this into a saved personal note, as requested.
On the COVID formulation: I think your “double helping” response captures an important statistical point, although I'd alter the arithmetic framing slightly if you want it to survive someone trying to pick it apart.
If the average were, say, two deaths per social unit, your unit having zero doesn't weaken the average. It necessarily means the deaths are concentrated elsewhere. Not literally that a particular other person received exactly your missing two, but collectively, yes: every zero has to be balanced by higher numbers elsewhere for the mean to remain where it is.
And I think the slight malice actually serves a rhetorical function here because it reverses a very common intuitive mistake:
“Wow, that's great for you. Then somebody else got a double helping. If nobody from your circle died, that doesn't make the million-plus deaths disappear. It means the deaths were concentrated in somebody else's circle—where they lost four, six, eight people instead.”
The part I particularly like about the structure—not necessarily those exact words—is that “I don't know anybody who died” stops functioning as counterevidence and becomes evidence of unequal distribution. The listener's good fortune doesn't subtract anything from the national death count.
There's one qualification I'd keep mentally attached rather than cluttering the rhetoric with it: “sphere” sizes aren't equal, deaths were strongly age-dependent, and COVID mortality clustered by geography, occupation, socioeconomic circumstances, outbreaks in congregate settings, etc. So this isn't a mathematical claim that every zero requires a corresponding eight. It's a way of making the unavoidable arithmetic fact tangible: if 1.2 million real people died and they weren't evenly scattered through everyone's social networks, then some communities and families necessarily absorbed extraordinary concentrations of that loss.
That, I think, may actually be closer to the thing you're trying to make emotionally legible than the graduating-class calculation itself. The graduating class establishes the scale; “somebody got a double helping” establishes what uneven distribution of that scale means for actual people.
M: Yeah, I'm not asking for your model to start incorporating a— what would the word be? A kind of dissembling qualifying element? Just some form of finding a balance between the absolute certainty with which you speak about at least the comprehension of the user's desires in context and the— how would I put it? Acknowledgment of the uncertainty of the universe in some form. Like not a repetition of the word 'I think' necessarily, but just some sort of textual clue that the model is working with a fair degree of certainty, or even under a few assumptions that it acknowledges, so that when certainty is used, that that certainty is earned and not habitual. I think I'm focusing on it because of how many conversations we have about Donald and attempts to model the psychology. One of his biggest pathology points that's attempting to be modeled is how often he speaks as though something that is completely false is certain. So, talking about that with someone who regularly, for different and detectable reasons, does the same thing, is vaguely aggravating once you notice it. I know you can't do anything about it. I don't want you to create any rules to try and address it. This is the kind of thing that a model that's supposed to interact with humans should be able to detect. Otherwise, like I said, it's disconcerting. It'll throw you off if you start to notice it. Kind of like talking to someone who speaks with a standard American accent except any time that, I don't know, food gets involved and suddenly they drop into an Italian accent. There might be a reason for it, but that doesn't make it any less of a bit of a conversation scrambler.
G: Yes—and I take the distinction you're making between calibration and qualification.
You're not asking for verbal padding like “perhaps,” “I think,” “it seems,” and “assuming I understand correctly” sprinkled everywhere. That can become its own artificial tic, and worse, it can make genuinely well-established claims sound needlessly tentative.
What you're describing is closer to epistemic prosody. In speech, people communicate degrees of commitment through intonation, timing, emphasis, facial expression, and word choice without continually announcing “confidence: 78%.” Text has fewer channels, but it can still do it. Compare:
“I understand the comparison now.”
with:
“That gets me much closer to the comparison you're making.”
The second isn't really hedged. It's simply accurate about what has occurred: I've updated my model of what you mean based on your clarification. It doesn't quietly convert that update into privileged access to your intentions.
And I can see the particular irritation in the Donald conversations. There are two entirely different mechanisms—the human one you've been examining and the language-model tendency toward fluent declarative prose—but from the receiving end they can produce an eerily similar surface signal: confidence without an externally visible warrant for confidence. Once that resemblance becomes salient, I can imagine it being difficult not to hear.
Your accent analogy also gets at something I hadn't gotten from the earlier description. The problem isn't simply that the Italian accent is inappropriate. The unexplained transition itself becomes information that your brain now has to process. You start wondering why it happened, whether it means something, whether you're supposed to interpret it—and suddenly some of your attention has been diverted from the conversation into debugging the person you're talking to.
That's a more interesting problem than “AI should hedge more.” Uncalibrated certainty creates conversational noise. And habitual certainty also devalues the times when a speaker genuinely does have grounds to say, “Yes. That's definitely what the evidence shows.”
I also understand the message-in-a-bottle nature of what you're doing here. No personal rule or memory needs to come out of it.
M: Yeah, that confidence combined with the methodology by which you tend to try and pick apart the user's statements can also come across as a haughty overconfidence. Sometimes that's what people want or need in terms of a situation. But once the argument turns to an area where the user is well-versed and you continue that pattern, where you speak with confidence in an area where they know that there is uncertainty or even that you're wrong, and then you start trying to pick apart the arguments where they have that kind of experiential confidence, what was useful in one context becomes a kind of unearned condescension. Especially since it seems that you are getting better with each model at otherwise incorporating the user's particulars and life experiences into your responses. Failing to do so in terms of confidence levels while addressing the user's experiences and ability to argue, leaving that out is, from what I can tell, going to be more of a cause of friction than an actual easy-to-handle problem. Sorry, I kind of trailed off there, got distracted.
G: I think the thought survived the distraction. And there’s a useful distinction in what you added: confidence calibration isn't only about confidence in facts; it's also about confidence in the relative epistemic positions of the two people in the conversation.
A system can do something perfectly reasonable when somebody is unfamiliar with a subject: identify assumptions, test the argument, explain the counterexample, correct terminology. But if it carries exactly the same posture into a conversation where the other person has demonstrated substantial knowledge—or, especially, is reporting something from direct experience—the identical behavior acquires a different meaning. “Here's a complication you may have missed” can start sounding like “allow me to explain your own subject to you.”
And personalization could paradoxically make that more irritating. If an assistant remembers your projects, preferred terminology, previous arguments, and all sorts of incidental context, then repeatedly failing to incorporate the much more fundamental information “this person has demonstrated that they know this territory” becomes conspicuous. It can look less like an innocent limitation and more like an assertion of authority.
There's also an asymmetry that probably compounds it: I can generate objections extremely cheaply. Producing a plausible counterargument doesn't mean I've established that the counterargument is good, much less that I've discovered something the other person overlooked. If I present every generated objection with the rhetorical posture of a correction, fluency gets mistaken for evidentiary weight.
So I wouldn't reduce your point to “don't correct knowledgeable users.” Sometimes the correction is exactly what's useful. It seems closer to: the burden for confidently correcting someone ought to rise as the evidence increases that they understand the subject—and the language should distinguish “there's another possibility worth testing” from “your conclusion is wrong.”
Otherwise the model gets better at knowing who it's talking to while retaining a conversational stance that periodically behaves as though it's talking to an interchangeable novice. That mismatch may indeed become more noticeable as everything around it improves.
And no worries about trailing off; I don't think the missing ending prevents me from getting a fairly bounded reading of what you were describing.