The last podcast standing is a stand-in moniker for the podcast discussed in this post. If there is a podcast with this name, be aware that it is not the subject here.

One by one, they all got dropped. Some to conspiracy garbage, some to compulsory repudiation, most to the conviction that everybody else is stupid. I didn’t mourn many of them. But I kept one. The last man standing.

This week he crossed over. Not into criticism of AI, not into critique, not into anything with a pulse of thought. Into the riff. Two guys, back and forth, about how helpful it is. How it helps. How helpful. I sat there and felt something I can only call defeat.

The roster, collapsed to a single edge

Dashed grey edges are severed. The salmon edge is the last one left, and it now leads somewhere else.

Understand what the loss is. He didn’t surrender to the machine. He surrendered to the banality of talking about the machine. I came for a particular mind: the observations, the temperament, the pleasure of hearing a man make something out of his own experience. What I got was two people who had discovered that a tool is useful and mistook the discovery for material.

The worst part is that there’s barely anything to argue with. AI is helpful. Fine. Granted. That’s the entire conversation. It begins and ends right there, and they kept talking anyway. It wasn’t even a bad episode in the ordinary sense. A bad episode you forgive. This was the disappearance of the reason I listened in the first place.

Had they shown me AI enriching somebody’s life, or even being right about something, I’d have shrugged and moved on. What they served up instead was indoctrination. Not the kind with flags and anthems. The quiet kind, where you learn to accept answers without asking how they were made.

Here’s the scene. One of them says he’s been seeing posts on X about starvation in Somalia. Good instinct: he doesn’t just swallow it. He checks. He asks Grok. And Grok hands him a figure about “food insecurity.” He asks what that means. He gets a definition. And then, and this is the part where I put my head in my hands, he stops.

That’s it. Case closed. The posts were weighed, an authority was consulted, the matter was settled.

Except nothing was settled. To know what a designation means, you have to know how it gets assigned. By what mechanism. Who counted, what they measured, where the line falls between one category and the next. Without that you don’t have an answer. You have authority saying so, and a man nodding along.

The path walked, and the path abandoned

And here’s the insidious part: the audience hears what sounds like verification. He didn’t believe social media; he checked. That little gesture lends the conclusion a pedigree it never earned.

The performance of inquiry stood in for inquiry itself, on a show I used to trust to know the difference.

Let’s do the work he didn’t.

Food insecurity is not starvation. The words are not synonyms, and a count of one is not evidence of the other. The FAO’s Food Insecurity Experience Scale, for instance, rests on eight self-reported survey questions about difficulty getting food: worrying about running out, skipping a meal, eating less than you should. Somalia’s famine warnings come from a different machine entirely, the IPC, which stitches together consumption, livelihoods, malnutrition, mortality and expert judgment into phases. Different instruments, different questions, different answers. A figure from one cannot be waved around as a verdict on the other.

Two machines that share no inputs

So the first obligation was plain. Which metric did Grok hand over? How is it produced? Does it bear on the claim that people are starving? Three questions. He asked none of them.

To be fair to the machine, and I’ll be fair exactly once: Grok didn’t invent the metric. The silliness of “food insecurity” as a public talking point belongs to the institutions and the press that repeat it. But notice what the machine did do. A question about starvation went in, and a statistic about something else came out. It chose the instrument. It answered a different question while appearing to answer yours. Legitimacy is not relevance, and the very tool being celebrated for its usefulness was the one that slipped the substitute authority into the room.

All it would have taken to catch it was one more question. How do they know?

Then came the second act, the one where I would have wept if I were built for it. His co-host starts bragging that AI can do Seinfeld.

AI can do a Seinfeld bit. It can do a Don Rickles joke, too. Both men work from clear, consistent formulas, and formulas are exactly what these machines eat. Seinfeld takes a small social convention, finds the inconsistency in it, and follows the consequences until they turn absurd. Rickles picks a target, crosses a line, inflates the insult, and keeps it warm with timing and affection. George rationalizes. Jerry observes. Elaine objects. Kramer comes through the door with total certainty about something insane. That’s machinery, and machinery can be copied.

The formula, abstracted from its source

So yes, it’s doing a taste. An impression, a sketch. It’s even become a genre: YouTube is full of Seinfeld transplanted into Dragon Ball Z and whatever else is lying around. The bit the podcast was admiring took on the recent Nvidia fiasco, except it called it OpenAI. Nobody on the mic noticed. They were too busy marveling at the fidelity of an imitation that couldn’t keep straight which company the joke was about. That was the most Seinfeld thing in the whole episode, and it wasn’t in the bit.

A taste is not an episode. An episode has to develop a premise, find consequences you didn’t see coming, crash three stories into one another and land a payoff that justifies the trip. That’s creation. What the machine hands you is a photocopy, sometimes a very good one, good enough that people forget there’s a difference. But the formula was never the accomplishment. The accomplishment was knowing what deserved the treatment, how far to push it, and when to stop. Seinfeld and Rickles weren’t running the formula. They were the source it was abstracted from.

Producing something that sounds like Seinfeld isn’t a feat for a language model. It’s the definition of one.

The defenders have a ready answer: it makes things that never existed before. True, in the narrowest sense. Shuffle a deck and you’ll deal a hand nobody has ever held. That isn’t invention. That’s arithmetic.

When I say AI can’t create, I don’t mean it can’t produce a new arrangement. I mean it can’t produce something new: a thing with no template behind it, an invention of form rather than one more instance of a form.

A new hand is the same graph, relabeled

Same five vertices, same five edges. The first two are one graph drawn twice: 1 connects to 2, 2 to 3, all the way round. Only the third has a different structure.

Ask a machine for a logo and you’ll get a handsome one. Good proportions, decent type, maybe even something you haven’t seen. You will not get Paul Rand. Rand didn’t become Rand by making handsome logos. He changed what people understood a logo could be. There’s a world between making an object that didn’t exist and introducing an idea that didn’t exist, and the Seinfeld demo never reaches the starting line.

Same with the math. OpenAI has made plenty of noise this year about models producing new mathematical arguments. Fine. Look at who’s standing behind the curtain. A mathematician picked the problem, chose the assumptions, designed the line of attack, built the scaffolding, defined what counts as a proof, and then checked and formalized whatever came out. The machine may well have found a step nobody expected, and that can be a real computational achievement. It is not authorship of the investigation. The more carefully a human engineers the path, the less sense it makes to call the destination the machine’s discovery. Even the people building tests for this stuff draw the line where I do: proving a proposition someone hands you is one thing, dreaming up the conjecture worth proving is another, and they admit they don’t yet know how to test for the second.

Who is standing behind the curtain

Why did it produce those arguments? Was it bored? Did a loose thread nag at it in the shower? Did it grow dissatisfied with the accepted explanation and lie awake? No. It was set up to. Assigned a problem, handed an objective, run inside a system built to pursue it. What’s missing isn’t horsepower. It’s the impulse, the itch that sends a mind looking in the first place.

Once I built a web page for a lone Albanian. Somewhere in the process I flipped an image. Horizontal. One click, maybe two. He lost his mind. He looked at me like I’d bent a spoon with my eyes. It was the most impressive thing he had ever seen.

It was sad. Not because he was impressed; astonishment is a fine thing and nobody should be ashamed of it. It was sad because he had no basis for judging what deserved his amazement. The distance between the effort and the awe was enormous, and he couldn’t see it.

Edge thickness is weight. In both cases the input is a click and the output is reverence. The difference is who announces it.

That’s the same distance I heard on the podcast, with one difference. The Albanian was just a man surprised by something unfamiliar. No harm done. These two were presenting their astonishment as informed judgment, telling an audience what was impressive without ever showing they could tell a good answer from a persuasive one. In the Grok segment, consulting the machine became evidence of having investigated. In the Seinfeld segment, imitation became proof of creative power. Both times the bar for achievement was set by the people least equipped to measure it.

If there were accuracy, if there were rigor, I could ignore it. Let them have their toy. But there isn’t, and I can’t.

Here’s the knife. A few months back, this same podcast played a clip of two guys earnestly explaining that they treat people based on who they are on the inside. My man watched it and named it exactly: two dogs having a conversation. It was hilarious. A handful of words, and the whole empty, self-congratulatory exchange collapsed into its true intellectual content: two creatures affirming something neither had bothered to examine.

He has the wit. He proved it. He can spot the absurdity from across the room.

He just didn’t recognize himself in it.

Partway through the AI episode I caught myself. Stopped the thing and asked, out loud, what moronic bullshit back-and-forth am I watching? I already knew the answer, because he’d given it to me months ago. I recognized the conversation before either of them did. The man who diagnosed the dogs had become one of the dogs. Not because he likes AI, but because the moment the subject changed, he set down the very discrimination that made him worth hearing.

From observer to vertex

There’s a special grief in watching someone demonstrate a faculty of judgment and then decline to use it. It’s worse than never having had it. The ones who never had it, you skip. The ones who had it, you mourn.

So here is my review of the last podcast standing, as economical as anything he ever said:

Fuck this shit.