Would you prefer the Claude-assisted version of this essay? Or the one that's wrong?

or, How to corral the truth with AI

It started when I saw this cool NASA animation of ocean currents around the world, which is just mesmerizing. Made in a visual style that bears a startling, impossible-to-miss similarity to Van Gogh’s starry night, the ECCO2 animation makes the scale and complexity of ocean currents immediately evident in the way no verbal explanation ever could.

So I saw that and thought, “yeah that’s awesome I definitely want to write a post about that.” And I did indeed write a post about it, and thought it came out pretty good, and was about to press play when that little inner self-preservation voice editors developed kicked in and told me “you should just check this for accuracy: this is not your field.”

Now, I could, of course, hassle hassle a fluid dynamicist to check this text I’d written for accuracy, and that would’ve taken a week and pulled his attention from other, much more relevant tasks that fluid dynamicists presumably spend their time on. And it would have gotten me further into giri debt* with that fluid dynamicist, who would one day ask for a favor in return, and rightly so.

But come on, it’s 2026, I’m obviously not going to ask a human, that would be lunatic behaviour. I went to Claude, and asked for a fact check.

So Fable goes off and does it thing and comes back a few minutes later to inform me I’ve sullied its inards with an intellectual turd. The thing I’d written, the robot very politely but unequivocally informed me, was —and this is a technical term— donkey garbage: vague on some points, but then just straight-up wrong on others, and worst of all wholly silent on key advances in the field. Bad enough that I’m not going to release it, because it’s not worth reading.

So, deep breath, I start in on fixing it with Claude. I inevitably get into a long back and forth with the robot to try to reconceive how a post like this could be written without embarrassing me in the eyes of professional scientists.

This back and forth is intense: it feels like a collaboration and I learn a fuckton doing it. What comes out at the other end of a long process of iteration with the robot is this:

NASA’s ECCO2 ocean current visualization is pretty amazing… but is it lying to you?

When you spend your time working on marine carbon dioxide removal, like I do, you find yourself thinking about mesoscale eddies a lot.

Eddies are the ocean’s swirls — rotating lenses of water somewhere between 10 and 500 kilometers across, spinning for weeks, sometimes for more than a year, before they fall apart. They are not exotic. They are pretty much everywhere, all the time. They account for enormous amounts energy: coherent mesoscale eddies hold something close to 80 percent of all the kinetic energy in the ocean.

Every scheme for moving carbon out of the atmosphere and into the ocean has to come to grips with these systems. But following the water is not straightforward. You need to know where a parcel of seawater went when. You have to figure out how much it mixed with the parcels of water on either side and above and below, and you have to figure out whether it surfaced again to breathe on the atmosphere. That is a tracking problem — a Lagrangian problem, a nerd would say — and eddies are responsible for the bulk of the moving and the stirring. You cannot certify a tonne of removed carbon without some account of the eddy field that tonne wandered into.

Pretty much nobody outside oceanography has any picture in their head for this at all.

The standard analogy is “ocean weather,” because eddies are chaotic. They have a predictability horizon of days to weeks, and they’re the reason the ocean is turbulent and hard to model. Unlike atmospheric weather, though, eddies are invisible from where you’re standing. We have satellites staring at every square kilometer of atmosphere and the forecast is always right there in your pocket. The ocean’s weather isn’t like that. It happens beneath an opaque surface, largely unobserved, and until embarrassingly recently we could barely see it at all.

That is why I found this NASA visualization so fascinating.

It was made with ECCO2, a simulation run out of NASA’s Jet Propulsion Laboratory that blends satellite altimetry, temperature and salinity readings from thousands of drifting robotic floats, and the equations of fluid motion into a seamless portrait of the ocean. The grid cells are about 18 kilometers on each side, which is pretty high resolution for a global model. The footage — Perpetual Ocean 2, released in early 2025 — covers 2021 through 2023.

The tricky bit is that ECCO2 is a hindcast: a reconstruction of an ocean as it was in the past. Or, maybe, as it ought to have been, according to physics. But it is not a forecast and never pretended to be.

I love the visualization because it shows just how many eddies there are, spattered just all over the place and across every ocean basin. And it gives you a gut-level feel for the sheer quantity of mass and energy involved in all that moving water.

But what you’re watching doesn’t actually correspond to the ocean on any given day. What’s the video shows is the tracks of imaginary particles dropped into a computer model of the ocean, each drawn with a trail a few days long. That trail length was a choice made by animators at NASA Goddard, not by physics. And the fine, hairlike filaments threading between the big swirls are not features of the ocean, or even of the model — they emerge because when you advect a line of particles through a swirling current, stretching and folding draws it out into threads much finer than the grid cell.

The eddies themselves, though, are real model features. At 18 kilometers, ECCO2 can resolve mesoscale eddies through most of the ocean — an Agulhas ring is 200 to 300 kilometers across, which is fifteen grid points.

What the model does not know is which eddies. Its fit to observations is loose enough that no particular swirl on your screen corresponds to a specific eddy that existed at that spot on that date. What it gets right is the character of the thing: where the fast currents run, how big the eddies grow, how often the Agulhas Current pinches off a ring. Really it’s closer to the ocean’s climate than its weather.

Which poses some strange epistemological questions. Can a visualization composed entirely of motions that never actually happened bring us genuine information about the world? In what sense can that information be “true” even though it’s the result not of observation but of software loosely constrained by observation?

The kind of seeing you do when you register the density of eddies on that screen is both true and illusory at once. The eddies are there! No — not these eddies. Others a lot like them.

For most purposes, that is fine. For mine, it’s kind of a problem. If I want to quantify how much carbon a specific intervention has taken to which depth, I need to be able to track a specific, real-world column of water going forward. ECCO2 definitely can’t do that.

Because, remember, ECCO2 is a hindcast, and none of this is a fair complaint to level at a hindcast. It does what it’s built to do: close budgets and provide a general picture of the way the ocean has circulated over the past thirty years.

So what tool will work?

Definitely not ocean general circulation models. The workhorses oceanographers have been building since the 1960s mostly cannot see eddies at all. The climate-scale versions run on cells around 100 kilometers per side, sometimes 25. Far better than the old days, when a few hundred kilometers was normal, but still coarse next to a 40-kilometer eddy. So rather than simulating eddies, these models represent their average effect through a scheme called parameterization — most famously Gent–McWilliams, which captures how eddies stir tracers along surfaces of constant density without ever modeling an eddy as such.

To get at real eddies in the real ocean, you need a different tool.

In December 2022, NASA and CNES launched SWOT — the Surface Water and Ocean Topography mission — which is more or less purpose-built for this problem. Conventional altimeters see along a narrow track and blur out anything below roughly 150 kilometers. SWOT sweeps a wide swath at 2-kilometer native resolution, which allows it to pick out coherent eddies down to about 15 kilometers, an order of magnitude better.

One result has been a global census of submesoscale dynamics in Nature. SWOT has already zeroed in on small features on the East Greenland shelf that nobody could previously see. And the assimilation work has started — feeding SWOT into regional models has already measurably improved their mesoscale structure.

So the picture is not really all that bleak. But it is unfinished. With SWOT, we finally have something close to eyes on the ocean’s real weather.

What we still don’t have is the thing mCDR actually needs, which is an eddy-resolving model, constrained by SWOT-class observations, that can tell you where a specific parcel of water is heading over the coming months rather than the coming days.

Nobody needs to track a water parcel precisely for a century — that would be impossible. But what we do need is to be able to follow it long enough to confirm when it has sunk past the winter mixed layer and joined a water mass that won’t come back up for a very long time. Once it’s down there, permanence is a function of ocean physics we understand reasonably well — ventilation timescales of decades to centuries, depending on where it ended up. The hard part is the first few months. And the hard part of the hard part is the eddies.

That gap — between the weeks we can forecast and the months we need to show transport out of the system — is the major stumbling block to making marine CDR work in the real world. Until you can say with confidence that the water you fertilized went down and stayed down, nobody sane is going to pay you to do it.

In the meantime: watch the video. Marvel at how intricate and energetic the ocean really is. No, not this ocean. The real ocean, which is different. But the same.

Now, if you feed that piece to Pangram, it’ll tell you that it’s like 91% robot-written, because it is. Some people will see this alone as disqualifying, but is it?

I mean, I could’ve just have published my original draft instead. Only when I wrote that original draft I simply overlooked SWOT, which as Claude quickly picked up, would have been a very misleading omission: it would make it seem like we weren’t doing anything to solve the problem, which isn’t true, in fact we’re way on our way to solving it.

So yes, I could have just published my original excrecence, but the text would have been wrong, it would have misled you, and it would have contributed, subtly, to degrading the public sphere, as well as generally making an ass out of me in the eyes of my professional peers.

Or I could, at the cost of a lot of time and a lot of hastle, bugged a bunch of fluid dynamicists and modelers and bombarded them with ignorant questions and wasted a ton of both their time and mine, and come to substantially the same conclusion after two weeks. I suppose that’s what some people think ought to happen, but I don’t see how that’s better that consulting a robot that’s certainly read more of the literature than any one expert has.

Dario’s country of geniuses in a data center doesn’t feel like the future to me. For topics like this one, it’s here now. Humanity has invented a technology that can make an ignorant asshole like me write an essay about fluid dynamics and get the details right. If the syntax is a little stilted, ultimately, doesn’t matter as much as that the damn thing is right.

I don’t know, maybe people think this frame narrative is more interesting than the robot output. It’s messy and human and a little bit redundant and certainly could stand to omit needless words. I don’t think a blog centered on just the machine outputs will ever be that engaging.

But I do think it’s worth bringing this discussion more into the open. Because if your ultimate goal is to inform a public debate, refusing on principle to use the tool that prevents you from making blunders is just perverse. When it comes to the ever-present journalistic need to identify mistakes and correct them, Claude Fable is already on a par with the best humans, and orders of magnitude faster.