Daniel's been chewing on the follow-up to the ship-to-ship transfer episode, and he's landed on something that sounds like a simple question but opens into a hole with no visible bottom. He wants to know whether anyone has actually tried to count the private image recognition models out there. His framing: detecting a ship at sea is almost trivially easy, because the ocean is this vast, uniform backdrop and anything rectangular and flat and a different color is either land or a boat. But identifying which kind of vessel it is requires custom training data, and that's where the difficulty lives. He connects it to Frigate, the open-source NVR system, where the model can tell you there's a bird in the frame, but it can't tell you whether it's a pigeon or a dove. The granularity just isn't there. So the question is: if YOLO is the foundational open-source model that anyone can build on, how many custom image recognition models have never been released? Governments, NGOs, companies working on proprietary internal data. And separately, beneath the generalist models, there's this whole ecosystem of highly specific models. He wants a sense of any estimates that have been put together for the breadth and diversity of what's been released publicly versus what's been built in private.
The short answer on the estimate question is that no comprehensive public estimate exists. And that's not because nobody's looked. It's because the thing being counted is deliberately not visible. But you can triangulate from a few directions, and the numbers you land on are staggering.
So the answer to "has anybody counted" is no, but the answer to "can we get a sense of scale" is yes, sort of. Let's start with why the two tasks are so different, because that's the foundation Daniel's question rests on.
Right. Detection versus identification. They're not adjacent points on a difficulty curve. They're different problems entirely. Ship detection in SAR imagery, specifically Sentinel-1 C-band SAR, is a solved problem to the point of being boring. The standard approach is something called CFAR, constant false alarm rate. You look at the ocean, which is this enormous, low-entropy background. Sea clutter has a fairly predictable statistical signature. A ship is a bright or dark anomaly against that background. You set a threshold, you sweep the image, anything that pokes above the threshold is a detection. The ocean does most of the work for you, because it's uniform. A rectangular, flat, uniform mass of a different color is either land, which you already know about, or it's a vessel. That's it. That's the whole trick.
Which is why Daniel called it one of the easiest tasks in satellite imaging. He's right. The contrast between vessel and water is enormous compared to, say, finding a specific building in an urban area, or identifying a crop type in a field where everything is green and textured.
The ocean is the best possible canvas for this. It's dark, it's smooth, it's statistically well-behaved. A ship is a hard-edged, bright or dark object sitting on top of it. CFAR has been around for decades. It's not even a machine learning approach in its classic form. It's just signal processing. You estimate the local noise floor, you set a threshold, you flag anything that exceeds it. You can do ship detection with a laptop and a Sentinel-1 scene.
And then you try to figure out what kind of ship it is, and suddenly you're in a completely different universe. That's the fine-grained visual classification problem, FGVC. You're no longer asking "is there an object here." You're asking "which of two hundred possible vessel types is this object." And the differences between classes are subtle. Hull shape. Superstructure layout. Length-to-beam ratio. Where the cranes are. Whether there's a helipad. A cargo ship and a tanker look almost identical from directly overhead at ten-meter resolution. A research vessel and a small naval auxiliary are separated by details you need a maritime expert to point out.
This is the pigeon-versus-dove problem. A general object detection model sees "bird." It doesn't see "rock dove versus mourning dove versus wood pigeon." The training data never told it those distinctions existed. It was trained on something like COCO, where "bird" is one class among eighty. The model learned to find birds. It never learned to tell them apart. And the reason it never learned is that nobody labeled the training data with species.
And that's exactly what's happening with Frigate. Frigate runs a YOLO-based model, usually something optimized for TensorRT on a small device. It detects person, car, dog, cat, bird. That's the taxonomy. The model was trained on a dataset where those are the classes. If you want Frigate to tell you whether the bird at your feeder is a house finch or a purple finch, you need a different model. You need a model trained on a species-labeled dataset. You need someone to have sat down and labeled thousands of images of finches with the correct species. That dataset exists for birds, by the way. CUB-200-2011, two hundred bird species, about twelve thousand images. It's tiny compared to the general object detection datasets. COCO has over two hundred thousand labeled images. ImageNet has millions. So the general models are trained on enormous data, and the specific models are trained on tiny data.
And the long-tail problem makes this worse. Real-world vessel type distributions are heavily skewed. A few classes dominate: cargo ships, tankers, fishing vessels. Those are the head of the distribution. Then there's this enormous tail of specialized vessels: cable layers, dredgers, research vessels, naval auxiliaries, offshore supply vessels, icebreakers. Hundreds of types. Each one is rare. If you train a model on a balanced dataset, it fails on the tail because it never saw enough examples. If you train on the real-world distribution, it's biased toward the head and it misclassifies the rare types as common ones. There's no free lunch. You need targeted, expensive data collection for each rare class.
The thing to understand is that the model architecture is not the bottleneck. YOLO is freely available. The descendants, the fine-tuning pipelines, the training frameworks, all open. The bottleneck is the labeled data that maps the model's output to the specific granularity you need. For ships, labeled SAR imagery of specific vessel types is expensive to produce. You need maritime experts to annotate. You need AIS transponder data to cross-reference, and AIS data for sensitive vessels is often unavailable or deliberately obscured. Sometimes you need on-site verification, which for a ship at sea means you have to go look at it. The data is the scarce resource, not the model.
So that's the mechanism. Detection is easy because the ocean is a high-contrast, low-entropy background and a ship is an anomaly. Identification is hard because you need fine-grained labels, and fine-grained labels are expensive, scarce, and skewed toward common classes. The Frigate analogy is exact: the model sits at the coarsest level of granularity because that's where the training data sits. To get more specific, you need custom data and a custom model.
Which brings us to Daniel's actual question. How many of those custom models exist? The ones that were never released. The ones governments and companies and NGOs built on proprietary internal data. The short answer is nobody knows. But let's triangulate.
The public model ecosystem is the lower bound. Hugging Face crossed three million models. That was the milestone everyone pointed to. But the vast majority of those are NLP models. Text generation, text classification, embeddings. The image recognition subset is much smaller. Still substantial, but smaller. And that's just what people chose to upload. That's the visible tip.
The private side is where the numbers get wild. Every company with a proprietary image dataset and a business need has likely trained a custom model. Retail shelf-scanning, so a company can audit what's on their shelves without sending a human down every aisle. Agricultural crop monitoring, where the model has to distinguish between crop varieties or disease states on proprietary drone imagery. Medical imaging diagnostics, where a hospital system trains a model to flag a specific pathology on their own scanner data. Industrial defect detection, where a factory trains a model to spot a specific manufacturing flaw on their production line. Security and surveillance, where a company trains a model to recognize specific threat objects or specific individuals. Autonomous vehicle perception, where every company has their own internal datasets of road scenes, pedestrians, traffic signs, edge cases. None of these are released. They're competitive advantages, or they're regulated, or they're just internal tools nobody thought to publish.
And then there's the government and NGO layer. Satellite imagery analysis for intelligence. Environmental monitoring. Disaster response. Maritime domain awareness. The ship-type identification models used in illicit transfer detection are exactly this. They exist. They work. They are not public. They're classified, or restricted, or the data they were trained on is sensitive, or the operational advantage they provide is something the owner doesn't want to give away. So the models that would actually answer the question Daniel started from, the ones that can tell a tanker from a cargo ship in SAR imagery, those are precisely the ones that will never show up on Hugging Face.
Here's the dynamic that makes the count essentially unknowable. YOLO is the foundation. It's freely available. Anyone can fine-tune it on custom data. The barrier to entry is not the model architecture. It's the data. And the number of organizations with a proprietary image dataset and a reason to train is enormous. Every hospital, every factory, every farm, every retailer, every government agency with a satellite imagery budget. Each one of those is potentially training a custom model. The number is plausibly in the hundreds of thousands to millions. And that's not counting the models that get trained, used, and thrown away. The experimental ones. The ones that didn't work. The ones that were fine-tuned for a single project and then abandoned.
The "trained and thrown away" category is bigger than people think. A research team gets a grant, collects a dataset, trains a model, writes a paper, maybe releases the paper but not the model. Or a company trains a model for a specific client engagement, delivers the result, and the model sits on a hard drive somewhere until the drive dies. The half-life of a private model is short.
And this has knock-on effect. The invisibility of private models has implications for AI safety, reproducibility, and governance. We can audit what's public. We can look at the released models and say, here's what this one does, here's what it was trained on, here are its failure modes. But the private ecosystem is a black box. We know it's there by its gravitational effects. We see the products, the services, the decisions that private models make. But we can't see the models themselves. It's the dark matter of AI. We infer its existence from the way the visible stuff behaves, but we can't observe it directly.
And the diversity is not just in number. It's in kind. Each private model encodes a particular taxonomy. A particular dataset. A particular operational context. A retail shelf-scanning model and a satellite imagery model share an architecture, maybe, but they encode completely different ways of seeing the world. The retail model knows what a specific brand of cereal looks like from a specific angle under specific lighting. The satellite model knows what a specific class of vessel looks like from directly overhead in SAR. They're both image recognition, but they're not the same language. They're not even the same dialect.
There's a misconception that private models are just slightly modified versions of public models. Fine-tuned YOLO, fine-tuned ResNet. And technically, yes, the architecture is shared. But the weights, the learned representations, the way the model carves up the visual world, those are fundamentally different. A model trained on proprietary shelf imagery has learned a taxonomy of products and packaging. A model trained on classified satellite imagery has learned a taxonomy of military equipment. They share an architecture, but they don't share a visual language. The private model is not a copy of the public one. It's a different thing that happens to use the same scaffolding.
So when Daniel asks about the breadth and diversity of image recognition models, the honest answer is that the public models are a small, biased sample of the total population. The public models are the ones where the data wasn't sensitive, the business need wasn't competitive, and someone had the time and inclination to package them up and upload them. That's a narrow slice. The private models are the ones where the data was proprietary, the application was classified, or the advantage was too valuable to share. And those are the ones that actually do the work in the real world.
For the ship-to-ship transfer problem specifically, this is why the detection-versus-identification gap persists. The detection models are public, because detection is easy and the data is free. Sentinel-1 imagery is open. The identification models are private, because identification is hard and the data is sensitive. The models that can tell a tanker from a cargo ship in SAR exist. They work. They're running right now inside government agencies and maritime intelligence companies. But they're not coming out. The data they were trained on is classified, or the operational advantage is proprietary, or the regulatory environment discourages release. So the public is stuck at the detection layer, and the identification layer stays submerged.
The iceberg metaphor is almost too neat here. The visible tip is the public model ecosystem. Hugging Face, the open-source releases, the academic papers with attached weights. The submerged mass is everything else. And the submerged mass is what actually shapes the water. The models that matter for maritime security, for intelligence, for competitive business advantage, those are all down there. And they're not coming up.
Here's the thing that makes this hard to think about. The submerged mass is growing. Training a custom model gets cheaper every year. The frameworks get easier. The hardware gets faster. The data collection gets more automated. So the barrier to entry keeps dropping. Every organization that was on the fence about training a custom model is now over the fence. The private model population is exploding, and we have no way to count it.
The answer to Daniel's question is: no, nobody has estimated it, and the reason nobody has estimated it is that the thing being estimated is deliberately invisible. But the lower bound is the public ecosystem, which is already huge, and the upper bound is essentially unbounded. Hundreds of thousands to millions is a reasonable guess, but it's a guess. The real number could be higher.
There's a philosophical point underneath this. Every private model is a frozen decision about how to categorize the world. Somebody decided that these classes matter and those classes don't. Somebody decided that this level of granularity is worth paying for and that level isn't. The private models are a record of those decisions, and most of them will never be seen by anyone outside the organization that made them. We're building an enormous, distributed, invisible taxonomy of the visual world, and we're doing it without any coordination or oversight.
That's the part that sticks with me. The public models are the ones where someone said, here, look at how I carved up the world. The private models are the ones where someone said, this is how I carve up the world, and it's none of your business. And the second category is vastly larger.
Hilbert: I had a job in the late nineties doing aerial survey photography for agricultural insurance. We flew a Cessna over farmland and took thousands of rolls of film. My job was to go through the contact sheets and circle anything that looked like a drowned field or a hail-damaged crop.
Hilbert: We didn't have models. We had me and a light table and a magnifying glass. And the weird thing is, I got really good at telling the difference between flood damage and hail damage. Flood damage is smooth and uniform. Hail damage is speckled. I could do it in about two seconds per frame. But I couldn't explain how I did it. It was just pattern recognition.
Hilbert: When you say nobody knows how many private models there are, I believe you. But I also think every one of those models is just somebody's version of me at a light table. Somebody looked at a lot of pictures and learned to tell the difference between two things that look almost the same. The model is just the part where they wrote it down.
Hilbert: I still have one of the contact sheets from that job. It's in a drawer. It's the only piece of my work from that era that I kept, because it's the only one where I can still remember exactly what I was thinking when I circled a particular field. I looked at it and I thought, that's not flood damage. That's hail. And I was right. The adjuster went out and confirmed it. But I couldn't have told you why I knew. I just knew.
Hilbert: I don't know if that helps. But that's what I thought of when you were talking about pigeons and doves.
The contact sheet is the model. That's what's so striking about what Hilbert's describing. The trained weights, the learned representation, the thing that distinguishes flood from hail, it was in his head. And when he circled the field, he was making an inference. The model existed. It just wasn't written down anywhere. And now it's in a drawer, and when he's gone, it's gone.
Which is the other end of the iceberg. Not just the private models that exist but are hidden. The models that existed and were never written down at all. Every human expert who ever looked at a thousand images and learned to tell two things apart, that's a model that never got saved. Hilbert's contact sheet is the only artifact of one of them, and it's not even the model itself. It's a single inference. A single circled field.
The scale of that is even more impossible to estimate. Every radiologist who learned to spot a specific pathology. Every mechanic who can diagnose an engine by sound. Every farmer who knows what a diseased leaf looks like. Every one of those is a trained model, and most of them were never written down, never formalized, never turned into weights. They just existed in a person's head, and they vanished when the person stopped doing the work.
Daniel's question about counting private models is really a question about counting knowledge. And the answer is that most of the knowledge was never counted, most of it was never written down, and most of it is gone. The private models are the ones that got written down but not released. The human models are the ones that never got written down at all. And both populations are invisible.
Hilbert: The field I circled was in Iowa. I remember that. It was a cornfield, which I thought was funny at the time. I was circling hail damage on a cornfield, and my name is Hilbert. It's not a joke. It's just a thing I remember.
Hilbert: The adjuster said I was right about the hail. He said the pattern was textbook. I asked him what textbook, and he said he didn't know. Somebody told him once. That's how it worked. The knowledge just moved from person to person. Nobody wrote it down.
The oral tradition of image recognition. That's what existed before the written-down version. And now we're in the written-down era, but most of the writing is private. So we've gone from invisible knowledge in people's heads to invisible knowledge in proprietary weights. The visibility problem didn't get solved. It just changed form.
The contact sheet in the drawer is the bridge between the two. It's the moment when the knowledge was almost written down. A circled field on a piece of film. If Hilbert had written down why he circled it, that would be a dataset. If he'd circled ten thousand fields and written down why, that would be a training set. And someone could have trained a model on it. But he didn't, and they didn't, and so the model stayed in his head until it didn't.
Hilbert: I could still do it, probably. If you put a contact sheet in front of me right now, I could tell you flood from hail. The eyes still work. But I don't have the contact sheets anymore. Just the one. And I wouldn't know where to get more. The company folded in two thousand one. The Cessna got sold. The film archive went to a warehouse somewhere, and I don't know what happened to it after that.
Hilbert: The model still exists, but the data doesn't. That's the other way models die. The weights are fine, but the inputs are gone. You can't run inference if there's nothing to look at.
That's a point we haven't even touched. The private models aren't just invisible. They're also mortal. A company trains a model, the company goes under, the model and the data both vanish. A government agency trains a model, the program gets cut, the model gets deleted or just sits on a retired server. The half-life of a private model is short. We've been talking about the iceberg as if it's stable, but it's actually melting and refreezing constantly.
The answer to Daniel's question keeps getting more complicated. Not only has nobody counted the private models, but the count is changing constantly. Models are being trained and discarded. Companies are folding and taking their models with them. Experts are retiring and taking their unrecorded knowledge with them. The invisible population is dynamic, and any estimate would be a snapshot of a moving target.
Yet the lower bound is still useful. We know the public models are at least in the hundreds of thousands, if you count everything on Hugging Face and the other repositories. We know the private models are plausibly in the hundreds of thousands to millions, based on the number of organizations with proprietary image data. And we know the human models, the ones that never got written down, are in the millions at minimum, because every expert in every visual domain is one. So Daniel's question about the breadth and diversity has a real answer, even if it's not a precise number. The breadth is enormous, the diversity is enormous, and most of it is invisible.
The specific models Daniel started from, the ones that identify vessel types in SAR imagery, those are a small fraction of the invisible mass. But they're the fraction that matters for maritime security, and they're the fraction that will never be released. So the public conversation about ship detection stays stuck at the detection layer, while the identification layer, the part that actually answers the question "is that a tanker transferring oil to a dark ship," stays hidden.
The open question is whether that invisibility is a problem we can solve, or just a fact we have to live with. The private models aren't coming out. The data they were trained on isn't coming out. The expertise that created them isn't coming out. So the public understanding of what's possible in image recognition will always lag behind what's actually being done. The iceberg will always be mostly underwater.
That's the thing to sit with. We can count the tip. We can guess at the mass below. But the mass below is what's actually shaping the water. The models that matter, the ones making decisions about ships and crops and tumors and threats, those are down there. And they're not coming up.
Thanks to Hilbert Flumingtop for producing the show. This has been My Weird Prompts. If you enjoyed this episode, leave a review on your podcast platform of choice. It helps other listeners find the show.
We'll be back soon.
