I’ve talked to a lot of people the last few weeks – administrators, faculty, and students alike! – who don’t know what on earth is going on with the “NC State AI Hub.” The rollout of the “hub” has been very strange!
(Apologies for putting the earlier series on hold for a second. I know that everyone is at the edge of their seats wondering how Big Tech will fall to the seductive promise of Artificial Intelligence, but I accidentally spent all weekend bewildered and haunted by the Hub, so I wanted to get this out.)
In order to better understand what’s going on here, I’d like to take a close look at one of its projects. As best as I can tell, this project is genuinely really good and valuable! That said, it should also indicate just how far away the Good Uses of “AI” (like this project) are from products like LLM-based chatbots and coding assistants. I think this distinction is very important for everyone thinking about this technology to understand, and I hope this letter helps! More soon… – Theodora
Sweet, Sweet Sweetpotatoes
The AI hub website has a link near the bottom to a list of articles which “Explore how the Pack leads with AI.” The second initiative on the list is titled “AI Meets Agriculture: Smarter Sorting for Sweetpotatoes.”
The article this links to is from nearly a year ago today, October 2025; the sweetpotato project is described first.
The write-up doesn’t get into too much detail; instead relying on phrasing like “AI-driven” to perform the work of connecting it to “AI.” I had to comb back through Williams’s publications – he and his collaborators are doing a lot of different things with machine learning and this project. Most usefully for our purposes, I think, I found a more implementation-oriented publication they released in January.
(A caveat: it is difficult to overstate the extent to which I am not an expert in any of this stuff. But frankly, I think this is important information, and I haven’t been able to find it anywhere else. If my ignorance upsets you as much as it upsets me, note that I am willing to entertain offers of funding to go back to school and get a computer science degree, as I am proving disappointingly bad at remembering to finish the YouTube tutorials on Python that I keep starting.)
This publication presents “the results of a model that can perform grading and provide yield estimates directly in the field faster than manual measurements.” Basically, they trained an “AI-powered” system to perform the sweetpotato image recognition; this publication makes this system accessible to other researchers and farmers interested in giving it a try.
Great! This sounds sincerely useful. What’s the AI part?
How this project uses AI
Here is a sentence that tells us the AI part:
“Detectron2, a library consisting of deep-learning object detection algorithms, was used to implement Mask R-CNN, an instance segmentation model.”
Cool. Hope this helped. Just kidding! Let’s break statement down a bit.
We don’t need to worry much about Detectron2 for now: it is essentially a platform for using image recognition models like the one Dr. Williams’s lab trained and released. Mask R-CNN, the model itself, is more to our point.
A Brief Discursion on “Models”
“Model” is an underratedly complex word, one that philosophers are arguing about as we speak. Here’s a brief account of the understanding I, at least, am working from:
A model is a representation of something. (Think “model plane,” “model architecture,” even “mental model.”) More specifically, a model is a representation not primarily intended to serve aesthetic or mimetic ends, but which is intended to contain and convey information.
Models don’t purport to explain reality – that’s theories’ business – but they do purport to help us get some perspective on phenomena: to understand what is going on, if not exactly why.
(If you’ve read my letter on AlphaFold – well, first of all, thank you! – but second of all, this might be a familiar theme.)
Now, a model can never represent an entire phenomenon, but it can represent some aspects of that phenomenon. This can lead them to be pretty abstract.
But they’re still representations. When we talk about a “weather model,” we’re not talking about something that looks like the weather in any way we’d recognize; we’re talking about a mathematical representation of the behavior of physical systems, designed to convey information.
A general-purpose image recognition model attempts to represent the relationship between an image and the objects it contains (image from Mask R-CNN GitHub page):
The model that Williams and his team made, then, doesn’t “represent” sweetpotatoes in the same way that the image on the NCSU website does, or that a diagram of the parts of a sweetpotato might. It’s not really a model of a sweetpotato at all, per se.
Rather, the model designed by Dr. Williams and his team is a specially-tuned version of a general-purpose image recognition model. In order to do this, the team began with an image recognition model (the aforementioned Mask R-CNN), then trained it to specifically recognize sweetpotatoes in a way that allows us to find pretty accurate size estimates. More specifically, it converts one kind of information – the information contained within a digital photograph of a sweetpotato – into another kind of information, namely, an estimate of the sweetpotato’s size.
AI Big and Small
This section is a little in the weeds, but if we want to understand how exactly a model like Mask R-CNN is related to “AI” writ large, we have to get pretty granular. The upshot for those disinclined to wade is that the AI used in this project is a very distant relative of the AI used by, for example, modern chatbots.
Okay! Now that we have the basic outlines of what an image recognition model is, let’s talk a little bit about how it works.
Like other models collected under the term “AI,” it’s an intricate mathematical function: an algorithm for taking inputs and returning outputs. As we said, in the case of image recognition models, the input is a digital image, and the output is textual and quantitative information about the probable contents of the image.
Mask R-CNN, the model used in the sweetpotato project, analyzes images in two stages: first, it uses an older approach to detect the general location of objects within images; then it uses “instance segmentation,” which actually breaks the image down pixel-by-pixel in order to return more accurate results.
The model behind the AI sweetpotato project uses Mask R-CNN, which is an “instance segmentation model.” In terms of what we call “AI” right now, this second step is the only aspect of Mask R-CNN that uses any of the techniques developed in the recent “deep learning revolution.” They do not generally use the “transformer” architecture that enabled LLMs, instead using “convolutional neural networks” – a machine learning technique dating from the 1980s, but which modern graphics cards (“GPUs”) made much, much faster. (The specific technique for doing this was developed in 2006.)
The sweetpotato initiative itself began in 2020 – about two years after the beginning of the “deep learning” revolution which began our current AI hype cycle, but two years before it reached a mass audience with ChatGPT (released in November of 2022).
Mask R-CNN
Their model is based on a pre-existing model called “Mask R-CNN.” Mask R-CNN is a small open-source model specifically designed to process computer images by separating and bounding different objects within the image. Researchers is can fine-tune the model themselves, training it on a data set specific to their needs.
It’s difficult to overstate how distinct this is from the “frontier models” undergirding ChatGPT and Claude. Commercial models are completely locked down: you cannot train them like you can Mask R-CNN. The illusion of commercial chatbot “memory” only exists because the entire conversation is invisibly fed back through the model every single time it’s prompted.
But even if you could train them, if you did it at all like OpenAI or Anthropic do, it would take millions of dollars worth of “compute,” or computing power. OpenAI spent $19 billion on R&D last year alone. The annual operating budget for the entire UNC system is just over $12 billion. Most of our “frontier”-scale open models are developed in China.
Models like those powering ChatGPT and Claude are also designed to do everything, in a way that makes them far less efficient than Mask R-CNN. Mask R-CNN does one thing: image recognition. You can’t talk to it. It can’t write you a paper. In terms of power and scale, trying to do what Mask R-CNN does with ChatGPT would be kind of like trying to take a Boeing 747 through a drive-through.
This has enormous practical implications. According to the original publication announcing Mask R-CNN, training, when performed a decade ago, took a couple of days on a single 8-GPU computer. While I haven’t found hardware information for this specific project, another paper refers to training from a similarly-sized dataset on an Nvidia GTX 3090 – a $250 video card, slightly more powerful than the one I have in my six-year-old desktop computer. And this is just the training process: actually using the model to recognize sweetpotatoes isn’t very compute-intensive, either. (Mask R-CNN is about 75MB to download from its official GitHub page.)
It also means that farmers could own and operate their own copies of the specialized Mask R-CNN model. They aren’t dependent on the whims of unprofitable corporations; they won’t have to subscribe or pay for tokens.
What We Talk About When, etc.
Though boosters deflect away from this fact whenever it seems to suit them, the core of the current “AI revolution,” the heart of the hype cycle, is LLM-based text generation. Without chatbots – and, to a lesser but real extent, other generative products like video generation – AI would not have received nearly $2 trillion in funding. When someone talks about “using AI,” they are, in all likelihood, talking about something powered by an LLM: they’re talking about chatbots like ChatGPT or coding assistants like Claude Code. Even “agentic” AI, the implementations responsible for the criminal cybersecurity breaches we’ve seen recently, is built on a foundation of LLM-based text generation: of running commands through the text-base terminal.
Now, someone talking about “AI” might be talking about software which has “AI” integrated into it; this might be a specialized model like those we’re talking about: image recognition, text-to-speech, audio transcription, etc. But the elision is deceptive. As we’ve seen, versions of most of these products existed before. Mask R-CNN is a decade old. And in the specific case of image recognition, most implementations haven’t found much use for the “transformer architecture” that sparked this whole thing.
So what does Mask R-CNN have to do with “AI” writ large?
Well, half of Mask R-CNN’s process is reliant on “deep learning,” an umbrella term that includes the processes used to train the LLMs powering ChatGPT and Claude. Deep learning dates back to the 1980s;l in the 2000s and 2010s, researchers began to realize that graphics chips could allow them to scale up their training, making it practically applicable.
But that’s about it. Mask R-CNN is a specialized image recognition model trained and run on relatively cheap local hardware. It isn’t trained on stolen material like LLMs. It doesn’t require apocalyptic data center buildouts. It doesn’t blow through unfathomable amounts of power to train and use. It doesn’t diminish your capacities for independent thinking or tell you to kill yourself.
It might be useful to compare deep learning with something like fertilizer. Say a new kind of fertilizer was developed, and that it’s really helping apples flourish. Great! But just because the fertilizer helped grow apples, and apples are good, doesn’t mean that everything grown with the fertilizer is good. Saying that we need to get our students Claude Pro subscriptions because Mask R-CNN is useful would be like saying that, because new fertilizers are helping improve apple harvests, it is imperative that our students have access to cigarettes rolled with tobacco that was grown with the same fertilizers.
When we interact with “AI,” we’re not interacting with the technology used to develop the models underpinning software: we’re interacting with specific consumer and enterprise products. The AI of this project is not remotely the same thing as Google Gemini or ChatGPT, and the implication that it is is disingenuous and manipulative.
The Genuine Intelligence Project is an initiative from the North Carolina branch of the American Association of University Professors. Check out our website, and follow us on Instagram and BlueSky!