Technology alone is never enough. For it to be useful, it must be accompanied by other elements, such as public understanding, good habits, and acceptance of collective responsibility for its consequences. Without this social halo, technology tends to be used ineffectively or incompletely. A good example of this might be mRNA vaccines, which were created a long time ago. COVID infectious disease. These were amazing medical achievements, but they didn’t land as well as expected because of widespread lack of understanding. It may not even be appropriate to call a technology a technology if it lacks the necessary elements to usefully introduce it into the human world. If we don’t understand how technology works, we risk succumbing to magical thinking.
Another way to say this is that you need to create a cartoon in your head of how the technology works. I don’t know enough about vaccines to make my own, but I have a comic book about vaccines, so I can get a general understanding. Enough to follow the news about vaccines and understand the technology’s development process, risks and expected future prospects. I have similar comics in my head about rockets, financial regulation, and nuclear power. They’re not perfect, but they give you enough intuition. Even experts sometimes use comics to talk to each other. Sometimes you can see the forest for the trees by looking at things in a simplified way.
I have some tension with many in the computer scientist community on this point. I think the comics we aired about AI are counterproductive. We brought artificial intelligence into the world with ideas that are both useless and perplexing. The worst of which is perhaps the sense of human obsolescence and doom that many of us convey. I have a hard time understanding why some of my colleagues say what they are doing could lead to the extinction of the human race, yet insist that it is still worth doing. Masu. It’s hard to understand this way of talking without wondering if AI is becoming a new kind of religion.
Besides the apocalyptic vibe, we don’t have a good explanation of what it is and how it works. Most non-technical people can better understand a cumbersome abstraction when you break it down into concrete parts that tell a story, but in the world of computer science, this can be difficult. We typically prefer to treat AI systems as a giant impenetrable continuum. Perhaps there is some resistance to figuring out what we are doing because we like to engage in mystery. The usual terminology that begins with the phrase “artificial intelligence” itself is based on the idea that we are creating new creatures rather than new tools. This concept is further reinforced by biological terms such as “neuron” and “neural network” and anthropomorphic terms such as “learning” and “training” that computer scientists use all the time. Another problem is that there is no set definition for “AI.” It is always possible to ignore certain comments about AI because they do not refer to other potential definitions of AI. This lack of stereotypes about the term coincides with a metaphysical sensibility that will soon transcend human frameworks.
Is there a way to describe AI that doesn’t imply human obsolescence or replacement? If we can talk about our technology in other ways, there may be a better path to bringing it into society. I don’t know. In a previous essay I wrote for this magazine, “AI Doesn’t Exist,” I argued for rethinking big model AI as a form of human collaboration rather than as a new creature in the field. This article explains how such AI works without going into the often arcane technical details, how the technology modifies human input, and how human input I would like to emphasize that it depends on This is not an introduction to computer science, but rather a story about cute objects in time and space that serve as metaphors for how we have learned how to manipulate information in new ways. We’ve found that most people can’t follow the common stories about how AI works, but not about other technologies. We hope you find the alternatives presented here useful.
You can draw a human-centered comic about large model AI in 4 steps. Each step is easy. However, when you accumulate them, it becomes easier to visualize them and use them as a thinking tool.
I. Tree
The very first step, in some ways the simplest step, may be the most difficult to explain. Let’s start with a question. How can you use a computer to tell if a photo contains a cat or a dog? The problem is that cats and dogs generally look similar. Both have eyes and noses, tails and paws, four legs and fur. It’s easy for a computer to measure an image and determine whether it’s bright or dark, blue or red. However, such measurements cannot distinguish between cats and dogs. Similar questions can be asked for other examples. For example, how can a program analyze whether a passage was likely written by William Shakespeare?
On a technical level, the basic answer is a complex web of statistics called neural networks. But the first thing to understand about this answer is that we are dealing with complex technology. Neural networks, the most basic gateway to AI, are something of a folk technology. When researchers say that AI has “emergent properties,” we often mean that we didn’t know what the network would do until we built it. There is. AI is not the only field like this. Medicine and economics are similar. In these areas, you try and try again to find more effective techniques. We don’t start with a master theory and use it to calculate ideal results. Still, you can manage complexity even if you can’t fully predict it.
Think fancy about differentiating cat photos from dog photos. Digital images are made up of pixels, so we need to do something more than just a list of pixels. One approach is to put a grid over the image that measures more than just color. For example, you can start by measuring the degree of color change in each grid square. You now have a number for each square that may represent the saliency of sharp edges within that patch of the image. Even one layer of these measurements cannot tell the difference between a cat and a dog. But you can put her second grid on top of the first grid and measure something about the first grid, then another grid, and then another grid. You can build a tower of layers with a measuring patch at the bottom of the image and each subsequent layer measuring the layer below it. This basic idea has been around for half a century, but only recently have we found the right tweaks to make it work. Who knows if there’s a better way?
Here, we will draw a manga that looks like an illustration for a children’s book. These tall grid structures can be thought of as large tree trunks extending out from the image. (Most of the photos are rectangular, so the trunk is probably rectangular, not circular.) Inside the tree, each small square on each grid is decorated with a number. Imagine yourself climbing a tree and as you climb he looks inside with an x-ray. The number found at the highest point depends on the number at the lower points.
Unfortunately, so far it is still not possible to distinguish between cats and dogs. But now we can start “training” the tree. (As you know, I hate the anthropomorphic term “training,” but I’ll leave it at that.) Imagine that the bottom of the tree was flat, and you could slide a photo underneath it. He then creates a collection of cat and dog photos, clearly and precisely labeled “cat” and “dog,” which he slides one by one down to the bottom layer. The measurements cascade upwards toward the top layer of the tree, the canopy layer that can be seen by someone riding in a helicopter. Initially, the results displayed by the canopy are inconsistent. But you can, for example, use a magic laser to fly into the tree and adjust the numbers on the various layers to get a better result. You can increase the number that turns out to be most helpful in distinguishing between cats and dogs. This process is not trivial because changing numbers in one layer can propagate changes to other layers. Finally, if successful, the canopy leaves will all have numbers 1 if there is a dog in the picture, and all numbers 2 if there is a cat in the picture.
Amazingly, we have created a tool – a trained tree – to distinguish between cats and dogs. Computer scientists refer to the grid elements at each level as “neurons” to suggest a connection to the biological brain, but the similarity is limited. Biological neurons are sometimes organized into “layers”, such as in the cortex, but this is not always the case. In fact, the cortex has fewer layers than an artificial neural network. However, AI found that adding more layers significantly improved performance. This is why the term “deep” is often used, as in “deep learning.” This means many layers.
