Help me understand AI

quote:

---

Does AI collect images from around the web and piece it together into a new image or is every element generated?

---

No, they are fed with information in order to create a pattern of behavior.

quote:

---

For example, would you ever see yourself in an image generated by AI?

---

No

quote:

---

Is every person, building fake unless you specify like, say the White House?

---

Yes

quote:

---

Only AI has a better more detailed memory and is better at drawing.

---

Everything but hands. And eyes.

For the most part, it's a very clever trick. Computers have always been able to spot patterns and predict what hypothetical things would look like based on that. They just have a lot more data to work with now, and more power to do the work.

AI doesn't really "know" anything. That's how you end up with a shockingly good rendition of, IDK, "Neil Armstrong planting an LSU flag on Mars" but with a "Z" instead of an "S."

I will say, the youngsters who don't take the "AI can write code" hype seriously and get CS degrees in spite of it will make a lot of money... just as those of us who disregarded the noise of 30 years ago about Indian programmers have.

Who could have predicted that Americans would be captivated by the cute little pictures that AI creates to the point that they wholesale ignore the threat that AI (even in its current form) actually is? It is the perfect representation of where we are as a society these days.

"Look! Shiny new toy!"

quote:

---

Does AI collect images from around the web and piece it together into a new image or is every element generated? For example, would you ever see yourself in an image generated by AI? Is every person, building fake unless you specify like, say the White House?

---

The answer is that's not quite how these things work. They train on tons of data, much of it from the Internet. But they don't store these as individual documents or photos. It stores the patterns it finds like what words typically come before and after other words. Or what people, places and things are in different types of photos.

quote:

---

For example, would you ever see yourself in an image generated by AI?

---

Are you famous enough to figure prominently in its training data? It does well with pictures of celebrities. But again, they aren't storing images in a typical database and just pulling faces out of a hat. They "draw" the image from a starting point of noise, so if it ends up drawing you, that's either coincidence, or you figured prominently in its training data, like celebrities would be in a lot more photos than nobodies.



This article has a pretty good laymen's terms explanation of the process for images. LINK

quote:

---

Does AI collect images from around the web and piece it together into a new image or is every element generated?

---

So the first thing to know is that there are different types of AI image generators. Right now, diffusion models are the most successful. But that may change one day. I bring this up because some of what I’m about to say may not apply to other types of models.

Diffusion models are neural networks (read: AI’s) that learn to create images from noise. Or more accurately - they learn to differentiate noise from an image. They are trained by taking a complete image and slowly adding random noise to it, like this:



During training, the model’s goal is to look at the image and estimate the noise that was added. After many rounds of training you’re left with a model that’s capable of looking at a noisy image and determining which parts are noise.

Then during actual image generation (referred to as inference) the model does the exact opposite. It starts with an image that consists of pure random noise. The model determines the “predicted” noise from its training and subtracts that from the actual image. It looks something like this:



One key thing to note is that the model denoises the image in multiple steps. So it doesn’t subtract all of the noise at once. At each step the model looks at the image, estimates the noise, then subtracts a percentage of that estimated noise. This allows the model to hone in on an image rather than trying to create something from nothing in one shot. The amount of noise subtracted at each step is determined by the noise schedule, which can vary depending on the particular model being used. Some more customizable models, like Stable Diffusion, allow the user to pick from various noise schedules.

The image dataset used for training is obviously very important. If you only train the model on images of cats, it will be very good at creating cats but that will be the only thing it produces from noise.

The captioning of the training images is also incredibly important. At each step of training the model is given an image with some amount of random noise added along with a caption describing the original image.

So you might have a noisy image of a cat with the caption “painting of a cat by Vincent Van Gogh” (I’m sticking with cats for consistency ). You might also have noisy image of a dog with the caption “painting of a dog by Vincent Van Gogh.” You might also have another noisy image of a cat with the caption “painting of a cat by Claude Monet.”

With these three images, the model learns four concepts: dogs, cats, the style of Vincent Van Gogh, and the style of Claude Monet. Now extrapolate that over hundreds of millions of images, and you can begin to see how these models learn so many concepts. The LAION dataset used to train Stable Diffusion consists of 5 billion images and captions.

quote:

---

Does AI collect images from around the web and piece it together into a new image or is every element generated?

---

I think I answered this above but it’s probably most accurate to say that the AI creates images based on knowledge it gained from images on the web.

quote:

---

For example, would you ever see yourself in an image generated by AI?

---

Only if there were enough images of you in the training dataset to actually influence image generation. Even then, you would have to prompt for it. If you Google image search your name, do you see a page full of your pictures? If not then.. probably not.

quote:

---

Is every person, building fake unless you specify like, say the White House?

---

Good question. If you prompt for “a woman,” you will get the AI’s generic vision for what a woman looks like based on its training data. So it will be some sort of average across all images that were captioned as “woman.” In this way you can actually get an idea of what the overall training data looks like. Prompting “woman” on a model trained from magazine clippings will look very different than the same prompt on a model trained from Facebook images, both in overall style and the appearance of the women themselves.

If you prompt “Jennifer Connelly” you will probably get an image that looks like Jennifer Connelly because her pictures are all over the web. If you prompt “Jennifer” you will get some sort of average approximation of all of the Jennifers in the training data, which will be skewed toward the Jennifers that show up most often in that data.

quote:

---

We’re not in true AI.

---

What do you mean?

That applies to us too, though. We learn, notice patterns, and react to stimuli, based entirely on our past experiences.

AI will never be worse than it is now.

This guy literally doesn't know a word of Spanish: LINK