What is Diffusion in AI Engineering?

In today's video, we'll see what diffusion-based models are. These are quickly replacing large language models for most tasks. Code generation. Image generation. Video generation. To understand diffusion based models, let's see how they generate output. Initially, you have the original query. “What is the best programing language in 2026? ” The LLM is going to be generating tokens one by one from left-to-right. So the final output completely depends on the fact that you have chosen popularity, and your responses are primarily dependent on the initial few tokens that you create. So when it comes to storytelling, writing, the start of your essay is going to be very important for the LLM. This is not how humans think. We can go back and make corrections. We can choose that instead of “popularity”. Let's go for “speed” or “safety”, and therefore our responses can change by going back and forth. But if we pass the same input to a diffusion-based model, then it's going to generate a bunch of blank tokens. The diffusion model now can iteratively improve on this output by replacing some of the tokens. Why is this so useful? Think about images. If you go from left-to-right while generating an image, it is quite likely that you have made a mistake over here or you here. In a autoregressive model, you can't really improve this. In case of a diffusion-based model, since the entire image is constantly being bubbled, is being improved, you can improve the image till it becomes acceptable. The other thing is, if you pass in the data again and again during training, if you pass it in 4x for autoregressive model, then it almost feels like fresh data to the autoregressive model. On the other hand, you have the diffusion-based model, which can have duplicates 100 times! Not 4x, 100x. So you can see that with lesser data, you can still train a diffusion-based model quite well. The reason is, once you pass this data through your neural network, it's going to update its weights. Some of the weights may be inaccurate or incorrect, or may have been changed too much. The second time the same data can be reused. And so to get the most ROI for the limited data that you have, you're looking at diffusion. Thanks for watching. Cheers.