I'm sure that all of us have seen how ChatGPT was evolving. It started with simple text generation, but then slowly it was adding more and more functionality. We have seen how image generation was added and how quickly it was improving. If you look at other models like Gemini or Grok, you can also see that they generate not only images and text but they can also generate video and they can even work with audio information. And there is an obvious question: why have all of these models developed so quickly at the same time?
Is it the same type of model or maybe the same family of models? When people speak about GenAI, they normally think about chatbots. But GenAI is a family of models which is larger than just generation of text. It also includes audio generation, image generation, video generation and also other types of generation. But these four are the main pillars. What unites all of them into one family is how they generate all these artifacts today. And the idea behind it is prediction. So they generate everything by prediction.
If we're talking about large language models, they predict the next token. Audio models predict the signal and generate it over time. When we speak about image generation, it starts from noise and then it predicts structure, how the structure should look. And if we move to video generation, it generates this structure not only in space but over time. So the answer to the question why everything advanced at the same time is the combination of a breakthrough in the technical foundation and also acceleration forces. So first of all, prediction-based generation became a thing across all of these modalities, and then another thing, which is maybe even the most important thing: since the momentum has been built after ChatGPT has been announced and people saw that it can really do magical things, a lot of talent and a lot of capital has been injected into this industry and all the things became significantly faster.
And another thing which is also very important: since these GenAI models share a very similar concept in their generation, it means that a breakthrough in one area will automatically or almost automatically lead to breakthroughs in all other areas, and the scientists who are working on audio generation can take some tricks and some best practices from the scientists who are working on large language models and vice versa. So as a result systems are becoming better and better. They can already generate amazing text and images, and the quality of audio and video generation is improving.
This was a note for today and I'll see you in the next one. Bye.
