Kronos — transformer that reads candlesticks as a language (part 3)

I built a system that makes money while I sleep and I still tinker with it almost every day because it's fun. And I'm obviously not going to give you that system because then it won't work as well for me. But I have no problem telling you how one of the secret tools actually works because then we can talk about in the comments and that's how we learn. I'm Giga Chen and this is part three. The final part to how Kronos, the AI model that can predict financial market prices works. And at the end of this video, you're going to know not only how this model works start to finish, but you're also going to have more knowledge about AI than just about anyone you'll bump into. If you've been following along, you know Kronos uses binary spherical quantization and then hierarchical encoding. Now for step three, we need to talk about the transformer. A transformer is a context engine. It's the same tech that powers chat GPT. A lot of people think that AI just takes an input and gives us an output, but it's not just simple data. In reality, you need to take in the environment or the context surrounding the data. For a while, we've had LSTMs. That's long short term memory recurrent neural networks. They use gates to try and pass information from one step to the next. But this is obviously sequential. You're missing the big picture. A transformer is a context engine because it's parallel, not sequential. The context engine processes everything at once. That's what makes it parallel. It's…